Daniel Pouzzner, firstname.lastname@example.org
Since the earliest electroencephalography experiments, large scale oscillations have been observed in the mammalian brain. More recently, they have been identified not only in the cerebral cortex and thalamus, but pervasively in the healthy basal ganglia. The basal ganglia mediated synchronization model, introduced here, implicates these oscillations in the combination of cortical association mechanisms with stimulus-response and reinforcement mechanisms in the basal ganglia. In the core mechanism of the model, oscillatory patterns in cortex are selected by and routed through the basal ganglia to the thalamus phase-coherently, then circulated back to widely separated areas of cortex, synchronizing those areas and functionally connecting them. Corticostriatal and striatonigral conduction delays are crucial to this mechanism, and evidence suggests that these delays are unusually long, and unusually varied, in arrangements that might facilitate learning of useful time alignments and associated resonant frequencies. Other structural arrangements in the basal ganglia show further specialization for this role, with convergence in the inputs from cortex, and divergence in many of the return paths to cortex, that systematically reflect corticocortical anatomical connectivity. The basal ganglia also target the dopaminergic, cholinergic, and serotonergic centers of the brainstem and basal forebrain, and the reticular nucleus of the thalamus, structures broadly implicated in the modulation of oscillatory network activity and expressions of plasticity. By learning to coordinate these various output channels, the basal ganglia are positioned to facilitate and synchronize activity in selected areas of cortex, broadly impart selective receptivity, attenuate and disconnect interfering activity, and recurrently process the resulting patterns of activity, channeling cognition and promoting goal fulfillment. Evidence suggests that this system is the most versatile and flexible example of a repeating architectural motif, in which subcortical and allocortical structures influence functional connectivity in cortical structures using spike-timing-dependent gain. Dysfunctions in the components of these highly distributed systems are associated with syndromes of perception, cognition, and behavior, notably the schizophrenias, some or all of which might fundamentally be disruptions of subcortically mediated neocortical synchronization.
Keywords: basal ganglia, convergence-divergence, thalamus, intralaminar, thalamocortical, cerebral cortex, synchrony, spike timing, entrainment, selection, functional connectivity, large scale networks, iteration, consciousness, schizophrenia
The cerebral cortex has long been styled the seat of higher thought, due to its size and disproportionate growth in mammalian phylogeny (Mountcastle 1998), and its astronomically large dimensionality (Tononi 2004). This size and dimensionality necessitate an exquisitely powerful coordination mechanism.
Here, I present and explore a hypothesis that phase-coherent signal conduction by the basal ganglia (BG) is fundamental to that mechanism. I term this model basal ganglia mediated synchronization (BGMS). In the core mechanism of the model, oscillatory patterns in cortex are selected by and routed through the basal ganglia to the thalamus phase-coherently, then circulated back to widely separated areas of cortex, synchronizing them using superficial facilitation and spike-timing-dependent gain, thereby establishing and sustaining functional connections. The BG thus act as essential organizers of cortical activity.
A particularly durable account of BG function is that they serve as a selection mechanism, resolving conflicting or ambiguous claims on computational and behavioral resources (Redgrave et al. 1999; Mink 1996; Graybiel 1998; Stephenson-Jones et al. 2011; Hikosaka et al. 2000). Similarly, the BG have been modeled as controllers of gates in cortex, selectively facilitating motor output (Chevalier and Deniau 1990; Hikosaka et al. 2000) and establishing contextually appropriate items in working memory (Frank et al. 2001; O'Reilly and Frank 2006).
At a more fundamental level, the BG are thought to develop a repertoire of compound stimulus-response relations through reinforcement learning, ultimately forming habits (Graybiel 1998, 2008; Markowitz et al. 2018). In this view, the BG transform cortical and subcortical inputs representing goals, elaborately contextualized by other cortical and subcortical inputs representing environmental state and recent history, into spatiotemporally complex, precise, widely distributed, often sequential adjustments to brain state, that are expected to promote internal and external (environmental) changes in furtherance of those goals. It has been previously suggested that a fundamental facility of the BG for precisely and flexibly triggered, structured, and directed neurodynamic gestures has far-reaching consequences (Graybiel et al. 1994; Graybiel 1997). This facility is at the heart of the proposal advanced here, because—as reviewed below—effective connectivity among the targets of the BG is strongly associated with the precise timing relationships of the activity within them.
According to this proposal, the BG and thalamus establish and reinforce effective connections in cortex by distributing precisely synchronized spike volleys to its feedback-recipient layers, imparting discriminative receptivity by spike-timing-dependent gain control. The proposition that spike synchronies are correlates of functional and effective connectivity, and represent associations, is supported by an array of evidence and integrative theory (von der Malsburg 1981, 1999; Bastos et al. 2015b; Bressler 1995; Damasio 1989; Fries 2005; Friston 2011; Hutchison et al. 2013; Kopell 2000; Meyer and Damasio 2009; Siegel et al. 2012; Singer 1993, 1999; Singer and Gray 1995; Varela et al. 2001; Wang 2010).
Synchronous spiking activity is normally oscillatory, and theta (~4-8 Hz), beta (~15-30 Hz), and gamma (~30-80 Hz) oscillations are particularly prominent (Wang 2010). Measurable signals reflecting these oscillations have been known for nearly a century (Berger 1929; Jasper 1937; Buzsáki et al. 2003, 2012; Olejniczak 2006), and local field potentials (LFPs) in particular can serve as proxies for massed neuronal membrane voltage fluctuations and synaptic currents (Haider et al. 2016).
Activity-driven plasticity mechanisms are crucially dependent on relationships of precise temporal coincidence among individual spikes and spike bursts (Song et al. 2000). These mechanisms can construct axon populations with precisely matched propagation delays (Gerstner et al. 1996), and indeed, synchronous spike volleys are thought to propagate coherently through chains of many directly linked neurons (Diesmann et al. 1999).
Neuronal oscillatory periods are intrinsically unstable; thus mutual incoherence is the pervasive and normal condition, except where functional connectivity mechanisms overcome it and establish synchronies (Fries 2005). Once established, synchrony has been shown to stabilize systems against disruption by noise, facilitating information transfer (Tabareau et al. 2010). In principle, coherence coding is intrinsically more robust than rate coding as a mechanism for the selective control of effective connectivity, because a neuronal module cannot a priori generate oscillations at the precise frequency and phase preferred by a particular receiving area. Rather, a transmitting module must obtain knowledge of these dynamic parameters from the receiving module, or from another module functionally connected to the receiver. In contrast, the representation of salience by rate coding per se requires no such dynamic system-level information. Coherence-based access control is particularly apt for generalized cognitive resources, which are subject to “greedy” access strategies by more specialized neuronal modules (van den Heuvel et al. 2012). And because spike generation is energetically costly in itself (Moujahid et al. 2014), timing-based codes have inherently useful metabolic advantages over rate codes.
Synchronies are pivotal in perceptual processing. For example, the relationship of oscillatory frequency and phase in interconnected sensory areas, measured by LFPs, has been shown to strongly influence their effective connectivity (Womelsdorf et al. 2007), and the resolution of competitions among sensory inputs can be predicted from the relationship of the LFP frequency and phase within each input to those prevailing within their common target (Fries et al. 1997, 2002). Attentional orientation is accompanied by LFP synchrony between frontal and posterior cortex: strong beta oscillation initiated in frontal cortex is associated with top-down orientation, and strong gamma oscillation initiated in posterior cortex is associated with bottom-up orientation (Buschman and Miller 2007; Bastos et al. 2015a). Top-down beta enhances bottom-up gamma through cross-frequency interactions, and this phenomenon is suggested to be a fundamental mechanism of attentional focus (Richter et al. 2017) and of top-down control in general (Bressler and Richter 2015). Some long range synchronies implicating prefrontal cortex (PFC) are associated with functional disconnection, manifesting as selective inattention (Sacchet et al. 2015).
Synchronies are also pivotal in the generation of behavior. Long range synchronies can be strongly predictive of behavioral decisions (Verhoef et al. 2011; Fiebelkorn and Kastner 2020‡), and planning and execution of voluntary movements are associated with characteristic synchronization of activity in shifting ensembles of neurons in primary motor cortex, separate from changes in their firing rates (Riehle et al. 1997). Indeed, some of the gesture selectivity of activity in primate motor neurons is apparent only in their synchronies (Hatsopoulos et al. 1998). And a recent study in humans (Fischer et al. 2020) suggests a similar primacy of spike synchrony in the combined cortico-BG dynamics underlying behavior.
The primacy of timing is also apparent in simpler animals. In the hawkmoth, it has been shown that muscles are coordinated with each other almost entirely by millisecond-scale spike timing relationships, with spike timing pervasively encoding 3 times as much information about behavior as does spike rate (Putney et al. 2019; Sponberg and Daniel 2012). Similarly, in Drosophila melanogaster, action selections can pivot on spike timing relationships (von Reyn et al. 2014).
In humans, the large scale architecture of neural synchronies has clear developmental correlates. Childhood improvements in cognitive performance are accompanied by increases in neural synchrony, while adolescence is accompanied by a temporary reduction in performance and synchrony, followed by oscillatory reorganization and still higher performance and synchrony in adulthood (Uhlhaas et al. 2009). Adolescence and young adulthood in humans are marked by uniquely prolonged episodes of myelination, particularly in prefrontal cortex (Miller et al. 2012), while cognitive decline associated with senescence in humans is marked by frontoposterior white matter deterioration, and concomitant deficits in the modulatory control of frontoposterior oscillatory synchrony (Hinault et al. 2019‡).
The functional prominence of temporal precision is suggested by a finding that temporal acuity and psychometric g (a measure of general cognitive performance) covary, with g predicted significantly better by acuity than by reaction time (Rammsayer and Brandler 2007). Similarly, uniformity of cadence in successive gestures within a self-paced rhythm task correlates significantly with performance on a test of general intelligence (Madison et al. 2009).
Characteristic synchronal abnormalities are associated with diseases such as schizophrenia (Sz), autism, Alzheimer's, and Parkinson's (Uhlhaas and Singer 2006, 2012; Hammond et al. 2007), and the reorganization of synchronal architecture in adolescence may be a trigger for the onset of Sz in those at risk (Uhlhaas and Singer 2010; Uhlhaas 2013).
Sz in particular is associated with pervasive physiological disruptions of the mechanisms underlying the generation and regulation of, and responses to, spike synchronies and functional connectivity (Friston 1995; Uhlhaas 2013; Pittman-Polletta et al. 2015), and multifariously implicates the BG (Robbins 1990; Graybiel 1997; Simpson et al. 2010; Wang et al. 2015; Grace 2016; Dandash et al. 2017; Mamah et al. 2007). Abnormal judgment in Sz of time intervals and sensory simultaneity (Martin et al. 2013; Schmidt et al. 2011; Ciullo et al. 2016), and highly significant motor deficits in Sz on tasks as simple as rapidly alternating finger taps (Silver et al. 2003), are further evidence of common timing-related mechanisms underlying sensory, motor, and cognitive processing. These abnormalities are likely to implicate the BG directly: for example, evidence suggests that phase-aligned synchronization of the BG with oscillatory activity in cortex is integral to precise judgments of time intervals (Gu et al. 2018).
Much of the large scale oscillatory activity in cortex is not purely intrinsic, and directly implicates subcortical structures, particularly the thalamus. The thalamus is a major target of BG output (Haber and Calzavara 2009), and the proposition that the BG have a prominent role in controlling long range cortical synchronies follows in part from evidence that the thalamus performs this function.
Thalamic control of cortical oscillation and synchronies follows naturally from the developmental relationship of thalamus to cortex. While a ballet of cortically intrinsic developmental processes parcels the cortex into its major cytoarchitectonic areas (Rakic 1988; Wang 2020), thalamocortical axons reach their pallial destinations before neurogenesis and migration of the receiving cortical neurons (López-Bendito and Molnár 2003; Paredes et al. 2016), and the basic architecture of cortex is thought to develop partly in response to patterns of activity in these axons (Katz and Shatz 1996). “Developmental exuberance”, entailing the robust proliferation of ephemeral long range links in cortex, is followed by a postnatal paring process driven in part by early patterns of thalamocortical activity (Innocenti and Price 2005; Price et al. 2006). Indeed, the manipulation of thalamocortical input patterns can dramatically alter cortical physiology and function (Rakic 1988). For example, uniquely visual attributes can be induced in cortical areas that normally subserve audition by rerouting retinal inputs to the thalamic auditory nuclei (Sharma et al. 2000).
These roles establishing the anatomical connectivity and intrinsic function of cortex position the thalamus uniquely to regulate cortical functional connectivity.
The thalamus is also uniquely positioned anatomically, at the base of the forebrain on the midline. This is an optimal situation for distributing synchronized spike volleys to far-flung loci in cortex, notwithstanding thalamocortical distance disparities due to sulci and gyri. It is striking that postnatally (week 4 in mice), the thalamocortical projection to a given functional area of cortex develops a uniform delay, in many areas less than 1 ms of maximum disparity, despite widely varying axon lengths; even intermodally, thalamocortical delays are often aligned within 2-3 ms (Salami et al. 2003; Steriade 1995). The central clustering of thalamic nuclei is noteworthy in itself: absent functional requirements and associated evolutionary pressures to the contrary, many of these nuclei might migrate toward the cortical areas with which they are intimate, realizing physiological efficiencies (Scannell 1999). Moreover, in many mammals the dorsal BG maintain rough radial symmetries centered on the thalamus, suggesting time alignment pressures like those that appear to influence the gross anatomy of the thalamus.
It has been shown clearly that the thalamus can control cortical oscillatory activity (Poulet et al. 2012), and that it can orchestrate lag-free (zero phase shift) long range synchronies in cortex (Ribary et al. 1991; Vicente et al. 2008; Saalmann et al. 2012). “Desynchronization” associated with mental activity in fact consists of focal, high-frequency (20-60 Hz) synchronization of distributed thalamocortical ensembles (Steriade et al. 1996). Long distance, multifocal (posterior visual, parietal, and frontal motor), lag-free synchronies in the beta band have been observed in association with visuomotor integration (Roelfsema et al. 1997), and similar lag-free beta synchronies, and precise antisynchronies, have been observed among loci in prefrontal and posterior parietal cortex in visual working memory (Dotson et al. 2014; Salazar et al. 2012). Phase-locked rhythmic stimulation of widely separated loci in frontal and parietal cortex can significantly improve working memory performance, strongly suggesting that extrinsic synchronous influences on cortical activity meaningfully affect cortical network configuration and associated mental faculties (Violante et al. 2017; Alagapan et al. 2019).
Projections from single thalamic nuclei to widely separated but directly interconnected cortical areas have been noted (Goldman-Rakic 1988; Saalmann et al. 2012), and there is evidence that intralaminar thalamocortical projections systematically reflect corticocortical connectivity, with individual axons branching multi-areally (Kaufman and Rosenquist 1985a; Van der Werf et al. 2002). The hypothesis has been advanced that midline and intralaminar thalamic nuclei in particular are the hub of a system to control cortical synchronies and associated effective connectivity (Saalmann 2014; Purpura and Schiff 1997), and the entire population of calbindin-positive neurons in the thalamus (Jones 2001), or indeed the thalamus as a whole (Halassa and Kastner 2017), has been proposed to function in this fashion.
Recently reported evidence supports this proposition. The mediodorsal nucleus is positioned to gate afferent inputs to PFC through direct connections to cortical interneurons (Delevich et al. 2015), and has been shown to control sustained functional connectivity in PFC (Schmitt et al. 2017; Nakajima and Halassa 2017) and shifts in PFC representations of rule context (Rikhye et al. 2018). Similarly, the pulvinar can synchronize oscillations, establishing functional connectivity associated with attentional engagement, between areas V4 and TEO in visual cortex (Saalmann et al. 2012), between V4 and the associative visual cortex in the lateral intraparietal area (LIP) (Saalmann et al. 2018‡), and between LIP and the frontal eye field (FEF) (Fiebelkorn et al. 2019). This influence of the pulvinar on cortical activity can be confined to modulations of aggregate inter-areal phase coherence, with no significant effects on local firing rates or local aggregate oscillatory power (Eradath et al. 2020‡).
As detailed later (§7.8), the interaction of physiologically distinct but spatiotemporally coincident inputs to cortex from intralaminar and non-intralaminar BG-recipient thalamus is a key mechanism within the model proposed here. In short, GABAergic output fibers from the BG, bearing phasically rhythmic activity, appose the distal dendrites of projection neurons in the intralaminar thalamus. These phasic inputs act as frequency- and phase-selective filters, favoring corticothalamic inputs with the preferred frequency and phase, while disfavoring others. The associated intralaminar thalamocortical projections then carry signals that broadcast those frequency and phase preferences to wide areas of cortex, with high temporal specificity. The multi-areal matrix cells (Jones 2001; Clascá et al. 2012) in non-intralaminar nuclei receive powerful, enveloping somatic inputs from collaterals of the same population of BG output fibers; thalamocortical projections from this population mainly target superficial layers, reinforcing selected activity and disfavoring unselected activity, in spatially delimited predominantly frontal cortical areas, with considerably less temporal specificity than the intralaminar paths. These distinct inputs to cortex interact, locally and inter-areally, with each other and with intrinsic cortical activity, arranging for dynamic recruitment of specific, contextually appropriate large scale cortical networks, and for their contextually appropriate dissolution.
It has long been appreciated that the cortex, striatum, pallidum/substantia nigra, and thalamus are arranged in loops placing each under the influence of the others (Alexander et al. 1986; Parent and Hazrati 1995a; Middleton and Strick 2000). As reviewed in detail later (§6.2), the pyramidal neurons of cortical layer 5 (L5) originate the primary input to the BG “direct path” centrally implicated in these loops, and are among the recipients of the output from the direct path via the thalamus. While subdivision of these loops into parallel circuits and constituent channels has been noted (Alexander et al. 1986, 1991), in toto the pathways of the BG exhibit remarkably varied patterns of convergence, divergence, and reconfiguration (Joel and Weiner 1994; Zheng and Wilson 2002; Hintiryan et al. 2016).
Diffuse projection fields from wide areas of cortex exhibit high convergence-divergence, and are thought to supply extensive context throughout the striatum (Calzavara et al. 2007; Mailly et al. 2013). Projections from interconnected cortical regions, including reciprocally interconnected pairs of individual neurons, systematically converge and interdigitate in the striatum (Van Hoesen et al. 1981; Selemon and Goldman-Rakic 1985; Parthasarathy et al. 1992; Flaherty and Graybiel 1994; Averbeck et al. 2014; Lei et al. 2004; Morishima and Kawaguchi 2006; Hintiryan et al. 2016; Hooks et al. 2018), and projections from interconnected areas have been shown to converge on individual striatal fast spiking interneurons (FSIs) (Ramanathan et al. 2002). These arrangements show that the BG are particularly concerned with corticocortical connectivity. Even before much of this evidence was uncovered, it was suggested that arrangements of convergence and interdigitation in the corticostriatal projection position the striatum to integrate, compare, or synchronize neural computations in distant areas of cortex (Mesulam 1990).
By having a sharp view of afferents from directly interconnected areas, simultaneous with a diffuse view of more widespread cortical activity, a striatal neighborhood is supplied with information upon which appropriate corticocortical connectivity decisions might be made as a function of present connectivity and context, with particular expertise for the functional domains implicated by those focal afferents. And given closed-loop circuitry, a striatal neighborhood convergently innervated by multiple cortical areas can impart oscillation from one of them to the others, with particular significance for directly interconnected areas, and areas linked via a common connectivity hub. Nonetheless, partial segregation of channels through the BG likely facilitates parallel processing of operations that require only partial coordination, with the degrees and directions of segregation tending to reflect the degrees and directions of non-interference and independence.
Parallelism in the BG provides for the simultaneous processing in the striatum of activity at multiple oscillatory frequencies in distinct regions, associated with distinct domains of skill acquisition and performance, with distinct expressions of plasticity in each region, and inter-regional coherence varying task-dependently (Thorn and Graybiel 2014). In cortex, too, evidence suggests that distributed functional networks are largely parallel, and entail interdigitation in circuit nodes, particularly in prefrontal and other associative areas (Goldman-Rakic 1988; Yeo et al. 2011; Livingstone and Hubel 1988), even while most areas have direct anatomical connections with each other (Markov et al. 2014). fMRI of spontaneous activity in resting humans has demonstrated corresponding integration, regionalization, and parallelism of cortico-BG networks (Di Martino et al. 2008).
Pathways through the basal ganglia exhibit an unusually broad range of conduction delays (Yoshida et al. 1993; Kitano et al. 1998). This diversity of delays plays a central role in the model introduced here, allowing the BG to meet disparate timing requirements at each stage of the BG-thalamocortical loop, and to tune the preferred frequencies and phases of oscillation at network loci implicated by a selection. Indeed, as detailed later (§3.7), closed loops through the dorsal striatum and globus pallidus have an average transmission delay corresponding to 40 Hz gamma oscillation, and loops through the the dorsal striatum and substantia nigra have an average delay corresponding to 20 Hz beta oscillation, with wide delay ranges along either path.
Prominent oscillatory activity in the BG, particularly beta oscillation, is associated with perception, attention, decision making, and working memory (Cannon et al. 2014), all of which implicate large scale brain networks. Experiments in rats have demonstrated rapid, coherent oscillatory reset and brief beta oscillation, spanning the BG, in response to sensory cues (Leventhal et al. 2012), and mechanisms intrinsic to the BG are posited to underlie the generation of some of these brief oscillatory episodes (Mirzaei et al. 2017). Closed-loop networks internal to the basal ganglia are intrinsically capable of generating and sustaining oscillations in the beta range (Bevan 2002; Tachibana et al. 2011; Mirzaei et al. 2017), and fast spiking interneurons in the striatum have distinctly oscillatory dynamics in the theta, beta, and gamma bands, which are systematically altered by dopamine (Berke 2009; van der Meer et al. 2010; Berke 2011; Chartove et al. 2020). These mechanisms might position the BG to generate contextually appropriate oscillatory responses to non-oscillatory inputs, and to tune input frequency preferences state-dependently.
The BG are among the most connected regions of the brain (van den Heuvel and Sporns 2011), and are densely integrated with cortical hubs (Middleton and Strick 2002; Vatansever et al. 2016; Averbeck et al. 2014). They participate in a particularly wide variety of large scale synchronized networks, with greater oscillatory specificity than cortical areas (Keitel and Gross 2016), suggesting primary oscillatory selection and generation.
BG influence on cortical activity is extensive, and can be strong. Cortical oscillatory dynamics, stability, and propensity for synchrony, are profoundly and specifically modulated by central supplies of dopamine, acetylcholine, and serotonin, all of which are integral to BG circuitry (Yetnikoff et al. 2014; Fallon 1988; Ioanas et al. 2020‡; Saunders et al. 2018; Yang and Seamans 1996; Towers and Hestrin 2008; Costa et al. 2006; Benchenane et al. 2010; Mesulam and Mufson 1984; Grove et al. 1986; Haber et al. 1990; Haber 1987; Sillito and Kemp 1983; Rodriguez et al. 2004; Muñoz and Rudy 2014; Howe et al. 2017; Baumgarten and Grozdanovic 2000; Neuman and Zebrowska 1992; Gervasoni et al. 2000; Carter et al. 2005). Artificial stimulation of the mouse striatum affects activity spanning the entire cerebral cortex (Lee et al. 2016), and rhythmic photostimulation of optogenetically manipulated mouse BG output structures can produce synchronous spike volleys in the motor thalamus and motor cortex (Kim et al. 2017). In monkeys, task-related oscillatory activity in the BG correlates strongly with oscillation in the implicated areas of thalamus (Schwab 2016, chapter 5), and BG oscillations like these appear to induce phase-locked oscillation in frontal cortical areas (Antzoulatos and Miller 2014; Williams et al. 2002).
The widespread influence of the BG on neocortical activity suggests a general role, integral to the normal operation of cell assemblies. This accords with the pervasive dynamical criticality of neocortex (Haider et al. 2006; Ma et al. 2019; Ahmadian and Miller 2019‡), which optimizes sensitivity to inputs, and dynamic control by those inputs of functional connections and disconnections (van Vreeswijk and Sompolinsky 1996; Vogels and Abbott 2009; Ahmadian and Miller 2019‡; Finlinson et al. 2019‡; Li et al. 2019). This arrangement has particular implications for the influence of the BG on the neocortex (Djurfeldt et al. 2001).
In the model introduced here, the BG tune circuits within the cortico-BG-thalamocortical loop so that selected signals return to cortex precisely coincident in time. Cortical pyramidal cells and cell assemblies targeted by the BG-recipient thalamus are arranged to respond selectively when their thalamocortical and corticocortical inputs are time-coincident (Llinás et al. 2002; Larkum et al. 2004; Pouille and Scanziani 2001; Volgushev et al. 1998), which arranges for sharp selectivity as a function of alignments in time. And indeed, habit learning is associated with the gradual emergence of widespread task-related spike synchronies and sharpened responses in the striatum (Barnes et al. 2005; Howe et al. 2011; Desrochers et al. 2015).
Sensitivity to widespread synchronies is also intrinsic to striatal physiology (Zheng and Wilson 2002), and subcortical projections from the BG-recipient thalamus to the striatum (Sidibé et al. 2002; Smith et al. 2004; McFarland and Haber 2000) suggest that dynamics and plasticity in the BG are driven in part by the synchronies present at their output. If the coherent relay of oscillatory signals with precise selectivity for time relations is a key function of the BG, as proposed here, then these arrangements are central to that facility: a contextually contingent, sharp, coordinated, widely distributed striatal spike volley could then evoke or reinforce widely distributed synchronous rhythmic spiking, and consequent functional connectivity, in downstream structures.
BG-facilitated burst firing in cortex might activate functional connections: Excitatory input from thalamus to the L1 (apical) dendrites of L5 pyramidal neurons, synchronous with somatic layer inputs to those neurons, promotes burst firing and concomitant long range synchronization (Larkum et al. 2004; Larkum 2013; Womelsdorf et al. 2014). BG input to the thalamus, affecting the temporal structure of activity there rather than its intensity, has been shown to be crucial for a similar type of pallial burst generation in songbirds (Kojima et al. 2013). These sorts of burst-induced transient long range synchronies are thought to provide for flexible information routing (Palmigiano et al. 2017), and dysfunctions of this mechanism, spuriously routing information, plausibly underlie the delusions and hallucinations associated with psychedelic drugs and schizophrenia (Carter et al. 2005; González-Maeso et al. 2007; Preller et al. 2018; Geyer and Vollenweider 2008; Ji et al. 2019; Giraldo-Chica et al. 2018), and plausibly implicate the BG (Walpola et al. 2020; ffytche et al. 2017).
In humans, the coherence of oscillatory activity in the BG, and the relationship of its phase to that in cortex, have been shown in some scenarios to be more predictive of movement than are BG mean firing rates (Fischer et al. 2020). And in tasks with shifting stimulus-response contingencies, the human BG have been shown to establish appropriate functional connections between prefrontal and posterior visual cortex (van Schouwenburg et al. 2010b).
Drawing on these findings, I propose the basal ganglia mediated synchronization (BGMS) model, and detail its mechanistic components and their relations below. In the BGMS model, the BG learn to recognize salient patterns of distributed, phase-correlated cortical activity, responding with synchronized spike volleys with functionally optimal delays, relayed via the thalamus, to the feedback-recipient layers of other areas of cortex, and back to those of the cortical areas of origin. These spike volleys reinforce activity in the areas of origin, promote contextually appropriate activity in allied areas, and establish and sustain selective long range oscillatory synchronies and consequent effective connectivity with other areas, both by elevating receptivity in the other areas by superficial facilitation and spike-timing-dependent gain, and—with stronger and more coherent activity—by promoting burst firing.
The long and diverse conduction delays of corticostriatal projection fibers provide for the temporal focusing of widely distributed, phase-locked but phase-dispersed cortical activity, rendering it coincident as it converges on sparse subsets of striatal spiny projection neurons (SPNs). Emergent task-related synchronies and response sharpening in the striatum, noted above (Barnes et al. 2005; Howe et al. 2011; Desrochers et al. 2015), suggest this dynamic. The similarly long and diverse delays of striatopallidal/striatonigral pathways provide for the temporal focusing of phase-skewed SPN outputs on sparse subsets of pallidal/nigral output neurons, and provide for additional phase shifting of BG outputs to the thalamus, to meet coincidence criteria in cortex associated with inter-areal phase relationships. The resulting combined delays through the BG direct path define preferred oscillatory periods and inter-areal phase relations.
A diversity of delays through the BG, focusing activity from distributed networks on particular striatal and pallidal projection neurons, is analogous to the “tidal wave” timing mechanism proposed by Braitenberg et al. (1997). This mechanism centers on the dynamics of granule cell input to the Purkinje cells of the cerebellar cortex via the system of parallel fibers, and has been demonstrated in vitro, in vivo, and in simulation (Braitenberg et al. 1997; Braitenberg 1961; Heck 1993, 1995; Heck et al. 2001; Heck and Sultan 2002; Sultan and Heck 2003). Indeed, a BGMS-like mechanism centered on the cerebellum has been demonstrated directly (Liu et al. 2020‡; McAfee et al. 2019; Popa et al. 2013).
Within the BGMS model, structures associated with the BG “indirect” and “hyperdirect” paths inhibit, desynchronize, or antisynchronize conflicting, aborted, irrelevant, and completed activity, and in general, delimit network activity, consistent with functions already proposed and demonstrated for these structures (Smith et al. 1998; Parent and Hazrati 1995b; Schmidt et al. 2013; Lee et al. 2016). Closed loops through the direct path, with tuned delays, inherently promote oscillations at preferred frequencies, while the indirect path must terminate these oscillations when they are no longer useful. Additionally, indirect path structures are considered (through the STN) to consolidate selections, amplifying localized BG activity (in particular, oscillations) to influence large surrounding areas in the BG.
BG influences on dopaminergic, cholinergic, and serotonergic centers, and the thalamic reticular nucleus, are proposed to be coordinated with (and partly by) these direct and indirect path outputs, promoting activity that contributes to the selected effective connections, attenuating or functionally disconnecting activity that conflicts with them, and modulating the dynamics within effective connections, to promote gainful computation and motor output. These mechanisms also greatly influence the expression of plasticity in the implicated structures, orienting neurophysiological investments to favor salient stimuli and behaviors, aligning selections with expectations and goals, and improving the immediacy, precision, and thoroughness of those selections.
As detailed throughout this paper, the BGMS model helps explain several historically mysterious aspects of BG and related physiology, among them:
All of these phenomena are, in principle, explained by the proposition that the intact BG recognize and select useful patterns of synchronized cortical activity, and route the predominant oscillations within them, chiefly via the thalamus, back to cortex, synchronizing widely separated areas and broadly promoting precisely discriminative receptivity to selected activity.
The BGMS model is hardly the first to ascribe to the basal ganglia the control of functional connectivity in cerebral cortex.
O'Reilly and Frank (2006), mentioned above, propose that the BG adjust large scale functional connectivity to fit context, controlling the formation, activation, and extinction of working memories. Stocco et al. (2010) propose that the BG act to control the routing of information within cortex, dynamically establishing bridges between “source” and “destination” regions to facilitate goal-directed cognition. In a similar vein, Hayworth and Marblestone (2018‡) propose such a role for the BG within a biologically inspired machine learning model, formulated to emphasize gating actions by the BG through the inhibition-disinhibition mechanism described by Chevalier and Deniau (1990), and inter-areal routing of information via the thalamic relay mechanism described by Guillery and Sherman (2002).
While those models do not expressly consider the modulation of oscillatory synchronies in the control of information routing, others do.
Fountas and Shanahan (2017‡) propose and simulate a model in which coalitions of oscillating cortical inputs to the BG change the course of information flow in the latter in a frequency-specific fashion, and are crucial to selection dynamics. Population-level synchronies, band pass filtering, and dynamically selective reinforcement of oscillatory frequencies have been shown to be a plausible mechanism for flexible, selective signal routing and functional connectivity in and beyond neocortex, even absent the direct involvement of subcortical structures such as the BG and cerebellum (Akam and Kullmann 2010; Sherfey et al. 2019‡). In particular, the model articulated by Sherfey et al. (2019‡), in which frequency-specific reinforcement of oscillations in PFC govern large scale functional connectivity underlying working memory, is compatible with the BGMS model.
My approach in introducing BGMS here is the obvious—to review the functional physiology of the BG and thalamocortical systems, much of it established in studies conducted in previous decades, recontextualized to the BGMS model. Following the foregoing introduction, I begin with a discussion of the role posited for the BG in gating motor output, and the relevance of precision spike timing in that role, followed by a review of BG path delays, proposed delay plasticity mechanisms, and patterns of convergence and divergence in these paths, all fundamental to the BGMS model. Following this is a discussion of the general relationship of the BG to cortex and thalamus from a signal processing perspective, and a detailed review of the areas of thalamus receiving BG direct path output, the areas of cortex receiving output from the BG via the thalamus, the principal pathways through the BG (direct, indirect, and striosomal), the modulatory functions of dopamine, acetylcholine, and serotonin in the thalamocortical system, and their integration into BG circuitry. Finally, I consider the roles of the BG in sensory perception and general cognitive coordination, discuss some relationships and contrasts with the cerebellar, hippocampal, and other analogously positioned systems, and extend the foregoing explorations to their logical conclusion, that the BG are mechanistically integral to and necessary for consciousness.
Within the theoretical framework of synchrony-mediated effective connectivity, von der Malsburg (1999) mused that “If there were mechanisms in the brain by which connections could directly excite or inhibit each other, fast retrieval of associatively stored connectivity patterns could be realized.” Implicitly, the BGMS model is a proposal that the BG, with the thalamus, implement such a mechanism, enabling patterns of activity in corticocortical connections to excite and inhibit other connections with nearly arbitrary flexibility. And as discussed later (§13), this system may be just one instance of a general architectural motif in the mammalian brain, with homologues in other vertebrates, in which subcortical and allocortical structures influence functional connectivity in cortical structures by spike-timing-dependent gain.
It has long been recognized that the BG are integral to movement performance (DeLong and Georgopoulos 2011; Chevalier and Deniau 1990). Selective disinhibition of tonically inhibited motor centers, concurrent with enhanced inhibition of unselected motor centers, is a prominent model for this involvement (Chevalier and Deniau 1990; Hikosaka et al. 2000). However, firing rate models do not fully describe the implicated mechanisms (Goldberg et al. 2012; Kojima et al. 2013).
Various lines of evidence underscore the complexity of these mechanisms. Removal of ostensibly inhibitory pallidal input to thalamus for treatment of Parkinson's disease (PD) does not result in an excess of movement (Brown and Eusebio 2008; Marsden and Obeso 1994; Kim et al. 2017), and manipulation of the oscillatory phase of STN stimulation optimizes alleviation of parkinsonian symptoms, without affecting STN unit firing rates (Holt et al. 2019). Direct and indirect path activation have effects on activity levels in BG direct path output structures opposite those predicted by the inhibition-release model (Lee et al. 2016), and BG direct path output structures and receiving thalamic structures often show simultaneous movement-related rate increases (Schwab et al. 2019‡).
All cortical output to the brainstem and spinal cord arises from pyramidal neurons in L5, whose apical dendrites are in L1 (Deschênes et al. 1994), apposed directly by the terminals of BG-recipient thalamic projection neurons (Kuramoto et al. 2009; Jinnai et al. 1987). These appositions are excitatory. The apical and proximal dendritic processes of these pyramidal neurons are thought to interact as a coincidence detector (or “vertical associator”) mechanism, with a window width of 20-30 ms, particularly gating burst generation (Larkum et al. 2004; Larkum 2013). This implies that BG facilitation of thalamocortical spiking has temporal specificity at least as fine as this time scale. Somatic coincidence detection in pyramidal neurons is subject to a much tighter window, ~4 ms (Pouille and Scanziani 2001; Volgushev et al. 1998), and evidence is reviewed later (§7.5) suggesting BG alignments at this much finer time scale, particularly implicating the intralaminar nuclei.
Motor performance entails patterns of synchronized activity in motor neurons (Riehle et al. 1997; Hatsopoulos et al. 1998), though interestingly, these synchronies are not consistently driven by motor neurons. For example, Granger causality analysis of LFPs in sensorimotor cortex suggests that sensory and inferior posterior parietal cortex (PPC) drive sustained beta oscillation in motor cortex during a sustained gesture (maintenance of a hand press) (Brovelli et al. 2004). Beta oscillatory synchrony in premotor cortex during delay periods appears to be extrinsically driven; this activity is selective for specific features of the forthcoming gesture, and is displaced by simultaneous bursting immediately before the onset of movement (Lebedev and Wise 2000). The BG are understood to be integral to these phenomena, and recent findings suggest the nature of this integration: The PPC has distinct projections to premotor cortex and striatum, with premotor cortex receiving signals that control movements, while signals to striatum reflect task-historical context, informing decision making (Hwang et al. 2019). This arrangement suggests that BG output to premotor cortex in such tasks is likewise driven by input from PPC. According to BGMS, for facilitatory decisions, this input will be phase-locked, and after learning, phase-aligned, to corticocortical inputs from PPC to premotor cortex.
Preparatory and sustained activity in premotor cortex depends crucially on excitatory inputs from the BG-recipient motor nuclei of the thalamus, and activity in those nuclei is likewise largely dependent on activity in premotor cortex (Guo et al. 2017). That the BG are directly implicated in these dynamics is suggested by the finding that oscillatory activity in the subthalamic nucleus (STN) couples coherently to gamma oscillations in motor cortex, in apparent preparation for action, with performance improving with increased anticipatory phase coupling (Fischer et al. 2020).
Control of gates in these corticocortical connections by the BG appears to depend on consistent rhythmicity in the implicated cortical activity. The delays of paths through the BG (e.g. 50 ± 15 ms via the substantia nigra to the frontal eye field, reviewed in detail later (§3)) are significantly longer than the corresponding corticocortical delays (e.g. 8-13 ms between area V4 in visual cortex and the frontal eye field (Gregoriou et al. 2009)), so that trans-BG spike volleys triggered by a given cortical spike volley return to cortex outside the coincidence window of that same cortical spike volley traveling corticocortically. Moreover, activation of the BG is dependent on synchronized cortical activity, due to physiology in the striatum (reviewed later (§4.8)). Thus, the BG can open a corticocortical gate, at the earliest, for the second in a series of synchronized spike volleys.
Volleys from a particular efferent area, in a particular scenario, must have a consistent characteristic cadence, even if only for a spike volley doublet, in order for coincident arrival to be possible (and, as proposed later (§3.4), learnable) for volleys traveling both corticocortically and through the BG to the same target area. Of particular relevance to this mechanism, behavior-correlated spiking in premotor and primary motor cortex has been found to always endure for at least one cycle of oscillation, and to often endure for only one (Churchland et al. 2012), representing the parsimonious spiking pattern for integration with the BG. Indeed, once the BG have learned a stimulus-response relation, with an associated optimal delay, activation of that relation intrinsically promotes oscillation at the particular frequency associated with that delay. Closed-loop delays, combined with implicated pyramidal neuron response delays, can be viewed as oscillatory periods (although the dynamical relationship between delays and network oscillations is, in general, more complicated (Ivanov et al. 2019‡)).
Sustained activity has also been proposed to be necessary for conscious cognition, entailing “dynamic mobilization” of long range functional networks (Dehaene and Naccache 2001; Dehaene and Changeux 2011). Irreducible delays in BG responses to preconscious cortical and thalamic activity might figure prominently in this dependency.
That the BG generally facilitate effective connectivity using multi-areally synchronized spike volleys is suggested indirectly by the finding that a solitary, precisely timed spike volley from the subthalamic nucleus (STN) to the substantia nigra reticular part (SNr) can be effective in stopping (preventing) behavior (Schmidt et al. 2013). This might be evidence that the disruption of the timing of BG output is enough to abolish its facilitatory effect, so that its timing is implicitly critical. STN axons appose neurons in the SNr throughout their input processes, both somatically and dendritically (Bevan et al. 1994; Tiroshi and Goldberg 2019), suggesting that paths through the STN are arranged to strongly modulate other inputs.
The precise timing of STN spikes has recently been shown to be more broadly significant. As noted above, STN activity couples to cortical gamma activity: the STN and motor cortex exhibit phase relationships specific to behavioral scenarios, shifting by 180 degrees for ipsilateral versus contralateral gripping, and measures of these phase relationships better predict behavioral performance than do STN mean firing rates, which were found to not change significantly over the course of behavior (Fischer et al. 2020).
As suggested in earlier accounts (e.g. Parent and Hazrati 1995b), the STN appears arranged to act as a crossroads, through which activity can spread from localized sectors of the BG to wider areas, implying competition and facilitating the completion of selections. Through its strong and highly divergent appositions on projection neurons in BG output nuclei, the STN might broadly entrain BG output to a winning rhythm, maximizing or indeed minimizing the efficacy of cortical activity converging with BG output. Once a winning rhythm is no longer useful (not contextually appropriate), the STN is positioned to disrupt it throughout the BG.
The intralaminar nuclei of the thalamus also target the STN, and through it, the SNr and its dorsal homologue, the globus pallidus internal part (GPi) (Sadikot et al. 1992a; Parent and Hazrati 1995b). As explored in depth later (§7), the intralaminar nuclei are themselves positioned to distribute oscillatory activity to cortex and BG very broadly and with high temporal fidelity, reflecting the combined effects of interacting inputs arising in cortex and the BG.
Among mammalian species, conduction velocities (CVs) for a given homologous projection vary widely, while alpha, beta, and gamma oscillatory frequencies are roughly constant, despite a 17,000-fold variability in brain volume (Buzsáki et al. 2013). Geometrically proportional scale-up of axonal propagation velocities appears to arrange for similar long range delays regardless of size, maintaining the compatibility of circuit synchrony mechanisms with the conserved and intrinsic dynamics of neurons and their microcircuits, with few exceptions (Buzsáki et al. 2013; but see Caminiti et al. 2009).
The BG are as beholden to these stable neuronal dynamics as is the rest of the brain, but according to the BGMS model, they must additionally align their responses to meet the timing requirements in each learned combination of scenario, efferent areas, and recipient areas, necessitating enormous spatiotemporal flexibility and precision.
Context-specific distributed neocortical activity patterns conform to stereotyped timing relations, with jitter of only 1-3 ms in many of the repeating firing sequences (Abeles et al. 1993); this activity, relayed through corticostriatal fibers, can be rendered coincident (and therefore effective) only by applying different delays to different inputs. Structurally intrinsic inter-areal delays (e.g. Gregoriou et al. 2009; Nowak and Bullier 1997; Schmolesky et al. 1998) must similarly (or equivalently) be compensated, for inter-areally distributed activity to be rendered coincident within the striatum.
The optimal delays for complete circuits from cortex, through subcortical structures, to other areas of cortex, vary significantly, reflecting both the dominant oscillatory frequency of the cortical activity, and the delay of the implicated corticocortical projections (Gregoriou et al. 2009; Nowak and Bullier 1997).
Particular large scale functional networks exhibit characteristic profiles of prevailing oscillatory frequencies, with the oscillatory activity of a particular area dynamically dependent on the network with which that area is functionally connected (Keitel and Gross 2016; Becker and Hervais-Adelman 2019‡), so that reinforcement of particular frequencies in cortex—particularly, by resonances determined by cortico-BG-thalamocortical delays—in itself can establish and stabilize corresponding large scale networks characterized by those frequencies.
There is evidence that synchronous oscillatory stimulation of direct path output structures such as GPi systematically modulates the amplitude of oscillation in related, indirectly connected structures such as STN, as a function of the phase of oscillations in the output structures relative to that in the related structures; favored phases produce large increases in the amplitude of oscillations in the related structures, while outputs at antiphase to the favored phases attenuate the related oscillations (Sanabria et al. 2020‡). While these effects may be due to experimental artifacts (antidromic activations) (Sanabria et al. 2020‡), they suggest that coherent relay of oscillations through the BG can have bidirectionally selective effects, pivoting on the precise timing of BG output.
Simulations suggest that small inter-areal phase delays of ~4 ms can be decisive in determining the direction of information transfer in reciprocal long range links, from phase-leading to phase-lagging areas (Palmigiano et al. 2017). And broad and systematic diversity in the delays attending cortical responses to sensory stimuli, apparent in the visual system of the monkey (Schmolesky et al. 1998), plausibly allow the BG to bias the salience of a selected dimension of the stimulus, by aligning the phase of BGMS signals to selectively reinforce activity associated with that dimension, and decouple activity associated with other dimensions.
In the BG, the obvious substrate for meeting time alignment requirements is the enormous variety of paths, delays, and time constants of striatal neurons, inputs, and outputs. A multiplicity of paths, exhibiting a multiplicity of delays, may assure that for any two cortical loci, there exist polysynaptic paths to the implicated thalamocortical neurons, exhibiting nearly optimal delays, that need only be strengthened to effect learning of appropriately selective, timed, and directed responses.
The general mechanisms whereby the BG accommodate these diverse timing requirements likely endow them with particularly rich representational power: When similar arrangements in cortex were simulated, an unanticipated result was that the number of distinct ephemeral neuronal assemblies greatly exceeded the number of neurons, and might even exceed the total number of synapses in the network (Izhikevich 2006).
Importantly, neuromodulatory projections from the brainstem and basal forebrain project not only to cortex and thalamus, but extensively to the BG. Thus oscillatory acceleration in cortex and thalamus is likely accompanied by acceleration in the BG. This might arrange to preserve the applicability of timing relationships learned by the BG at widely varying levels of arousal.
Plasticity mechanisms in the central nervous system are exquisitely sensitive to timing relationships, at time scales of several or even fractional milliseconds within a ±20 ms window, so that in many neurons, faster paths of communication are consolidated, and slower paths are culled (Markram et al. 1997; Bi and Poo 1998, 2001; Song et al. 2000). Spike-timing-dependent plasticity (STDP) in conjunction with coherent oscillatory activity may build temporally coherent circuits by grouping axons with precisely matching delays (Gerstner et al. 1996). However, the relationship of these mechanisms to the delay of polysynaptic BG paths is complicated, given evidence that STDP in striatal SPNs is reversed (Fino et al. 2005). Nonetheless, striatal FSIs exhibit typical STDP (Fino et al. 2008). The implications of this are explored below.
fibers exhibit fairly slow average CV, measured
to be about 3 m/s in macaque
(delay range 2.6 - 14.4 ms), in marked contrast to corticopeduncular fibers,
measured to average greater than 20 m/s (delay range 0.75-3.6 ms) (Turner and DeLong 2000).
Striatopallidal fibers are markedly slower still, measuring under 1 m/s in
(Tremblay and Filion 1989).
The typical striatopallidal CV is so slow that at peak spike rate
(~80 Hz (Kimura et al. 1990)), apparently more than one action
potential can be propagating simultaneously on the same axon.
Not only are corticostriatal and striatopallidal/
It is significant that CVs are slow and diverse in both the corticostriatal and striatopallidal/striatonigral projections. Locus-specific phase disparities, associated with converging corticostriatal inputs from widely separated but functionally connected loci, can be compensated by distinct conduction delays in their respective corticostriatal projection fibers. Activation of a multi-areal cortical ensemble can then produce spatiotemporally coincident activity at particular striatal FSIs and SPNs, despite phase skews in the ensemble at the cortical level. In essence, the spatiotemporal pattern of activation in cortex is convolved with the function embodied by the corticostriatal projection, so that particular cortical activation patterns are focused on particular cells in the striatum, which can learn to respond to them. Separately and subsequently, the output from SPNs is subjected to slow and diverse CVs in the striatopallidal projection, by which additional phase corrections can be applied to align outputs from converging but phase-skewed striatal neighborhoods, and by which additional delays can be inserted to optimally phase-align BG output as transmitted to the thalamus. By these delays, according to the BGMS model, BG output spikes are temporally aligned to promote thalamocortical activity associated with selected connections, and inhibit competing activity. The BG then promote oscillation at the period characteristic of the activated path delay, and recognize and promote that oscillation with the various and specific nonzero inter-areal phase relationships characteristic of a given large scale functional network, by shifting back input from phase-leading areas to align in striatum with that from phase-trailing areas, then in the striatonigral/striatopallidal stage, adding more delay to paths looping back to phase-trailing areas. As noted above, simulations suggest that such control of inter-areal phases in itself controls the direction of information flow in functional networks (Palmigiano et al. 2017).
As suggested above, diverse striatopallidal delays allow for coactivated SPNs to align their inputs to jointly targeted pallidal and thalamic cells. The primacy of synchrony rather than aggregate rate in these relations likely has profound consequences: fragmentary and spurious activation of BG projection neurons is overwhelmingly likely to produce robustly incoherent and thus ineffectual inputs to the thalamus and other BG targets, assuring that only synchronized and unconflicted outputs can promote thalamocortical activations. These relations, which imply a moderate tolerance for spurious activations, may maximize the versatility of BG projection neurons by dynamic functional pluripotentiality mechanisms of the sort studied in cortex (Izhikevich 2006; Rigotti et al. 2013).
As mentioned above, the STDP of SPNs is apparently reversed: synapses are strengthened that activate in the ~20 ms after activity in other synapses has induced postsynaptic discharge, and synapses are weakened that bear activity in a similar time window preceding discharge (Fino et al. 2005). This arrangement seems to systematically maximize the delay of paths through striatum, and it may also tend to maximize the variety, and the consequent breadth of associativity, of SPN afferents. Striatal FSIs show normal STDP relations, tending to minimize delays by strengthening the synapses that bear the earliest activity correlated with discharge (Fino et al. 2008). Striatal physiology thus appears to promote dispersion, while minimizing the delay of FSIs, which—as reviewed later (§4.6)—consistently activate before SPNs in their vicinity, and precisely control the timing of SPN discharge through powerful appositions. Selective reinforcement of slow cortico-SPN paths and fast cortico-FSI paths might arrange so that SPNs are rebounding from a GABAergic cortico-FSI-SPN spike volley precisely when that same spike volley arrives at the soma (with greater dispersion) via glutamatergic cortico-SPN paths.
In rats trained on a maze task until habit formation, then given extinction training, and finally retrained on the original task, ensembles of SPNs in the dorsal striatum formed, narrowed, and changed their responses to fire synchronously at the beginning and end of the task, then reverted, and finally reestablished their synchronous responses, respectively, with a high correlation of response synchrony to behavioral performance (Barnes et al. 2005). Similarly in monkeys, over the course of self-initiated, reward-motivated learning, large numbers of neurons in the dorsal striatum developed phasic responses aligned with the beginning and end of saccade sequences (Desrochers et al. 2015). In rat ventral striatum, a shift in the patterns of phasic activity from local islands of high gamma synchrony, to beta synchrony spanning wide areas and both SPNs and FSIs, accompanies skill acquisition and habit formation (Howe et al. 2011). These studies provide strong evidence that striatal plasticity entails the formation of widely distributed constellations of FSI-SPN assemblies that learn to discharge in synchrony as a function of context.
Presumed cholinergic interneurons in the striatum, recognized electrophysiologically by their tonic firing patterns, may be key components of a time alignment learning mechanism in the BG. Over the course of skill acquisition, progressively larger proportions of these sparsely distributed interneurons, over very wide areas of striatum, have been seen to pause in brief, precise synchrony in response to salient sensory stimuli, with this response dependent on dopamine supply (Graybiel et al. 1994). In general, corticostriatal long-term potentiation (LTP) depends on the temporal coincidence of cholinergic interneuron pause, phasic dopamine activation, and SPN depolarization (Zhang et al. 2019b‡). These interneurons have been implicated in the learning of changes in instrumental contingencies, and that learning is dependent on activity in thalamostriatal projections originating in the intralaminar nuclei (Bradfield et al. 2013). Moreover, precisely synchronized stimulation of these interneurons directly induces dopamine release through cholinergic receptors on dopaminergic axons, independent of somatic activation of midbrain DA neurons (Threlfell et al. 2012), suggesting that synchronous BG output per se, as measured by activity in intralaminar afferents, is intrinsically reinforced in the striatum.
As discussed in greater detail later (§6.11), striatal matrix is extensively and preferentially targeted by inputs from intralaminar thalamus, apposing both SPNs and FSIs. Thus, mechanisms of striatal plasticity are positioned to monitor and respond to the synchronies that the BG generate in thalamus, and so presumptively in cortex. Dopamine-dependent and dopamine-inducing activity in striatal cholinergic interneurons, innervated by these thalamostriatal projections, might act to strengthen striatal synapses that contribute to the production of synchronous thalamocortical activity associated with reward. Related mechanisms may similarly drive plasticity in other BG structures targeted by the intralaminar nuclei, notably the GP and STN (Sadikot et al. 1992a).
Consistent with the proposition that pallidothalamic and nigrothalamic axons collateralize to orchestrate tightly coherent long range synchronies via the thalamus, the delay of these segments is comparatively short and uniform: in macaques, antidromic response from thalamus to SNr was found to average 1.56 ± 0.44 ms (Kitano et al. 1998), and an earlier study (Harnois and Filion 1982), on squirrel monkeys, found similar antidromic delays from thalamus to GPi, tightly clustered about an average of 1.3 ms from ventral anterior (VA) and ventral lateral (VL) sites, and 1.6 ms from centromedian (CM) sites, arising from a CV of 6 m/s. Similarly, as noted earlier (§1.5), the thalamocortical projection appears to be tuned for rapidity and exquisitely precise (sub-millisecond) alignment of the projection to any given area of cortex; selective myelination of the portion of thalamocortical axons within cerebral white matter, the length of which varies two-fold within a target area, appears to account for this (Salami et al. 2003).
In cats, the delays for antidromic stimulation of thalamocortical projections from VA-VL and ventromedial (VM) thalamic nuclei to Brodmann areas 4, 6, 8, and 5 were found to average from 2.3 ms (VA/VL to area 4, primary motor) to 4.2 ms (VM to area 8, motor association cortex), with almost all measured delays falling below 6 ms, and significant and systematic, but small, shifts in delay as a function of thalamic nuclear origin (Steriade 1995). In the intralaminar nuclei, which in the BGMS model are a crucial broadcast hub for timing information (reviewed in detail later (§7)), antidromically measured thalamocortical delay is less than 500 µs, indicating conduction velocities (CVs) of 40-50 m/s (Glenn and Steriade 1982; Steriade et al. 1993).
In songbirds, an average delay of 5.1 ± 0.48 ms, range 3.5-7.9 ms, has been measured between arrival of a pallidothalamic terminal spike and the next discharge by the targeted thalamic projection neuron (Goldberg et al. 2012). Assuming this figure is roughly representative of the figure in mammals, and given the results summarized above from Yoshida et al. (1993) and Kitano et al. (1998), the total transmission delay for a closed loop through the BG can be estimated. This delay is about 25 ms (40 Hz) for a typical path through GPi and VA/VL to primary motor, and about 50 ms (20 Hz) for a typical path through SNr to a frontal eye field in area 8, with a large range of possible delays, roughly 19-31 ms (32-53 Hz) and 35-65 ms (15-29 Hz) respectively for average ±1 standard deviation.
These relationships suggest that cortical activity routed through the BG and thalamus is typically delayed by a single cycle upon its return to cortex, but perhaps in some scenarios higher frequency oscillation might be delayed by two or more cycles, and in some scenarios there might be cross-frequency interactions, particularly in the SNr. The prominence of beta-like delays through the SNr comports with the prominence of beta-like oscillations in visual top-down control signaling, which itself involves cross-frequency interactions in visual cortex (Richter et al. 2017; Lee et al. 2013).
Because position along the distal-proximal dimension of dendritic processes introduces a graded delay, thereby altering the phase shift imparted by the inputs upon the neuron's output (Goldberg et al. 2007), fine tuning of BG path delays may be possible within the spatially extensive terminal and dendritic processes of corticostriatal (Mailly et al. 2013) and striatopallidal (Levesque and Parent 2005) neurons. Striatopallidal axons penetrate perpendicular to the dendritic disks of pallidal output neurons, emitting thin (diameter 100-200 nm), unmyelinated (hence particularly low CV) collaterals parallel to the disks, repeatedly synapsing with the same target neuron (Goldberg and Bergman 2011; Difiglia et al. 1982). The diameters and termination patterns of these thin ramifications match those of the similarly positioned parallel fibers of the cerebellar cortex, particularly in the upper molecular layer, where nearly all fibers are 100 to 250 nm in diameter (Sultan 2000).
In the substantia nigra, striatal inputs constitute a large majority of inputs, and appose dendrites at various distances from the soma, with only a small fraction apposing the soma directly, and most apposing small (distal, slow-conducting) dendrites (Bevan et al. 1994). Conduction through nigral dendrites entails delays of up to ~12 ms relative to proximally apposed inputs (which predominantly arise from GPe and STN) (Tiroshi and Goldberg 2019), which is a plausible range for phase tuning of direct path outputs, and closely matches the range of phase tuning by parallel fibers of the cerebellar cortex (Heck and Sultan 2002). Moreover, the frequency preferences of networks may respond in a nonlinear fashion to much smaller adjustments of path delays, with synchrony depending strongly on precise matching of delays (Ivanov et al. 2019‡). In short, localized selective strengthening of appositions distributed along the length of dendrites might adjust path delays and associated oscillatory frequency preferences, in natural response to reinforcement.
Axonal CV plasticity, which has only recently been appreciated (Fields 2015), is largely unknown in its mechanistic particulars, but might also operate in the BG.
Whether optimization of conduction delays is by competition between distinct fiber paths, or between distinct synapses along the same fiber path, reinforcement-driven persistent modulation of synaptic efficacy could optimize not only the output rates (the efficacy with which a particular input evokes an output), but the fine time structure of the outputs, to the degree that reinforcement is a function of fine time structure. Axonal CV plasticity might operate in conjunction with these mechanisms, responding to the same (or to coordinated) reinforcement signals. Moreover, the traversed neurons themselves may exhibit a diversity of intrinsic time constants, similar to an arrangement that has been described in PFC (Bernacchia et al. 2011). Indeed, the activation of both PFC and striatal neurons shows a finely graded diversity of delays, though path variety may be the underlying mechanism (Jin et al. 2009).
Behaviors must be precisely paced to meet contextual requirements, and the BG are clearly integral to the performance of these behaviors. Moreover, as discussed here, BG circuitry entails diverse time constants and diverse, often lengthy conduction delays. However, the complex delay mechanisms of the BG direct path appear to not be central to the mechanisms underlying precise and variable pacing of overt behavior: patterns of proportional temporal scaling in neural activity, putatively associated with pacing, are likely generated in cortex, and are much less apparent in thalamus (Wang et al. 2018).
BGMS crucially entails the coherent transmission of corticostriatal spike volleys, through BG and thalamic relays, back to cortex, with routing and delays providing for spatiotemporal coincidence with corticocortical spike volleys associated with the selected effective connections. Above is a discussion of the various mechanisms that might underlie these temporal alignments. Below, I discuss the myriad patterns of convergence and divergence in the BG that underlie the capacity of the BG to distribute spike volleys coherently, widely, flexibly, and specifically.
Divergence in the paths from corticostriatal neurons through striatal fast spiking interneurons and spiny projection neurons, pallidal and nigral projection neurons, and thalamocortical projection neurons, suggest geometric expansion of activity from a single cortical column to a scope encompassing large areas of cortex. Well over 109 cortical neurons might be influenced by the output of a single striatally projecting neuron in cortex. The axonal processes of each corticostriatal neuron distribute sparsely through large regions of striatum, spanning on average 4%, and up to 14%, of total volume, forming on average ~800 synaptic boutons, likely apposing nearly as many distinct striatal neurons (Zheng and Wilson 2002).
While more than 90% of striatal neurons are spiny projection neurons, 3-5% are fast spiking interneurons (Koós and Tepper 1999). If corticostriatal neurons innervate SPNs and FSIs with similar preference, this suggests that each innervates on average ~24 FSIs (though there are indications of specialization in corticostriatal targeting of FSIs (Ramanathan et al. 2002)). Each FSI projects to ~300 SPNs (Koós and Tepper 1999), each SPN projects to ~100 pallidal neurons (Yelnik et al. 1996; Goldberg and Bergman 2011), each pallidal neuron projects to ~250 thalamic neurons (Parent et al. 2001), and each thalamic neuron projects to more than 100 cortical neurons (Parent and Parent 2005) (likely far more (Rubio-Garrido et al. 2009)). With a cortical neuron population in lower primates of approximately 109 (Herculano-Houzel et al. 2007; Azevedo et al. 2009), these divergence ratios suggest that a single corticostriatally projecting neuron can influence all of the cortical neurons within the relevant bounds of segregation. This influence is further fortified by intrinsic mechanisms in superficial cortical layers, described later (§6.7), that horizontally spread oscillations.
Striatal activity closely reflects cortical activity (Peters et al. 2019‡). As emphasized earlier (§1.7), interconnected cortical regions systematically converge and interdigitate, even while the projection of each cortical region diverges in a spotty, widely distributed pattern (Van Hoesen et al. 1981; Selemon and Goldman-Rakic 1985; Parthasarathy et al. 1992; Flaherty and Graybiel 1994; Hintiryan et al. 2016; Hooks et al. 2018). Direct path SPNs are preferentially innervated by neurons in cortex that are reciprocally interconnected over long ranges at the single unit level (Lei et al. 2004; Morishima and Kawaguchi 2006), and projections from these “intratelencephalic” cortical neurons, with widely separated but interconnected origins, show a particular tendency to converge in striatum (Hooks et al. 2018).
Convergence of densely interconnected cortical areas to single striatal FSIs is common; a study in rats found that nearly half of FSIs innervated by primary somatosensory or primary motor cortex receive projections from both (Ramanathan et al. 2002). Moreover, FSIs show a significant preference for direct path SPNs, with functional connectivity demonstrated for roughly half of identified direct path FSI-SPN pairs, but roughly a third of indirect path pairs (Gittis et al. 2010).
As reviewed earlier (§1.3), synchronization of activity in two areas signifies that those areas are functionally connected, and asynchrony or antisynchrony signifies functional disconnection. This prompts the expectation that functionally connected areas projecting convergently to an FSI robustly entrain that FSI, which imparts their shared cortical rhythm to the SPNs it innervates. By this mechanism, effective connections might act through the BG to directly excite further connections, or to inhibit connections, as envisioned by von der Malsburg (1999). On the other hand, when the afferents to an FSI are active but unsynchronized, the FSI is likely arrhythmically activated, imparting an incoherent inhibitory spike pattern to those SPNs, thereby preventing rhythmic discharge. Indeed, as discussed in greater detail later (§6.10), individual cortical cells subject to conflicting synchronies are themselves likely to exhibit arrhythmic spiking patterns (Gómez-Laberge et al. 2016). Antisynchronized afferent activity might have similar results, activating the FSI at twice the fundamental frequency, likely imparting a spike pattern to the SPNs that is particularly efficient at inhibiting discharge. In a third mode of operation, afferents to the FSI from one area bear strong oscillatory activity, while other afferents bear significantly weaker activity that may or may not be rhythmic. In this case, the FSI might impart the strong oscillatory activity to the SPNs, while the weak afferent activity has relatively little effect on FSI spiking, so that strong localized cortical oscillation is selected for effective connection to other areas.
The physiology of striatal FSIs in normal behaving animals, and their relationships with cortical and striatal projection neurons, are complex, specialized, and nuanced (Berke 2011). The temporal structure of FSI spiking closely conforms to that of afferent activity, aligning precisely with the trough of extracellular afferent LFP, regardless of band (Sharott et al. 2009, 2012; Howe et al. 2011). The phasic activation of each FSI is strongly but idiosyncratically related to ongoing behavior, and in particular, is independent of activity in other FSIs (Berke 2008). Nearly half of corticostriatal synaptic inputs to FSIs are robust, apposing somata or proximal dendrites (Lapper et al. 1992), and corticostriatal axons commonly form several synaptic boutons targeting a single FSI, indicating selective innervation and stronger coupling (Ramanathan et al. 2002). Consistent with the observed idiosyncrasy and independence of FSI responses to cortical activity, synaptic inputs to FSI somata are few, and FSI dendrites are almost entirely devoid of spines (Kita et al. 1990)
FSI projections to SPNs are robust (Koós and Tepper 1999), but FSI activation has been found to modulate SPN activity, rather than simply inhibiting or releasing it (Gage et al. 2010). In behaving rats, FSIs and nearby SPNs are simultaneously active in various stages of task learning and performance, at precisely opposite phases, at both beta and gamma frequencies (Howe et al. 2011). Simulation of normal in vivo conditions in the striatum shows formation of small assemblies of synchronized SPNs, with FSI activation increasing the firing rates of connected SPNs (Humphries et al. 2009). Pharmacological blockade of FSIs in sensorimotor striatum does not substantially change the average firing rates of nearby SPNs, but induces severe dystonia (Gittis et al. 2011), demonstrating that FSI regulation of the temporal structure of SPN activity is crucial to normal behavior.
Tourette syndrome is associated with abnormally low density of presumed FSIs in the striatum (Kalanithi et al. 2005). Indeed, the cancellation of inapt behaviors has been associated with GABAergic feedback projections from the GPe that selectively target FSIs (Mallet et al. 2016; Deffains et al. 2016). Pathological sparseness in FSI afferents to an SPN might result in entrainment of that SPN to cortical activity that would normally be inhibited by another FSI. The resulting spurious SPN discharges, phase-locked to localized cortical activity, then might induce spurious reinforcement and connections in cortex, manifesting as tics and other compulsions.
Whereas SPNs exhibit highly heterogeneous responses to dopamine, FSIs exhibit a largely uniform, dose-dependent response to drugs that manipulate DA, reducing firing rate in response to DA antagonism, and increasing it in response to DA agonism; likewise, FSI activity is positively correlated with locomotor activity, while SPN activity shows highly variable relations (Wiltschko et al. 2010).
FSIs exhibit significantly lower firing thresholds than do SPNs relative to the intensity of cortical activity; consequently, activation of SPNs is preceded by, and spatially embedded within, an encompassing area of activated FSIs governing their output (Parthasarathy and Graybiel 1997). In a study inducing focused, synchronized activity in primary motor cortex, nearly all (88%) of the FSIs in the center of the zone of striatal activation were activated, and nearly as many (78%) of the FSIs in a penumbra were activated; FSIs showed a robust and disproportionate response, comprising 22% of the responding striatal neuron population, while representing <5% of striatal neurons (Berretta et al. 1997).
Each SPN receives inputs from several (estimated 4-27) FSIs (Koós and Tepper 1999), suggesting that FSI recruitment in a striatal neighborhood reliably imparts strong modulatory input to all of the SPNs in that neighborhood. While there is some evidence that FSI prevalence in the striatal population follows a gradient, with highest concentration in the dorsal and lateral striatum and lowest in the medial and ventral striatum (Kita et al. 1990; Bennett and Bolam 1994; Berke et al. 2004), more recent evidence demonstrates FSI effects and connectivity in VS similar to those in dorsal striatum (Taverna et al. 2007; Howe et al. 2011), and the appearance of a striatal FSI density gradient may be an artifact of spatially correlated cytological heterogeneity in the FSI population (Tepper et al. 2008).
It has been shown in awake behaving rats that the activity of FSIs in the sensorimotor striatum rises shortly before, and peaks during, initiation of behavior reflecting a decision, and that FSI activity precedes that of coactivating neurons in primary motor cortex (Gage et al. 2010). FSI activity is erratic and bursty in the resting animal, but transitions to rhythmically regular activity at the moment a cue is presented, and remains regular throughout the delay period until movement execution, at which point the erratic bursting resumes (Berke 2011; Lau et al. 2010).
FSIs are electrically woven together into a loose, sparse continuum by gap junctions (Kita et al. 1990; Koós and Tepper 1999), that in simulation modestly encourage synchronization of neighboring FSIs, while modestly damping their activity unless afferent activity is well-synchronized (Hjorth et al. 2009). Gap junctions are also thought to be crucial for the regularization of FSI activity during cued delay periods, noted above, in response to transitions of corticostriatal input from random to patterned (Berke 2011; Lau et al. 2010). These phenomena further suggest an arrangement in which FSIs operate as a matrix, comprehensively regulating the temporal structure of SPN spike activity, with particular sensitivity to synchrony in the corticostriatal projection.
Striatal FSIs contain parvalbumin, can sustain firing rates of 200 Hz with little or no adaptation, have narrow action potentials (shorter than 500 µs), do not feed back to the inputs of other FSIs, and do not receive inputs from SPNs (Koós and Tepper 1999; Mallet et al. 2005; Taverna et al. 2007). SPNs and FSIs produce similar inhibitory post-synaptic currents (Koós et al. 2004), but FSI inputs to SPNs are directed to somata and proximal dendrites, where they can exert a more decisive and precise effect on the target, whereas corticostriatal inputs to SPNs, and SPN inputs to other SPNs, are directed to distal dendrites (Bennett and Bolam 1994).
SPNs may exhibit a low pass characteristic (Stern et al. 1997), so that even while the SPNs are highly sensitive to synchrony in their excitatory afferents (discussed below), the fine timing of the spikes they produce could be determined almost entirely by the FSIs. The influence of FSIs on the SPNs they target entails not only retardation of SPN phase, but phase advancement, through a rebound effect that reduces the firing threshold of the targeted SPN; the effect is most pronounced 50-60 ms after the FSI spike; SPN depolarization is advanced by ~4 ms when FSI spikes reach the SPN 30-70 ms before excitatory afferent spiking reaches the SPN (Bracci and Panzeri 2005).
Convergence is physiologically inescapable in paths through cortex, striatum, and pallidum. There are roughly ten times as many pyramidal cells in cortex projecting to the striatum, as there are medium spiny projection neurons, with each SPN afferented by roughly 10,000 distinct cortical neurons (Kincaid et al. 1998; Zheng and Wilson 2002). The massive convergence to single SPNs, and their high firing threshold, arrange so that SPNs fire only when their afferent activity is substantial, synchronous, and distributed broadly across dendrites (Zheng and Wilson 2002). As noted earlier (§1.8), the sensitivity of the striatum to synchrony in its inputs is particularly consequential if striatal output induces synchronies (as in the BGMS model), because the striatum is then positioned to iteratively process information encoded as patterns of synchrony.
Though relatively sparse, corticostriatal projections from GABAergic interneurons to SPNs (Melzer et al. 2017), particularly from cortical FSIs, are an additional mechanism that may regulate SPN sensitivity to synchrony. As reviewed in detail later (§7.5), cortical FSIs have been shown to be part of a coincidence detection mechanism with a very narrow window (Pouille and Scanziani 2001). Depending on the pattern of apposition of these cortical FSIs, they might also function like striatal FSIs, precisely controlling the timing of SPN output, as described above.
Corticostriatal neurons seldom fire periodically, but rather, their activity is aperiodic but phase-locked to oscillation in their cell membranes (Stern et al. 1997); thus their converged input to SPNs can exhibit substantial periodicity, but only when cortical activity is robust and synchronized. The aperiodicity, low spontaneous rates, and narrowly discriminative activity of individual corticostriatal neurons, suggest that few SPNs will be active at a given moment, and many will be silent (Turner and DeLong 2000). In the awake, resting animal, a large majority of SPNs are silent (Sandstrom and Rebec 2003), and some SPNs remain silent even in the awake, behaving animal, with no apparent physiological distinctions to explain the silence (Mahon et al. 2006).
SPNs exhibit no persistent or membrane-intrinsic frequency preference; rather, the aggregate intensity of afferent activity (simulated in vitro by injection of constant current) establishes a firing rate, and the phase of that firing preferentially follows that of afferent components at frequencies near the established rate, with particularly sharp frequency selectivity at beta frequencies (Beatty et al. 2015). This signifies that, as the aggregate afferent activity to an SPN increases to and beyond the firing threshold, the phase of that firing will be preferentially determined by progressively higher-frequency synchronized components of that afferent activity. The functional significance of this arrangement is unknown, but might be relevant to scenarios in which an SPN is not subject to regulation by FSIs (if such scenarios occur at all in the normal striatum, which seems doubtful), and in any case seems to be a virtually inevitable biophysical dynamic.
During slow wave sleep and drowsiness, SPNs fluctuate between “Up” and “Down” states, characterized by subthreshold depolarization and hyperpolarization respectively, and they only discharge when in the Up state (Wilson and Kawaguchi 1996; Wilson 1993; Mahon et al. 2006; Kitano et al. 2002). Impulsive dendritic bombardment, largely that arising from the corticostriatal population, pushes SPNs to the Up state, and quiescence in those inputs returns SPNs to the Down state (Kasanetz et al. 2006). Because SPNs can only discharge when in the Up state, patterns of SPN activation reflect the initial pattern of input bombardment, with the associated population of Up SPNs responding to their cortical inputs as long as they are Up, while Down SPNs are silent (Kasanetz et al. 2006). With this arrangement, trans-striatal routes can be activated by a single strong impulse, then be held open by weaker activity originating in the same corticostriatal population, passing rhythmic energy from those inputs, sculpted by striatal FSIs, to BG output structures.
However, the functional significance of this arrangement is unclear, because Up/Down bimodality seems to be characteristic of drowsiness, slow wave sleep, and anesthesia, and not of the awake state (Mahon et al. 2006). Indeed, evidence that striatal activity in the awake state continually tracks context (Arcizet and Krauzlis 2018; Weglage et al. 2020‡) is consistent with an arrangement in which SPNs are tonically Up in the awake state.
Curiously, the Purkinje cells of the cerebellum also exhibit Up and Down states, with transitions triggered by impulsive input currents, and discharge only in the Up state (Loewenstein et al. 2005). And like striatal SPNs, Purkinje cells receive enormously convergent inputs, respond only when that input is synchronized (Sultan and Heck 2003), and are key to a BGMS-like mechanism (Popa et al. 2013; Liu et al. 2020‡; McAfee et al. 2019).
SPNs exhibit uncorrelated activity even when they are immediate neighbors, suggested to be due to an arrangement in which each SPN receives axons from a unique, sparse subset of corticostriatal neurons (Kincaid et al. 1998; Zheng and Wilson 2002; Wilson 2013). This also follows from the firing patterns of direct path corticostriatal neurons, which are highly idiosyncratic (Turner and DeLong 2000). SPN axon collaterals synapse upon the distal dendrites of other SPNs, with each SPN inhibited by up to 500 other SPNs, but these inputs are sparse, weak, unreciprocated, and asynchronous, and do not induce correlated activity among neighborhoods of interconnected SPNs (Tepper et al. 2008; Wilson 2013). Instead, their location suggests they interact with excitatory afferents, enhancing the combinatorial power of the corticostriatal and thalamostriatal projection systems.
Convergence in the projections from striatum to the pallidal segments and SNr is inevitable given the further reduction in volume and cell count. In human, volume ratios from the striatum are 12:1 to the GPe, 21:1 to the GPi, 24:1 to the SNr, and 6:1 from the striatum to GPe, GPi, and SNr combined (Yelnik 2002), with cell count ratio estimates of 97:1, 400:1, and 210:1, respectively, for a combined ratio of 57:1 overall, and 138:1 in the direct path (Kreczmanski et al. 2007; Hardman et al. 2002). In rats, the volume ratios are 7:1, 112:1, and 19:1, for GPe, GPi (EP), and SNr, respectively, for a combined ratio of 5:1, and the cell count ratio estimates are 61:1, 880:1, and 106:1, respectively, for a combined ratio of 37:1 (Oorschot 1996).
Each pallidal projection neuron forms a large 1.5 mm2 dendritic disk perpendicular to incident striatopallidal axons, innervated by 3,000-10,000 SPNs (Yelnik et al. 1984; Goldberg and Bergman 2011). The dendritic processes of projection neurons in the SNr have variable forms, with an extent similar to that of pallidal dendrites, resulting in similar convergent innervation by SPNs (François et al. 1987). These arrangements imply a notional convergence ratio from corticostriatal neurons, to SPNs, to pallidal projection neurons, as high as 108. There is only slight convergence in the pallidothalamic projection, but extensive convergence in the thalamocortical projection (Rubio-Garrido et al. 2009) suggests a notional convergence ratio substantially greater than 109 for the full loop back to cortex.
The striatopallidal projection, like the corticostriatal projection, also entails divergence. Each SPN axon forms 200-300 synapses, sparsely distributed through a large volume of pallidum, with 1-10 synapses formed with a given pallidal dendrite (Yelnik et al. 1996; Goldberg and Bergman 2011), implying that an SPN projects to 20-300 pallidal neurons. Moreover, with a tracer injection in the striatum encompassing a small cell population, a hundredfold increase is seen in the volume of pallidum labeled by the tracer, while larger injections increase the density, but not the volume, of the labeled area (Yelnik et al. 1996), confirming an arrangement of simultaneous, extensive divergence and convergence like that of the corticostriatal projection.
Experiments in primates show that the projection from striatum to the GPi entails reconvergence, such that divergence in the projection from a cortical locus to multiple loci in the striatum is followed by convergence from those striatal loci to a single pallidal locus (Flaherty and Graybiel 1994). Graybiel (1998) suggested that the striatum thus acts as a dynamically configurable hidden layer. The BGMS model further proposes that this arrangement subserves selection and activation of effective connections in cortex. The path from a cortical locus, by diverging to many distinct FSI neighborhoods, then reconverging to a single pallidal locus, can be subjected to any of a variety of spike timings, representing a variety of candidate effective connections, while suppressing action through that pallidal locus when multiple SPNs impart conflicting activity upon it, as suggested above.
Experiments in vitro demonstrate that striatal afferents to GP can control the precise timing of firing by the targeted cells, and that these cells can follow striatal oscillatory inputs up to the gamma range (Rav-Acha et al. 2005; Stanford 2002).
The relationship of FSIs to SPNs may help elucidate the role proposed for the BG in competitive selection (Redgrave et al. 1999). An output neuron in GPi or SNr bombarded by mutually incoherent SPNs would be incapable of imparting a coherent temporal pattern to its thalamic targets. If coherence in this path is crucial, as suggested by the BGMS model, then only those GPi/SNr neurons with predominantly coherent activity in their afferents can participate in activation of a motor or cognitive connection, while activation of clashing connections tends to be suppressed.
GPi activity increases when direct path SPNs are activated, and decreases when indirect path SPNs are activated, even while direct and indirect path activation are associated with widespread cortical activity increases and decreases, respectively (Lee et al. 2016). As repeatedly emphasized in this paper, this suggests that oscillatory modulation and facilitation by spike-timing-dependent gain is among the core functions of the GABAergic output neurons of the BG direct path.
As discussed at greater length later (§5.2), spike timing within, and activation of, pallidal output neurons is almost completely independent in the healthy brain (Nini et al. 1995; Nevet et al. 2007), so any coordination or synchronization must be sparsely distributed. Preliminary results from a study of functionally connected cortical, pallidal, and thalamic areas, demonstrates as much, with strong phasic correlation of LFPs but almost no correlation of individual neuron spiking with those LFPs (Schwab 2016, chapter 5). According to the BGMS model, broadly and densely synchronized BG output would produce spurious and pathologically persistent effective connectivity in cortex, arresting task progress. Indeed, the independence of BG output neurons breaks down in parkinsonism, in which task progress is retarded or arrested (Stanford 2002; Wilson 2013; Deister et al. 2013; Hammond et al. 2007; Nevet et al. 2007).
PD has been shown in magnetoencephalography (MEG) studies to be associated with progressively greater functional connectivity in cortex, determined by measures of synchronization likelihood, both intra- and inter-areal, in both the alpha and, in moderate and advanced disease, the beta bands (Stoffers et al. 2008). It is associated with global disruption and progressive inefficiency of functional connectivity (Olde Dubbelink et al. 2014), and with the gradual emergence of psychosis as the disease progresses (ffytche et al. 2017). MEG evidence further suggests that a key consequence of pathological synchrony in parkinsonism is a contracted repertoire of functional constellations, progressively associated with deficits of cognitive and behavioral flexibility (Sorrentino et al. 2019‡). Pathological synchrony in parkinsonism may be rooted in the striatum: in vitro experimentation with, and physiologically realistic simulation of, the dopamine-depleted striatum has demonstrated spontaneous pervasive formation of clusters of synchronized SPNs (Humphries et al. 2009).
The main BG output projections from GPi and SNr appose neurons in the motor and association thalamus in giant inhibitory terminals, with multiple synapses, exerting powerful and precise inhibitory control of individually targeted cells, with GPi and SNr projections well-compartmented from each other (Bodor et al. 2008). Many of these projection neurons contain parvalbumin, with especially high density in the dorsal GP (Cote et al. 1991).
Pallidal axons branch extensively within thalamus, into 10-15 collaterals with highly confined terminal varicosities (Parent et al. 2000), so that each pallidal neuron projects to 200-300 neurons in thalamus; these are the most widely arborized neurons in the BG (Parent et al. 2001). The somata and primary dendrites of GPi- and SNr-recipient thalamocortical neurons outside the intralaminar nuclei are contacted almost exclusively by these afferents, with very dense terminal processes, suggesting that the GPi and SNr exercise predominant control over activity in these cells (Kultas-Ilinsky and Ilinsky 1990; Ilinsky et al. 1997). As discussed in detail later (§6.3), BG projections to the intralaminar thalamus do not follow this pattern, but instead predominantly appose small and medium dendrites (Sidibé et al. 2002).
In the pallidothalamic projection, a degree of convergence on single thalamocortical cells has been noted (Ilinsky et al. 1997). As discussed at greater length later (§5.2), convergence of independently oscillating pallidal projection neurons produces aggregate input that bears the characteristics of Gaussian noise, so that pallidal output can reflect inputs to pallidum with high fidelity. Perhaps this is a crucial advantage that evolutionarily stabilizes this arrangement in mammals, which have a surficial cerebral cortex and closely aligned thalamocortical conduction delays (Salami et al. 2003; Steriade 1995), while birds lack a surficial pallium, and their thalamic projection cells receive only a single calyceal BG input (Luo and Perkel 1999).
Given the high typical tonic discharge rate of pallidal and nigral neurons, thorough disinhibition or entrainment of targeted thalamocortical neurons requires coordination of multiple pallidal projection neurons. As with the convergence of several corticostriatal neurons on a single FSI, several FSIs on a single SPN, or several SPNs on a single pallidal projection neuron, this suggests several activation scenarios. If all BG output neurons targeting a thalamocortical projection neuron are phasically silenced coincidentally, their common target would likely be activated directly by rebound, and thereafter would tend to be activated synchronous with corticothalamic input. If instead those several inputs remain active, but are phasically synchronized, this would likely entrain their common target, even in the absence of corticothalamic input, and with great vigor in the presence of excitatory input exhibiting the favored frequency and phase. In a third scenario, some of the inputs are phasically silenced, coincident with others that are phasically synchronized, with the likely result that their common target is entrained by the active BG inputs. Until one of these coordinated arrangements is learned, modulation of one or several GABAergic afferents likely has lesser but significant post-synaptic effects, which can bootstrap learning in paths associated with the other afferents.
The dynamic in the intralaminar thalamus is somewhat different: because pallido- and nigrothalamic terminals there predominantly appose dendrites, their likely effect is to modulate receptivity to particular corticothalamic inputs with which they share a dendrite, so that only those inputs with the selected timing (frequency and phase) can affect the soma and, consequently, affect the somatically targeted L3 and L5 pyramidal neurons. As explored later (§7.6), this arrangement may maximize combinatorial power in the relationship of the BG to the intralaminar thalamus.
The various BG activation scenarios and pathways can be viewed as biasing the probabilities that various selected functional activations and connections will be established or continued. When BG input to a thalamic area is thoroughly synchronized, that biasing is strong, whereas partial synchrony produces weaker biasing. If an SPN activation is a vote for the establishment or continuation of some family of functional connections and activations, then the GPi/SNr and the BG-recipient thalamus are positioned to tally those votes, due to convergence. Thus, cortical connectivity decisions that are broadly and consistently supported within the striatum (particularly, habits) are strongly biased, while narrowly supported decisions weakly bias connectivity, and conflicts are akin to mutual vote cancellation. Breadth and consistency of support can be viewed as representative of decision confidence. And because of divergence in the BG, activation of an SPN can simultaneously contribute to strong biasing of some connections, but weak biasing of others.
The path from BG-recipient thalamus back to cortex exhibits striking divergence. As reviewed later (§6.4), neurons of the rat VL, VA, and VM nuclei project profusely and with massive overlap to L1, with individual neurons collateralizing to widely separated areas (Rubio-Garrido et al. 2009; Kuramoto et al. 2009), and the intralaminar nuclei in all common laboratory mammals innervate nearly the entire cerebral cortex (Van der Werf et al. 2002; Scannell 1999). Centromedian and parafascicular (PF) axons that reach cortex arborize diffusely and widely, with a single axon from CM forming on average over 800 synaptic boutons in cortex, a count that may miss many poorly stained axonal processes in L1 (Parent and Parent 2005). Even considering the possibility of extensive multiple terminations on the same target neuron, which is relatively unlikely in a diffuse projection, the number of cortical neurons innervated by a single BG-recipient thalamocortical neuron might be conservatively estimated at >100. As estimated above, the cumulative notional divergence ratio from a single corticostriatal neuron, through striatal FSIs, SPNs, pallidal projection neurons, and thalamocortical neurons, may exceed 109.
According to the BGMS model, the information represented by a pattern of activation in a particular cortical area passes to other receptive cortical areas chiefly via direct and indirect corticocortical connections between them. The information reaches the thalamus via the BG only incidentally, and in drastically reduced and fragmentary form. While the striatum is continually supplied with inputs that span the entire cortex (Parent and Hazrati 1995a; Hintiryan et al. 2016; Peters et al. 2019‡; Grandjean et al. 2017), only a small fraction of the information borne by these inputs can emerge from the BG, due to the >1000:1 reduction in neuron count from the corticostriatal population to the output neuron populations in the GPi and SNr (Kincaid et al. 1998; Zheng and Wilson 2002; Goldberg and Bergman 2011).
By a similar rationale, noting a 100:1 ratio of visual cortex neurons to pulvinar neurons in macaque, Van Essen (2005) suggested that the associative thalamus itself generally operates in a modulatory role, managing information transfers that are fundamentally corticocortical. This proposition is further supported by results, noted earlier (§1.6), suggesting that the thalamic mediodorsal nucleus regulates functional connectivity in PFC rather than acting as an information relay (Schmitt et al. 2017).
Due to the general irregularity and independence of firing patterns in individual BG projection cells, the entropy of the BG paths is substantial (Wilson 2013), suggesting that decisions represented by BG output are highly flexible and can be quite nuanced. The neurons projecting from the BG to the thalamus are noted for their continual and independent high frequency discharge patterns (Brown et al. 2001; Stanford 2002), and this activity must be functionally crucial, given its inherent metabolic burden, simultaneous with remarkable evolutionary stability, spanning hundreds of millions of years and all known vertebrate taxa (Stephenson-Jones et al. 2012). It has been suggested that these signals are particularly suited to act as carriers for motor commands (Brown et al. 2001); in the BGMS model these signals act as carriers for control signals spanning all domains. And because thalamocortical neurons are themselves tonically active during waking and paradoxical (REM) sleep, with some (e.g. in the rostral intralaminar nuclei) capable of following high frequency (100-300 Hz) spike volleys (Steriade and Llinás 1988; Glenn and Steriade 1982), the transthalamic BG influence on cortical activity is continual.
Because each BG output neuron tonically oscillates at an independent frequency, aggregate tonic BG output statistically resembles Gaussian noise, suggesting that oscillatory modulation (by inputs from the striatum, in particular) can produce output signals with high oscillatory fidelity. This is akin to dithering techniques that use additive noise with a triangular probability distribution to reduce waveform distortion, in systems that represent intrinsically continuous signals using quantized digital schemes (Lipshitz 1992). Convergence in mammals of several BG output neurons to single thalamocortical neurons (Ilinsky et al. 1997), and the multitude of thalamocortical neurons innervating each neighborhood in cortex (Rubio-Garrido et al. 2009), comport with such an arrangement. In the BGMS model, the resulting high temporal resolution lets the BG produce aggregate thalamocortical activity that is precisely coincident with converging corticocortical activity. Indeed, computational simulations suggest that convergence and mutual independence in BG output neurons are indispensable for precisely timed BG-induced spike generation in the thalamus, and for the avoidance of spurious spiking (Nejad et al. 2018‡).
These arrangements are closely related to the “stochastic resonance” mechanism suggested by simulations, whereby neural network sensitivity and waveform fidelity may be enhanced by the pervasive injection of noise tuned to effect network criticality (McDonnell and Ward 2011; Vázquez-Rodríguez et al. 2017; Krauss et al. 2019). There is evidence that oscillatory waveforms in brains are often non-sinusoidal, conforming to various source-specific stereotypes (Cole and Voytek 2017); functional significance has been ascribed to the fine time structure of spike “packets” exhibiting source-specific stereotypes over time spans of 50-200 ms (Luczak et al. 2015), and in general, to the information-carrying capacity of dynamic variations in inter-spike intervals (Li and Tsien 2017). To the degree that waveform harmonics and the fine time structure of spiking are functionally significant, waveform fidelity is likewise significant.
Stereotyped non-sinusoidality in cortical oscillatory waveforms, such as the sawtooth waveforms of motor cortical beta oscillations (Cole and Voytek 2017), may facilitate the learning and production of sharply time-coincident spike volleys in striatum (discussed in detail earlier (§3.4)). But beyond the facilitation of tightly synchronized spike volleys, waveform structure on short timescales might be exploited by the striatum to selectively filter inputs, because the diverse delays of the corticostriatal projection (Yoshida et al. 1993; Kitano et al. 1998), to which synchronized and converging cortical inputs are subject, in concert with plastic variations in corticostriatal synaptic efficacy, might realize finite impulse response (FIR) filters, engendering preferences that favor some waveforms while disfavoring others. Because of similar arrangements in the cerebellum (Heck and Sultan 2002), it too might realize FIR filters with associated selectivities.
In neocortex, individual neurons in a state of wakefulness exhibit almost completely random discharge patterns (Softky and Koch 1993; Stiefel et al. 2013). Computational modeling suggests that top-down synchronizing influences on a population of cortical neurons (of the sort exerted by thalamocortical projections, reviewed in detail later (§6)) profoundly impact their aggregate oscillation, evident in the LFP, with highly selective attentional effects, even while individual cells within the population continue to exhibit nearly Poissonian random firing patterns (Ardid et al. 2010). Indeed, simulations suggest that the stochasticity and brevity of synchronies characteristic of biological neural networks result in particularly effective modulation of information flow among the synchronized areas; even brief episodes of synchrony, lasting only a few cycles, may suffice for efficient information transfer, with directionality from phase-leading to phase-lagging areas (Palmigiano et al. 2017). More generally, as explored in some detail later (§11.6), noisiness in the brain may crucially aid problem solving.
Behaviorally consequential aggregate oscillatory synchronies, in the absence of significant correlations in the spiking activity of the individual contributing neurons, are apparent in the relationship of the BG to the thalamus. In recent experiments with monkeys, it was found that movement-related LFP oscillations in GPi and its target area in thalamus (ventral lateral, anterior part, VLa) were strongly and likely causally correlated, for the duration of each trial, with a time lag from GPi to thalamus shorter than 10 ms, even while individual neuronal firing patterns in GPi showed little correlation to GPi LFP, and virtually no correlation to LFP in thalamus (Schwab 2016, chapter 5). These results suggest that the neurons discharging synchronously are sparsely embedded within a much larger number of neurons whose discharges are not correlated, or that the LFP synchrony is due to a coherent but weak influence on large numbers of those neurons, or some combination. While this is expected from the known physiology of the GPi, discussed above, and at greater length earlier (§4.15), it is doubtless methodologically frustrating.
In any case, because the BG form closed loops with cortex and with themselves, they are well-positioned to select and reinforce, or indeed generate and sustain, large scale aggregate oscillations. It is suggestive that even without tunable delays, artificial recurrent neural networks can learn to oscillate at various specific frequencies as a precise function of non-oscillatory input patterns (Sussillo and Barak 2013). And as explored earlier (§3), the immense diversity of BG path delays suggests that closed-loop circuits through the BG can be readily tuned to prefer particular frequencies. This in itself might be an important selection mechanism.
A variety of evidence suggests that the BG process and preserve oscillatory time structure, positioning them to manipulate cortical synchronies. For some time it has been appreciated that cortico-BG circuits in a state of health show synchronized oscillations across the full spectrum of power bands, from the “ultra-slow” (0.05 Hz) to the “ultra-fast” (300 Hz), with robust oscillatory activity in the striatum and STN of alert behaving animals (primate and rodent) that is modulated by behavioral tasks (Boraud et al. 2005). In PD patients treated with levodopa, STN oscillation in the high gamma band, starting immediately before and accompanying movement, appears to entrain cortex, with the BG leading cortex by 20 ms (Williams et al. 2002; Litvak et al. 2012). In normal monkeys, task-related beta band oscillations in PFC follow and, according to Granger analysis, are caused by, activity in the striatum; this striatal activity, and that of its targets, sustain a spatially focused phase lock, with no inter-areal delay at beta (Antzoulatos and Miller 2014).
BG output responds quickly to sensory stimuli, accompanies and is sustained during delays, and precedes behavioral responses (Nambu et al. 1990). The striatum synchronizes with cortical theta (Berke et al. 2004) and gamma (Jenkinson et al. 2013; Berke 2009) oscillation, and populations of neurons within each of the successive and parallel nuclei of the BG can synchronize with cortical beta oscillation, each nucleus exhibiting a task-related characteristic phase relationship with cortical oscillation that becomes consistent and precise with task mastery, and is most pronounced at the moment of task-critical decision (Leventhal et al. 2012). Moreover, BG beta synchrony with cortical oscillation associated with a task-relevant sensory cue is established with an entraining phase reset that is sharp and immediate, within tens of milliseconds following presentation of an auditory stimulus (Leventhal et al. 2012).
In cortex, GABAergic fast spiking inhibitory interneurons (FSIs) play a dominant role in the induction and control of oscillatory activity in the beta and gamma bands, exerting fine control over phase (Hasenstaub et al. 2005). Projections from the BG-recipient thalamus to these cortical FSIs (Delevich et al. 2015; Rikhye et al. 2018; Peyrache et al. 2011) provide a path implicating the BG directly in these dynamics. Similarly in thalamus, GABAergic projections from the reticular nucleus (TRN) are believed to be crucial to the induction of the intense, globally synchronized spike bursts known as sleep spindles (Contreras et al. 1997), and extensive BG inputs spanning the TRN (Hazrati and Parent 1991; Shammah-Lagnado et al. 1996; Antal et al. 2014) implicate the BG directly in TRN regulatory mechanisms.
GABA, classically viewed as an inhibitory neurotransmitter, has a biphasic excitatory effect in certain circumstances, as a function both of the intensity of GABAergic release, and of the timing relationship between that release and the post-synaptic activity with which it interacts; GABA activity can either inhibit or enhance NMDA-dependent synaptic plasticity as a function of that timing relationship (Staley et al. 1995; Lambert and Grover 1995). Consistent with these in vitro and in vivo results, computer simulation suggests that synchronized rhythmic activity in cortical FSIs can substantially raise the sensitivity or gain of their pyramidal targets, even to constant (non-rhythmic) current injections (Tiesinga et al. 2004).
There is evidence of some of these effects, particularly biphasic activation and entrainment, in the GABAergic innervation of the thalamus by the BG (Goldberg et al. 2013; Bodor et al. 2008; Kim et al. 2017). These effects are particularly accessible to experimental probing in songbirds, where BG-recipient neurons in thalamus exhibit physiological similarity to mammalian thalamocortical cells, but unlike mammals, each receives only a single pallidal/nigral fiber, terminating in a calyx enveloping the soma (Luo and Perkel 1999). Studies in songbird thalamus have found coherent oscillatory entrainment at pallidothalamic terminals, and synchronous post-synaptic oscillation driven by pallidal input in the absence of excitatory presynaptic input (Person and Perkel 2005; Doupe et al. 2005; Leblois et al. 2009).
Simultaneous phasic intensification of ostensibly inhibitory pallidal and nigral output and activity in their thalamic targets has also been noted (Goldberg et al. 2013; Lee et al. 2016; Guo et al. 2017 “Extended Data Figure 10”). This has several possible explanations, among which are the effects described above, and the actions of dopaminergic, cholinergic, and serotonergic nuclei, which facilitate responsive oscillation, and are integral to BG circuitry (these paths and effects are reviewed later (§9)). It may also be explained by coexpression of excitatory neurotransmitters in the pallidothalamic projection, or indeed within the terminal processes of individual axons therein, which could be particularly effective at entraining a target. Indeed, several studies have found a glutamatergic component within the pallidothalamic and nigrothalamic projections (Kha et al. 2000, 2001; Conte-Perales et al. 2011; Yamaguchi et al. 2013; Antal et al. 2014).
The tonic level of activity in BG-recipient thalamus is similar to that in cerebellum-recipient thalamus, even though the latter is subject to tonic excitatory input, and the two compartments show no apparent distinctions in cholinergic or TRN innervation (Nakamura et al. 2014). This apparent paradox may be explained not only by the effects described above, but by systematic cytological preferences, in which the BG and cerebellum target cytologically distinct thalamic populations, with distinct physiology and connectivity (Kuramoto et al. 2009; Jones 2001). However, it seems clear that much of the explanation is in the nature of the BG input itself, given findings noted earlier (§2.2), that no excess of movement follows from PD treatments in which BG inputs to thalamus are removed (Brown and Eusebio 2008; Marsden and Obeso 1994; Kim et al. 2017).
Notably, just as thalamic activity increases simultaneous with increases in GPi activity, GPi metabolism and spiking activity increase simultaneous with activation of the direct path spiny projection neurons (SPNs) in the striatum that target it (Lee et al. 2016; Phillips et al. 2020‡), despite similar ostensibly inhibitory chemistry in the striatopallidal projection. Moreover, physiologically realistic modeling suggests that striatal FSI activation, ostensibly inhibiting connected SPNs, increases the firing rates of those SPNs (Humphries et al. 2009). These are the relationships needed for oscillatory relay through the successive stages of the BG, and are incompatible with models in which BG actions are limited to inhibition and release. Remarkably, even in the striosomal path through the striatum to the dopaminergic centers of the ventral midbrain, there is evidence that GABA acts by a non-inhibitory mechanism, with striosomes preferentially encoding reward-predictive cues (as does dopamine, discussed later (§9.2)) (Bloem et al. 2017).
In the BGMS model, it is through the direct path that signals arising in cortex are focused, selected, and coherently distributed, establishing long range connections implicating the areas originating the selected activity. The circuitry and scope of the direct path are thus of paramount importance in the model.
Large areas of the motor, limbic, association, and intralaminar thalamus are BG-recipient (Haber and Calzavara 2009; Groenewegen and Berendse 1994; Sidibé et al. 2002; Smith et al. 2004), projecting to feedback-recipient layers in cortex, particularly L1, L3, L5, and L6 (Clascá et al. 2012; Markov and Kennedy 2013). The primate pulvinar extensively influences visual cortex (Saalmann et al. 2012, 2018‡; Fiebelkorn et al. 2019), and is directly influenced by the superior colliculus (Stepniewska et al. 2000; Wurtz et al. 2005), itself systematically BG-recipient (Hikosaka and Wurtz 1983; Parent and Hazrati 1995a).
Direct path targets include the near entirety of frontal cortex, with some variation by phylogeny, via thalamic mediodorsal (MD), ventral anterior, ventral lateral, ventromedial, and ventral posterolateral pars oralis (VPLo) nuclei (Middleton and Strick 2002; Sidibé et al. 1997; Haber and Calzavara 2009; Sakai et al. 1996; Herkenham 1979). In primates, agranular insular and anterior cingulate cortex are targeted via MD (Ray and Price 1993), and additional targets include high-order visuocognitive areas in inferotemporal cortex via the magnocellular part of VA (VAmc) (Middleton and Strick 1996), and anterior intraparietal (Clower et al. 2005) cortex. These latter two areas are at the end of the ventral and dorsal visual streams, respectively, proposed to be associated with perceptual cognition relating to physical objects (Baizer et al. 1991), and have been found to show choice-predictive beta band synchronization in a 3D-shape discrimination task (Verhoef et al. 2011), plausibly implicating the BG directly (Leventhal et al. 2012). In chimpanzee, limited experiments have demonstrated projections from the MD, VA, and VL nuclei to posterior cortical areas 19 and 39 (Tigges et al. 1983). In humans, the dorsal and ventral visual streams in these areas are densely interconnected (Takemura et al. 2016), suggesting a role for BG-mediated inter-stream effective connectivity control in the middle stages of visual processing.
Extensive but diffuse projections from BG-recipient intralaminar nuclei (centromedian, parafascicular, paracentral (PC), and central lateral (CL)), discussed in detail later (§7), have been found to reach nearly the entire cerebral cortex, and furthermore project intrathalamically and to the basal forebrain (Kaufman and Rosenquist 1985b; Scannell 1999; Van der Werf et al. 2002).
Consistent with these widespread projections of BG-recipient thalamus, artificial stimulation of direct path SPNs in the striatum significantly and consistently increases activity throughout the entire cortex (Lee et al. 2016).
The laminar functional architecture of cortex, and the layer-specific targeting of corticocortical projections, can help explain the functional relationship of the BG to cortex. An arrangement of cortical areas in primates has been described in which hierarchies are defined by long range feedforward and feedback relations, with characteristic laminar origins and destinations that grow more distinct as hierarchical distance grows (Barone et al. 2000). In primates, feedforward projections tend to arise from L5 and deep L3, adjacent to L4 (the middle granular layer), and project to L4, while feedback projections arise from L6 and upper L3, and project to L1/L2, upper L3, and L6 (Barone et al. 2000; Markov and Kennedy 2013).
The activity borne by hierarchical projections tends to conform to stereotyped oscillatory frequency bands, with gamma and theta in the feedforward direction and beta in the feedback direction; feedback activity can enhance feedforward activity (Bastos et al. 2015a; Bressler and Richter 2015; Richter et al. 2017; Lee et al. 2013). The functions ascribed to feedback projections include attentional orientation by biasing competition, and disambiguation and hypothesis-driven interpretation of high resolution feedforward inputs (“biasing inference”), while feedforward projections introduce environmental state information into hypothetical representations (models), promoting rectification of their inaccuracies and inadequacies (Markov and Kennedy 2013).
Recent evidence and analysis indicates that activity in the alpha and beta bands is more prominent in deep cortical layers (L5/L6) than is gamma band activity, and mediates top-down control and inhibitory mechanisms, while the gamma band is more prominent in superficial layers (L2/L3), and underlies bottom-up inputs and the local maintenance of working memory representations (Bastos et al. 2018; Lundqvist et al. 2018a; Miller et al. 2018). As discussed in greater detail later (§7.5), the prominence of lower band activity in deep layers may be significant for inputs to those layers from intralaminar thalamic nuclei, which appear arranged for temporally precise but spatially diffuse distribution of spike volleys. Such inputs are likely to be particularly selective when interacting with lower frequency activity. There is, however, also evidence from human stereo-electroencephalography that long range phase coupling at low frequencies (<30 Hz) is preferentially localized to superficial layers, while very high gamma activity (> 100 Hz) correlations are localized to deep layers (Arnulfo et al. 2018‡), suggesting that temporal precision in thalamocortical projections to deep layers may be crucial for these paths to match phases into the high gamma frequency range.
Cortical inputs to striatal matrix, including the direct and indirect paths, arise from L3 and L5 (Gerfen 1989; Kincaid and Wilson 1996; Reiner et al. 2003), like corticocortical feedforward paths. This suggests that the BG receive detailed, highly specific, environmentally and contextually informative inputs. However, BG-recipient thalamocortical axons terminate chiefly in L1, L3, deep L5, and L6, and completely avoid L4 (Kuramoto et al. 2009; Jinnai et al. 1987; Parent and Parent 2005; Kaufman and Rosenquist 1985a; Berendse and Groenewegen 1991). Thus, the termination pattern of BG-recipient thalamus in cortex is like that of corticocortical feedback paths, consistent with the putative role of the BG as a modulator and selection mechanism, and suggests a key role in hypothetical modeling of the environment. Evidence suggests that neurons in L5 have broad receptive fields, while narrower receptive fields predominate in L2/L3, providing for extensive combinatorial coverage of stimulus dimensions (Xie et al. 2016; Li et al. 2016). These patterns suggest that disambiguating inputs are particularly useful in L5.
The cortical input to BG-recipient thalamus arises both from L6, and from collaterals of motor output to the brainstem and spinal cord arising from L5 (Deschênes et al. 1994), pooling afferents whose origins resemble those of corticocortical feedback and feedforward projections. This might have consequences for interactions with BG output, discussed later (§6.10). Also discussed later (§6.11), projections from the thalamus back to striatum relay this pooled input, which intermingles convergently with the more plentiful cortical inputs.
Curiously, the deepest pyramidal neurons of the neocortex, those of layer 6b, are mostly driven by long range intracortical projections originating in layers 5 and 6a, and are almost devoid of thalamocortical (and therefore of BG) inputs, even while some of them strongly target the thalamus, through which they may form large scale thalamo-cortico-cortical circuits (Zolnik et al. 2020). It nonetheless seems likely that these neurons are targeted by diffuse intralaminar thalamic projections (Parent and Parent 2005), subjecting them to BG influence by spike-timing-dependent gain.
Some projections of the PFC to posterior areas have been found to resemble those of the BG-recipient thalamus, terminating most densely in L1 and avoiding L4, in the pattern of feedback projections (Selemon and Goldman-Rakic 1988). Transthalamic paths from cortex through “core” neurons in the thalamus, on the other hand, have been found to originate chiefly in L5, and to terminate in L4 and deep L3 (Jones 2001; Rouiller and Welker 2000), like corticocortical feedforward paths. Similarly, input to the cerebellum from the neocortex arises from deep L5 (Glickstein et al. 1985; Schmahmann and Pandya 1997), and its output targets the core population in thalamus and, through them, middle layers in cortex (Kuramoto et al. 2009; García-Cabezas and Barbas 2014), so that many paths through the cerebellum also resemble corticocortical feedforward paths.
Movement is difficult or impossible to evoke by electrical stimulation of BG-recipient loci in the motor thalamus, even while such movement can be readily evoked from nearby loci receiving cerebellar output (Vitek et al. 1996; Buford et al. 1996; Nambu 2008; but see Kim et al. 2017). As noted above, cerebellum-recipient cells project mainly to middle cortical layers, while BG-recipient cells project mainly to deep and superficial layers. In primate, BG-recipient neurons in the motor thalamus may be consistently within the calbindin-positive population, associated with widely distributed and divergent cortical modulation, while cerebellum-recipient neurons are consistently within the parvalbumin-positive population, associated with specific and narrowly circumscribed topographic projections (Jones 2001; Kuramoto et al. 2009).
While movement can be evoked by microstimulation of the intralaminar nuclei (Schlag et al. 1974), BG input there, as in the nigrotectal projection, is mostly directed to the dendrites of projection neurons (Sidibé et al. 2002; Behan et al. 1987), where it can only modulate other, excitatory inputs. This contrasts with calyx-like BG terminals in the motor thalamus, that exercise predominant control over their targets, and can directly induce rebound firing and bursting (Bodor et al. 2008; Kim et al. 2017).
Outside the intralaminar nuclei, projections of the BG-recipient thalamus terminate most densely in L1 (the molecular layer, consisting mostly of the apical dendrites of pyramidal neurons), while projections of cerebellum-recipient thalamus terminate chiefly within L2-L5; the two terminal fields overlap and intermingle in cortex so that pyramidal neurons are simultaneously under the influence of the BG and cerebellum (Kuramoto et al. 2009; Jinnai et al. 1987). The rat and cat ventromedial nucleus, a major recipient of SNr output, has been noted for directing its output, covering large portions of the cerebral cortex, almost exclusively to L1 (Herkenham 1979; Glenn et al. 1982). Recent experiments in the rat have demonstrated that thalamic projections to superficial cortex are comprehensive, extensively overlap, broadly arborize tangentially, and are variously intra- and inter-areally divergent and convergent (Rubio-Garrido et al. 2009). Each square mm of superficial cortex was found to be innervated by an average of ~4500 thalamocortical neurons, and the most profuse nuclei of origin were the VL, VA, and VM, each of which was noted for terminal fields targeting widely separated areas in cortex. Primary sensory nuclei were found to be completely absent from the projection to superficial cortex.
Broadly branching axons, preference for superficial cortex, and the absence of contributions by primary sensory thalamus, suggest a mesoscopic modulatory function for these projections. Electrophysiological experiments and simulations support this. In vivo experiments on optogenetically manipulated mice show that the VM nucleus is directly implicated in large scale cortical activation (Honjoh et al. 2018). In vitro experiments in rat suggest that apical inputs to L5 neurons have a negligible direct effect on the soma, due to severe attenuation, and shunting by back-propagating action potentials (Larkum et al. 2004). Mathematical modeling of these projections suggests that they can serve a purely modulatory function, biasing inference by selectively amplifying activity in the pyramidal neurons they target, without disrupting information flowing through their more basal inputs, and without contributing information to their outputs beyond that implicit to selective amplification (Kay et al. 2019‡). Beta (and higher) oscillation in the BG, reaching superficial cortex, apparently does not in itself manifest as oscillation at the somata of receiving cortical projection neurons, but rather appears to be low-pass filtered with a cutoff frequency of 5 Hz (Rivlin-Etzion et al. 2008).
When thalamocortical projections provide subthreshold input to the L1 apical dendrites of L5 pyramidal neurons, that input appears to selectively increase the effective gain of the pyramidal neurons, such that temporally coincident input (within 20-30 ms) to their somata induces bursting output (Larkum et al. 2004; Kay et al. 2019‡). According to the BGMS model, BG-influenced activity in these thalamocortical projections, exhibiting multi-areal synchrony, reinforces activity in selected cortical areas, both those that triggered the BG response, and associated areas that are contextually relevant. If corticocortical bursting itself effects long range synchronization (Womelsdorf et al. 2014), then the BG may instantiate long range synchronies by controlling and coordinating the location and timing of cortical bursting and burst receptivity. Paths through the intralaminar nuclei to deep cortical layers are likely crucial to the temporal coordination and selectivity of these responses, as discussed in detail later (§7).
Recent evidence demonstrates that projections of the mediodorsal nucleus to PFC directly drive activity in cortical fast spiking interneurons, suggesting a mechanism whereby MD can mediate inhibition of conflicting or contextually extraneous activity (Delevich et al. 2015; Rikhye et al. 2018). As discussed earlier (§5.4), cortical FSIs exert a robust and precise influence on oscillatory activity in their targets, and may elevate receptivity to activity with the preferred frequency and phase. Thus, paths from the BG to cortical FSIs via the thalamus are a likely substrate for the precise selections that have long (Redgrave et al. 1999) been attributed to the BG.
The density and overlap of the thalamocortical projection to L1 (Rubio-Garrido et al. 2009) suggest that BG output associated with consolidated skills recruits comprehensive modulation of targeted cortical areas. Beyond the domain of well-worn skills, there are intrinsic mechanisms in cortex that may heal intra-areal gaps in modulation. It has been proposed that activity entering cortex through L1 spreads horizontally through L2 and L3, yielding mesoscopic facilitation of firing in L5 pyramidal cells, thereby controlling long range effective connectivity (Roland 2002). Lag-free lateral spread of oscillation up to high gamma has been demonstrated in vitro, working through interactions in L2 and L3, entailing an ensemble dynamic of gap junctions and GABAergic fast spiking interneurons (Tamás et al. 2000).
The neurons of L2 (the external granular layer) have larger receptive fields and a higher incidence of combined feature selectivity than L3 neurons, and their projections are exclusively corticocortical (Gur and Snodderly 2008; Markov and Kennedy 2013). The apical dendrites of the small pyramidal neurons of L2 intermingle in L1 with the apical dendrites of L5 pyramidal neurons, spreading 100-200 µm laterally (Meller et al. 1968; Noback and Purpura 1961). Because L1 and L2 are adjacent, the implicated thalamocortical appositions are more proximal, so likely subject to markedly less of the attenuation and filtration characterizing apical inputs to L5 neurons. Putative high frequency BG modulation of thalamocortical projections to L1 might therefore induce correlated discharges in L2 pyramidal neurons, spreading this high frequency activity laterally within the superficial layers. Moreover, in vitro experiments with pyramidal neurons from these layers have demonstrated nonlinear coincidence detection dynamics, with windows only 4-7 ms long (Volgushev et al. 1998), suggesting temporal specificity in L2/L3 mechanisms sufficient to bias competition among conflicting gamma oscillations.
Discontinuities have been found in the local horizontal linkages of cortex, particularly in L2 and L3, that are posited to tessellate sensory areas along boundaries of similar function and features (Rockland and Lund 1983; Ojima et al. 1991; DeFelipe et al. 1986; Juliano et al. 1990). Similar tessellation, into “stripes” 2-3 mm long and 200-400 µm wide, has been described in PFC (Levitt et al. 1993; Pucak et al. 1996), and is posited to define the limiting spatial resolution with which the BG can modulate cortical activity (Frank et al. 2001). Spreading, but spatially restricted, synchronized activation of superficial modulatory layers has semantically valid effects with a cortical layout in which mental categories and analogs are represented by precise, spatially graded semantic continuities—feature maps—not only in sensory receptive fields, but throughout the cerebral cortex, as evidence suggests (Rao et al. 1999; Huth et al. 2012; Simmons and Barsalou 2003; Rajalingham and DiCarlo 2019; Lettieri et al. 2019; Zhang et al. 2019a‡).
The precise significance of these areal activations is complicated by multidimensionality and mixed selectivity in feature maps. A canonical example of a multidimensional map is the primary visual cortex, wherein the representations of ocular dominance, orientation, and spatial frequency, are folded into small local maps embedded within an overall map arranged in registration with the two major dimensions that represent visual field position (Yu et al. 2005). In associative areas of cortex, which are more densely BG-recipient, multidimensionality and mixed selectivity are yet more prominent (Rigotti et al. 2013; Huth et al. 2016).
Striosomal BG paths (reviewed later (§8.3)) are major modulators of the supply of dopamine (DA) to the forebrain. While DA promotes oscillatory responses to activity in proximally apposed afferents to pyramidal neurons, it has been found to attenuate receptivity to inputs on apical dendrites, which may “focus” or “sharpen” the effects of inputs to those cells (Yang and Seamans 1996). Moreover, the GABAergic interactions among L2/L3 basket FSIs that are crucial for the elimination of phase lags in laterally spreading oscillation (Tamás et al. 2000) are depressed by DA (Towers and Hestrin 2008). This suggests that phasic DA induces phase lags in L2/L3 that increase with distance from the locus of excitatory input, perhaps producing a temporal center-surround effect that effectively focuses cortical responses. Because expressions of plasticity are pervasively spike-timing-dependent (Song et al. 2000), this phase lag control mechanism may also have important consequences for the formation and refinement of cortical feature maps.
These arrangements suggest a corollary to the central proposition of the BGMS model: not only do the BG control effective connectivity in cortex, they also separately control the dynamic characteristics of effective connections. BG influences on cholinergic and serotonergic centers, reviewed later (§9), extend this control.
Most of the corticothalamic population arises from L6, with small terminals apposing distal dendrites in thalamus, and reciprocation is particularly prominent in this projection (Rouiller and Welker 2000). The corticothalamic projection topographically and comprehensively reciprocates the thalamocortical projection, with consistent rules such that each thalamic locus that originates a given type of projection to cortex has a corresponding corticothalamic type reciprocated (Deschênes et al. 1998). The population of corticothalamic fibers is considerably more numerous than the thalamocortical population, by roughly a factor of 10 (Deschênes et al. 1998), and some of these fibers exhibit terminal fields that spread laterally, to encompass neighboring reciprocal receptive fields (Rouiller and Welker 2000).
In an in vitro study in rat, the delay of the corticothalamic projection from L6, and its variability, were found to be significantly greater than those of the thalamocortical projection, 5.2 ± 1.0 ms and 2.1 ± 0.55 ms respectively (Beierlein and Connors 2002). While thalamocortical delays are essentially fixed and very tightly aligned (Salami et al. 2003), L6 corticothalamic delays evidence supernormality. This entails reduction of delay and threshold below baseline after the relative refractory period (Swadlow et al. 1980). Supernormality was found to persist for roughly 100 ms following a discharge, and with a 40 Hz stimulus it reduced corticothalamic delay by up to 12% (Beierlein and Connors 2002). While the functional significance of supernormality in the corticothalamic projection is elusive, it might arrange for advancement of spike timing in cortex as activity intensifies, either matching supernormality in implicated corticocortical projections, or producing other useful timing-related effects, such as phase-of-firing intensity encoding (Masquelier et al. 2009).
The numerosity, structure, and variability of the L6 corticothalamic projection suggest it may be subject to some of the same pressures producing convergence, divergence, and variability in the corticostriatal and striatopallidal projections—particularly, the need for a supply of inputs with appropriate characteristics to meet complex and widely varying topological and spike alignment requirements in paths through thalamus terminating in feedback-recipient layers of cortex. These arrangements were discussed at much greater length earlier (§3).
Dissociation of the functions of corticothalamic and corresponding thalamocortical projections has recently been demonstrated in rats. Using narrowly targeted pharmacological manipulations of the mediodorsal thalamic nucleus and the reciprocally linked area of frontal cortex, Alcaraz et al. (2018) show that inhibition of the corticothalamic cell population disrupts behavioral responsiveness to changes in reward magnitudes, while inhibition of the thalamocortical population disrupts behavioral responsiveness to changes in cause-effect associations. These patterns appear to be consistent with the proposition that corticothalamic projections to motor, associative, and limbic thalamus carry a supply of convergent inputs that is modulated by extrinsic inputs, particularly from the BG, producing selective and sensical thalamocortical responses.
A full account of the function of the corticothalamic projection has proved elusive (Goldberg et al. 2013). According to the BGMS model, the corticothalamic projections from L5 and L6 to BG-recipient thalamus arrange for BG output to be able to select, in each channel, which frequency and phase of cortical activity is to be reinforced and which are to be inhibited. Evidence and modeling suggest that conflicting rhythms in the afferent activity to a cortical neuron shift its discharge pattern away from rhythmic regularity and toward randomness (Gómez-Laberge et al. 2016). Such a response might work to assure that corticothalamic afferents from conflicted cortical loci, converging with BG afferents, can produce postsynaptic activity entrained by those BG afferents, for any particular frequency and phase of BG-selected activity. Moreover, as noted above, corticothalamic fibers are far more numerous than thalamocortical ones, and pool afferents from L5 and L6. This suggests a relatively high degree of convergence in the corticothalamic projection, further working to assure a supply of suitable excitatory afferent activity, and including activity associated with both feedforward and feedback projections.
While the main input to the matrix compartment of striatum arises from cortical L3 and L5, as reviewed above, there are profuse projections from BG-recipient thalamus to striatum, implicitly relaying input from L6. Inputs from cortex and thalamus converge on individual SPNs, with similar axodendritic patterns, but with cortical inputs more numerous (Huerta-Ocampo et al. 2014). Projections from thalamic VA/VL to striatum converge with functionally corresponding projections from cortex (McFarland and Haber 2000). Evidence from primates shows that striatal FSIs are likely broadly targeted by projections from intralaminar thalamus (Sidibé and Smith 1999). Intralaminar projections have also been noted for supplying the striatum with information relating to salient sensory events (Matsumoto et al. 2001), and are implicated in the learning of changes in instrumental contingencies, through projections to cholinergic interneurons (Bradfield et al. 2013). In vitro experiments demonstrate projections from the thalamus to striatal SPNs, with distinctive synaptic properties, such that postsynaptic response generation is likelier than for corticostriatal synapses, but repetitive stimulation depresses postsynaptic depolarization (Ding et al. 2008). In awake monkeys, activation of projections from intralaminar thalamus to striatum has complex effects, with SPN discharge induced only by rapid bursts from thalamus, and long latencies peaking 100-200 ms after intralaminar stimulation (Nanda et al. 2009).
Significantly, intralaminar thalamostriatal projections strongly prefer the matrix compartment (Sadikot et al. 1992b) to which the direct and indirect paths are confined. Thalamostriatal and thalamocortical projection neurons in the intralaminar nuclei are intermingled, and many axons branch to innervate both striatum and cortex (Deschênes et al. 1996; Parent and Parent 2005; Kaufman and Rosenquist 1985a), so that synchronies in the striatal projections of these nuclei are presumptively representative of synchronies in their cortical projections. Thus, thalamostriatal inputs implicitly reflect the current effect of BG output upon the cortex. These subcortical feedback loops may facilitate regulation of BG output to bring modulatory results into conformity with intentions, both dynamically and, as discussed earlier (§3.5), by driving the expression of plasticity.
They may also provide for sequential elaboration of BG output, adjusting thalamocortical modulations with greater speed and precision than is possible within cortico-BG loops. Intralaminar thalamic projections to the globus pallidus, substantia nigra, and subthalamic nucleus (Sadikot et al. 1992a) may serve similar roles, exploiting loop delays that are much shorter than the propagation delays of the corticostriatal and striatopallidal projections (Kitano et al. 1998; Harnois and Filion 1982). Indeed, for some tasks, cortex is crucial for initial acquisition but not necessary for subsequent performance (Kawai et al. 2015; Wolff et al. 2019‡; Dhawale et al. 2019‡). Moreover, tight and bidirectional integration of the BG indirect path with the cerebellum through subcortical pathways has been noted (Bostan and Strick 2010; Bostan et al. 2013; Bostan and Strick 2018; Milardi et al. 2016), and seems likely to be prominent in mechanisms underlying performance of rapid, precise, sequential cognition and behavior, including the production of rhythmic behavior with cadences that are precisely consistent, but continuously variable.
Compared to the motor and association nuclei, the intralaminar nuclei of the thalamus are small, but they have exceptional characteristics and functions placing them at the very center of cognitive coordination and awareness. Moreover, among thalamic nuclei, the intralaminars are uniquely intimate with the BG, and indeed have been described as an integral part of the BG system (Parent and Parent 2005). In the BGMS model, the intralaminar nuclei, through their broad projections and dynamical characteristics, work as a high fidelity broadcast mechanism whereby long range effective connectivity, and therefore cognition, are oriented by spike-timing-dependent gain.
Some studies in rat and cat report intralaminar thalamocortical projections principally targeting L1 (Royce and Mourey 1985; Royce et al. 1989), like the projections of the BG-recipient populations in the MD, VA, VL, VM, and VPLo nuclei, whereas other studies in primate, rat, and cat report intralaminar projections principally to L5 and L6, where individual axons branch widely and arborize massively to appose the somata and proximal dendrites of great numbers of pyramidal neurons, and may terminate in L1 only more sparingly (Parent and Parent 2005; Kaufman and Rosenquist 1985a; Berendse and Groenewegen 1991; Llinás et al. 2002). The disparities among these studies have been suggested to relate to actual physiological distinctions among the species at issue, made all the more likely by the particularly active recent evolutionary history of the intralaminar nuclei and cerebral cortex (Royce and Mourey 1985), but may simply be methodological artifacts.
While intralaminar projection fibers to frontal cortex are greatly outnumbered by those from non-intralaminar BG-recipient thalamic nuclei (Barbas et al. 1991; Schell and Strick 1984), intralaminar projections are strikingly widespread, encompassing nearly the entire neocortex. Experiments in rats and cats demonstrate that the CM/PF nuclei, comprising the caudal group, project to motor, frontal eye fields (FEF), orbitofrontal, anterior limbic, cingulate, parietal, and visual cortex, and to many structures of the medial temporal lobe, though not to the hippocampus proper (Royce and Mourey 1985; Berendse and Groenewegen 1991). In the same two species, the rostral CL and PC nuclei project widely and without consistent topography to the FEF, anterior cingulate, insular, parietal areas 5 and 7, visual, and auditory cortex (Kaufman and Rosenquist 1985a; Royce et al. 1989; Berendse and Groenewegen 1991). Studies in cat (Cunningham and Levay 1986) and macaque (Doty 1983) have identified sparse but distinct projections from the rostral intralaminar nuclei to L1, L5, and L6 of primary visual cortex (area 17). Collating many of these results, a metastudy pooling thalamocortical and corticothalamic projections in cat concluded that the intralaminar nuclei connect very widely with most of visual, auditory, motor, and prefrontal cortex; though nearly all of these connections were characterized as weak or sparse, of 53 cortical areas studied, only 7 (the contiguous primary, posterior, ventroposterior, and temporal auditory fields, the posterior suprasylvian area of visual cortex, and the hippocampus/subiculum) were not reported to be connected with any of the BG-recipient intralaminar nuclei (Scannell 1999).
Single axons from CL and PC have been noted to branch multi-areally to innervate visual and parietal association cortex, suggesting a general function for the intralaminar nuclei, rather than specific functions in the spatial processing of visual information (Kaufman and Rosenquist 1985a).
The broad cortical projection field of the intralaminar nuclei, their extreme divergence, and their intimacy with oscillatory dynamics, were demonstrated by the “recruiting response” reported in early experiments in cats. Oscillatory activity spanning nearly the entire cerebral cortex, most strongly in frontal areas, was evoked with electrical stimulation centered anywhere within the intralaminar region (Morison and Dempsey 1941; Dempsey and Morison 1941). The ventral anterior, mediodorsal, and ventromedial nuclei, prominent in the system of superficially projecting BG-recipient thalamus detailed earlier (§6), exhibit similar indications of large scale connectivity. The VA nucleus in particular has also been implicated in the generation of the recruiting response (Skinner and Lindsley 1967).
CNS insults that bilaterally destroy not only the rostrocaudal extent of the intralaminar nuclei, but also the adjacent MD nucleus, are consistently associated with the permanent vegetative state (Schiff 2010). Pharmacological manipulation of the intralaminar nuclei can rapidly abolish or restore wakefulness (Alkire et al. 2008), and a special indispensability to consciousness has been proposed for these nuclei (Bogen 1995; Baars 1995). Indeed, primates rendered unconscious with propofol anesthesia can be promptly roused to wakefulness solely by high frequency electrical stimulation of the intralaminar thalamus (Redinbaugh et al. 2020; Donoghue et al. 2019‡).
Sleep spindles, which entail tightly synchronized responses spanning large areas of cortex, also demonstrate the broad scope of intralaminar projections. In spindling, activity in the corticothalamic projection and thalamic reticular nucleus are thought to drive thalamocortical cells to simultaneous discharge in nuclei spanning much of the thalamus, particularly through highly divergent projections from the rostral reticular nucleus through the BG-recipient intralaminar and association nuclei (Contreras et al. 1997).
The CL and PC nuclei in cats contain neurons whose activity is uniquely related to all kinds of eye movements, fast or slow, self-initiated or evoked, to stimuli and movements characterized visuotopically, allocentrically, by direction of gaze, and various combinations thereof, to eye position, and to polysensory context and vigilance (Schlag et al. 1974, 1980). Activity in these neurons precedes saccade onset by 50-400 ms, and continues during the saccade, whether the saccade is self-initiated or visually evoked, with each neuron showing a consistent but idiosyncratic pattern (Schlag et al. 1974; Schlag-Rey and Schlag 1977; Schlag et al. 1980). While each completed saccade is accompanied by a consistent pattern of activation in some of these neurons, the reverse is not always the case—the same pattern of activation in an intralaminar neuron is sometimes seen in the absence of an executed saccade (Schlag et al. 1974). Nonetheless, microstimulation in the CL and PC nuclei consistently evokes conjugate saccades with a delay of 35 ms for large deviations, suggesting primary involvement in saccade generation (Maldonado et al. 1980).
Similar to CL and PC, evidence suggests that activation of the parafascicular nucleus can by itself generate turning and orienting movements, though this apparently occurs not via its projections to cortex, but rather, via its projections to the STN, implicating STN and other BG components in movement initiation (Watson et al. 2018‡).
Unlike other BG-recipient populations in thalamus, the CM and PF nuclei are densely parvalbumin-positive (Jones and Hendry 1989). In other thalamic nuclei, as noted earlier (§6.3), parvalbumin is associated with putative “core” or “driving” neurons, which are not BG-recipient. In other brain organs, notably the cerebral cortex, striatum, GP, and SNr, parvalbumin is associated with fast-firing, fatigue-resistant neurons. Via the caudal intralaminar nuclei, the BG complete loops within which spike timing is largely determined by parvalbumin-containing, fast-firing, non-fatiguing neurons (Mallet et al. 2005; Bennett and Bolam 1994; Cote et al. 1991), targeting somata and proximal dendrites of pyramidal neurons in deep cortex as described above.
The rostral intralaminar nuclei are densely calbindin-positive (Jones and Hendry 1989), like non-intralaminar BG-recipient thalamus, but the laterodorsal part of the CL nucleus has been shown in adult cats to contain a population of neurons projecting to parietal association cortex that, during wakefulness and REM sleep, regularly emit bursts of 3-4 spikes with interspike intervals (ISIs) shorter than 1.3 ms, at a burst rate of 20-40 Hz, with no apparent signs of fatigue, and an antidromic thalamocortical delay ⩽ 500 µs, indicating conduction velocities (CVs) of 40-50 m/s (Glenn and Steriade 1982; Steriade et al. 1993). During the spindling characteristic of stage 2 sleep, bursts in these cells were found to be even more intense, 8-9 spikes with ISIs as low as 1 ms. The CVs of their axons, uniquely fast among thalamocortical cell classes (Steriade et al. 1993), are roughly 50 times faster than those of striatopallidal axons, which in awake cynomolgus monkeys exhibit CVs under 1 m/s (Tremblay and Filion 1989).
Because BG inputs to intralaminar nuclei are collaterals of inputs to other thalamic nuclei (Parent et al. 2001), the information received from the BG by the intralaminar nuclei presumably duplicates that received by non-intralaminar cells. But high fidelity relay by neurons of the intralaminar thalamus, combined with pyramidal somatic layer targeting, appears to arrange for particularly narrow selectivity through spike synchrony effects. Indeed, feed-forward inhibition in cortex, implicating fast-spiking interneurons, arranges for an extremely narrow coincidence detection window for proximally apposed afferents to pyramidal neurons, -1.5 to +2.4 ms for effective spike summation, even while the coincidence requirement in distal inputs was found to be much looser, -8.6 to +12.3 ms (Pouille and Scanziani 2001). Even absent the influence of FSIs, pyramidal neurons stimulated somatically in vitro have been shown to act as nonlinear coincidence detectors with windows only 4-7 ms wide, that become narrower with rising oscillatory frequency, with the timing of discharges tightly correlated to the timing of somatic membrane potential oscillation (Volgushev et al. 1998).
Evidence discussed earlier (§6.2) implicating L5/L6 in top-down control, with prominent activity in the alpha and beta bands (Bastos et al. 2018), underscores the significance of temporally precise spike relay and distribution by the intralaminar nuclei to these layers. Oscillatory periods ⩾ ~50 ms predominate in L5/L6, much longer than the ~4 ms coincidence window for proximally apposed inputs, so that intralaminar thalamic inputs with high temporal fidelity appear arranged to enable precise selections.
As noted above, BG inputs to primate caudal intralaminar thalamus overwhelmingly appose dendrites, not somata (Sidibé et al. 2002). These appositions are not homogeneous, in that over 80% of SNr inputs to PF in monkey were found to appose small or medium, mostly distal, dendrites, with none apposing somata, while over 75% of GPi inputs to CM were found to appose medium or large, mostly proximal, dendrites, and 5% to appose somata. These patterns of apposition clearly result in looser coupling between the BG and intralaminar thalamus than does the perisomatic, calyceal pattern seen in non-intralaminar BG-recipient thalamus (Bodor et al. 2008). Perhaps more important, because corticothalamic inputs to these nuclei are also predominantly through small distally apposed terminals (Rouiller and Welker 2000), BG inputs are positioned as frequency- and phase-selective filters, with the potential for extensive presomatic nonlinear computation, enhancing computational power and combinatorial flexibility. Thus a single intralaminar neuron might participate in a vast variety of scenarios characterized by distinct corticothalamic and nigrothalamic input patterns, each producing somatic discharges, but by different combinations of dendritic inputs.
While inputs from L6 predominate in BG-recipient motor/association thalamus, in BG-recipient caudal intralaminar thalamus it is L5 inputs that predominate (Van der Werf et al. 2002; Balercia et al. 1996; Cornwall and Phillipson 1988; Royce 1983a, 1983b). This mirrors targeting of L5 in thalamocortical projections from this area, discussed above, and moreover shares its laminar origin with many corticocortical projections (Reiner et al. 2003). This is significant, because it suggests that corticocortical projections are systematically accompanied by trans-intralaminar paths, sharing exactly the same origins and targets, and subject to temporally precise gating by the BG direct path, in which L5 is similarly predominant in inputs to striatum, as reviewed earlier (§6.2).
Widespread intralaminar projections appear arranged to broadcast a temporally precise but spatially diffuse signal to most of cortex, while non-intralaminar projections to superficial layers have mesoscopic spatial specificity, more restricted (principally frontal) areal targets, and relatively crude temporal specificity (though their appositions on cortical FSIs (Delevich et al. 2015; Rikhye et al. 2018) likely provide for temporal precision). At the heart of the BGMS model is the proposition that the BG coherently modulate these two influences, so that their convergence and inter-areal linkage in cortex provide for spatiotemporal specificity and consequent precision in the control of effective connectivity. By interacting with intrinsic cortical activity, these inputs rapidly and dynamically recruit specific large scale networks. A corollary of this view is that the nuclei of the thalamus act as attentional spotlights, with selectivity rooted in both temporal and spatial specificity, while the BG are prominent in the orientation of those spotlights.
The BGMS proposal can be summarized as follows: When an input pattern triggers a selection in the striatum, the timing of striatal output tracks the prevailing timing of the input pattern, and the GPi, VP, and SNr impart that timing to the thalamus, with striatopallidal and striatonigral delays tuned for optimum effect (optimality being a function of cortical rhythms and corticocortical conduction delays, discussed in detail earlier (§3)). The intralaminar nuclei, through widespread diffuse projections throughout the cortical column (excepting only L4), impart discriminative receptivity to any activity that is precisely synchronous with that prevailing in the input pattern that stimulated the BG response, and narrowly reinforce its generation in its loci of origin. The non-intralaminar nuclei, through dense, mesoscopically specific, largely closed-loop projections, chiefly to L1, fortify activity in selected areas, particularly those contributing to the input pattern. When this fortification is strong, and coincident with substantial activity in the corresponding proximally apposed afferents, bursting is promoted (Larkum et al. 2004), further promoting establishment of effective connections (Womelsdorf et al. 2014).
Closed-loop paths through non-intralaminar nuclei largely implicate areas in frontal cortex, which are the densest targets of non-intralaminar BG-recipient thalamocortical projections, but other association areas in primates, notably in parietal and temporal cortex, are also implicated. All of these areas are thought to originate feedback signals with top-down control over their targets. By this narrative, the BG direct path establishes and fortifies top-down control connections from both ends, with the MD, VA, and VL nuclei fortifying the top end of the connection, and the CM, PF, PC, and CL nuclei tuning both ends to complete the connections. An additional function of BG-modulated thalamocortical afferents to L1 is that they open gates for feedforward signals, suggested by evidence (Bastos et al. 2018; Lundqvist et al. 2018a) that activity in L2/L3 is associated with bottom-up inputs.
Open loop direct paths through non-intralaminar nuclei may serve to complete activation of a distributed cortical ensemble that is only partly activated when it first triggers a striatal response, particularly when the triggering pattern largely originates in sensory cortex. Closed loop paths through intralaminar nuclei may tighten synchrony throughout the selected ensemble, and provide reinforcement that is highly selective, due to the narrow coincidence windows associated with proximal inputs to pyramidal neurons.
Notably, trans-thalamic inputs may actively inhibit and disconnect activity that is not synchronous (particularly, that is antisynchronous) with the thalamocortical signal, by feedforward inhibition via cortical FSIs. Evidence noted earlier (§6.6) strongly suggests that the path from MD to PFC entails such a mechanism (Delevich et al. 2015; Rikhye et al. 2018). And it is clear from the response to sleep spindles (Peyrache et al. 2011) that both pyramidal neurons and FSIs are targeted by thalamocortical projections. Feedforward inhibition associated with this arrangement enforces extremely short windows of summational receptivity (Pouille and Scanziani 2001).
It may be important that intralaminar projections, which target most of the cortex, are subject to extremely narrow coincidence windows. With wider windows, the intralaminar broadcast mechanism seems prone to establishment of spurious connections. Indeed, schizophrenia involves abnormal enlargement of these coincidence windows (Lewis et al. 2005; Gonzalez-Burgos et al. 2015), while lesioning and deactivation of intralaminar nuclei has been found to relieve hallucinations and delusions associated with Sz and other psychoses (Hassler 1982).
Sz is also characterized by enlargement of the time window within which visual stimuli are judged to be simultaneous (Schmidt et al. 2011), and by abnormalities in the simultaneity criteria for implicit audiovisual fusion (Martin et al. 2013). Beyond Sz, loosening of simultaneity criteria, and deficient perception of short time intervals, may be characteristic of psychosis generally (Schmidt et al. 2011; Ciullo et al. 2016).
As evident from their function in vision and saccades, the BG-recipient intralaminar nuclei are a jumble of perceptual and motoric function, with activity in individual neurons highly correlated with both. Roles for these nuclei in executive control, working memory, and general cognitive flexibility—capacities that are most developed in humans—have also been shown (Van der Werf et al. 2002). Over the course of mammalian evolution, the intralaminar nuclei, particularly the posterior group, have undergone relative expansion and elaboration, reaching their greatest extent in primates, and in humans particularly (Macchi and Bentivoglio 1986; Royce and Mourey 1985; Herkenham 1986).
Psychosurgical results in humans give further evidence that these nuclei can originate driving inputs to motoric, perceptual, cognitive, and motivational centers. Treatment of Gilles de la Tourette syndrome (GTS) by stereotactic ablation or rhythmic electrical stimulation of the rostral (Rickards et al. 2008) or caudal (Houeto et al. 2005; Servello et al. 2008) intralaminar nuclei has produced substantial and sustained abatement, in some cases almost complete remission, of compulsive behavior (tics) in many patients. Similarly, severe or extreme symptoms of obsessive compulsive disorder (OCD) have been substantially, consistently, and sustainably alleviated by unilateral lesioning of the right intralaminar nuclei (Hassler 1982), or by rhythmic electrical stimulation localized to the inferior thalamic peduncle, inactivating connectivity between intralaminar nuclei and orbitofrontal cortex (Jiménez-Ponce et al. 2009). GTS and OCD involve extensive BG abnormalities (Graybiel and Rauch 2000; Albin and Mink 2006; Kalanithi et al. 2005), so alleviation of symptoms by IL inactivation suggests functional prominence of the intralaminar nuclei in BG dynamics, and may be evidence of key involvement in the transmission of BG output to cortex.
Functional deficits in Sz are intimately related to the functional roles of the intralaminar nuclei. Eye tracking and saccade control are dysfunctional, suggesting particular deficits in anticipatory control and the suppression of distractors (Levy et al. 1994; Fukushima et al. 1988; Hutton et al. 2002), and aberrant connectivity between the intralaminar nuclei and PFC has also been described (Lambe et al. 2006). Sz has been found to be associated with significant relative reduction in volume and metabolic hypofunction in the centromedian nucleus, in addition to the MD nucleus and pulvinar, in a study that found no significant effects by these measures in other thalamic nuclei (Kemether et al. 2003; Hazlett et al. 2004).
The BG-recipient intralaminar thalamus expresses D2 dopamine receptors at particularly high density (Rieck et al. 2004), and these receptors are targeted by antipsychotic drugs, usually with ameliorative effect for positive symptoms (Nordström et al. 1993; Kay et al. 1987). There is evidence from experimental clinical practice that lesioning of the mediodorsal and rostral intralaminar nuclei can permanently eliminate delusions and somatosensory, auditory, and visual hallucinations associated with Sz, while rhythmic (20 and 50 Hz) electrical stimulation of these areas can abolish symptoms promptly (Hassler 1982).
That some hallucinations and visuocognitive deficits in Sz may involve BG interaction with the intralaminar nuclei is further suggested by the common occurrence in PD of visual and other hallucinations and delusions (Barnes and David 2001; ffytche et al. 2017) and impaired shifting and maintenance of visual attention (Wright et al. 1990). PD is marked by abnormally strong coupling within BG loops (Hammond et al. 2007), and hallucinations incidental to PD appear to be associated with pathological coupling of visual areas with the “default mode network” (Yao et al. 2014; Shine et al. 2015; Walpola et al. 2020). In short, hallucinations and delusions incidental to PD might be due in large part to pathologically synchronized BG output, inducing pathological persistence and widespread synchronization of neocortical activity, which could functionally connect spurious activity in visual cortex to hub areas. Consistent with this account, extensive thalamic cell loss in PD specific to the caudal intralaminar nuclei (Henderson et al. 2000) suggests that, among thalamic areas, these nuclei bear the brunt of the abnormal dynamics characteristic of the disease. PD and Huntington's disease are also both associated with voluntary saccade deficiencies, including abnormal distractibility in Huntington's (Bronstein and Kennard 1985; Lasker et al. 1987, 1988), resembling some of the oculomotor abnormalities associated with Sz.
Auditory hallucinations are commonly associated with Sz (de Leede-Smith and Barkus 2013; McCarthy-Jones et al. 2014), and also sometimes occur in advanced PD (ffytche et al. 2017). Many of the brain areas implicated in these hallucinations are within or intimate with the BG (Shergill et al. 2000). In cat, connections of the parafascicular nucleus with secondary auditory cortex and the anterior auditory field have been demonstrated (Scannell 1999), but as noted above (§7.3), no direct connections have been found between the intralaminar nuclei and the primary and several adjoining auditory fields (Scannell 1999). This lacuna is intriguing, in that it suggests that intralaminar input may be detrimental to signal integrity there, outweighing the benefits that evolutionarily stabilize intralaminar innervation elsewhere.
Evidence from EEG studies indicates that activity in cortical auditory areas synchronizes precisely with regular features of rapidly changing auditory stimuli, independent of and indeed in the absence of attention; beyond auditory cortex, in widely distributed areas such as frontal and parietal cortex, attention is required for sustained increases in activity in response to such auditory patterns (Herrmann and Johnsrude 2018). Consistent with local, automatic processing, activity in secondary auditory cortex shows a delay relative to primary cortex consistent with direct, feedforward signal propagation (Barth and MacDonald 1996). Bringing these disparate facts together: perhaps the prominence of auditory hallucinations in Sz is a result of top-down regulatory influences that are particularly weak (both normally and in Sz), intrinsically dysregulated activity in non-BG-recipient auditory areas, and intrinsic and extrinsic dysregulation in BG-recipient auditory areas, pervasively implicating GABA signaling (discussed at greater length later (§12.5)).
As noted above, sleep spindling, generating broadly synchronized responses in cortex, particularly implicates the BG-recipient intralaminar and association nuclei of the thalamus (Contreras et al. 1997). Spindling is thought to be crucial for consolidation during sleep of new associations (Tamminen et al. 2010; Genzel et al. 2014). Moreover, a direct association has been demonstrated between the prevalence of fast parietal spindles during stage 2 and slow wave sleep, and fluid intelligence (Fang et al. 2017).
A consistent pattern of deficient spindle activity in stage 2 sleep has been demonstrated in Sz, with severity of symptoms correlated to degree of deficiency (Ferrarelli et al. 2007, 2010a; Wamsley et al. 2012). Because synaptic homeostasis mechanisms largely operate at the level of individual microcircuits, neurons, and synapses (Turrigiano 2011), spindling deficits may cause progressive deterioration of the long range circuits that are the physiological basis of effective connectivity in wakefulness. Such deterioration is, in any case, characteristic of Sz (Lim et al. 1999; Mori et al. 2007; Collin et al. 2014; de Leeuw et al. 2015). As reviewed later (§12.5), it is the hub areas of cortex that are most implicated in the circuit deterioration characteristic of Sz. These are the areas most clearly implicated in fluid intelligence, as explored later (§14.3.4).
Sleep spindles have been found to preferentially recruit FSIs in PFC, more than pyramidal projection cells there (Peyrache et al. 2011). This is likely a consequence of feedforward inhibition in response to the lengthy ultra-high frequency bursts associated with spindling, importantly demonstrating that thalamocortical projections appose both FSIs and pyramidal cells in cortex. Impairment of GABA synthesis in intrinsic FSIs of DLPFC, and consequent deficiencies in cortical projection neuron synchronization and loosening of spike coincidence criteria, have been implicated in Sz (Lewis et al. 2005; Gonzalez-Burgos et al. 2015). PFC FSI response patterns are also modified by dopamine inputs (Tierney et al. 2008), which are abnormal in Sz (Grace 2016). The consequences of severe deficiencies in sleep spindling, simultaneous with disruption of feedforward inhibition by cortical FSIs, may disrupt BGMS with particular potency. Whether spindle and PFC FSI deficiencies are part of the etiology of Sz, or are sequelae, remains to be determined and may vary. It is probably significant that both can result directly from GABA dysfunction.
Evidence that the intralaminar nuclei are profusely innervated by the BG and integral to BG circuitry, that they are innervated by and proximally appose L5 pyramidal neurons, that these appositions are subject to stringent (<4 ms) coincidence requirements, and that spike bursts from highly energetic intralaminar neurons in a state of wakefulness last only 4-5 ms and recur at a rate of 20-40 Hz, suggest that BG output associated with well-practiced behavior and cognition is precisely aligned on this timescale. While the timing of spikes in projections to superficial cortex is surely significant, it is in the projections to somatic layers that timing appears most critical, and that the potential for timing-based selectivity is most apparent.
While BGMS as discussed in this paper most directly implicates the direct path, BG circuitry beyond the direct path is just as functionally crucial, and indeed is even more extensive and broadly connected than the direct path. The striatum is the common component in all these circuits. The striatum is a particularly complex brain organ, structured simultaneously along multiple schemes overlaid upon, and interacting with, each other in intricate patterns (Graybiel 1990; Kreitzer 2009; Tepper et al. 2010; Bolam et al. 2000; Märtin et al. 2019). Its striosome-matrix compartmentation, and its direct-indirect dichotomy, both bear upon the present hypothesis.
Among corticostriatal projection neurons, there is evidence that most direct path cells, but not most indirect path cells, are reciprocally connected over long ranges at the single unit level, and are a specialized population dedicated to intracortical connectivity and striatal innervation (“intratelencephalic”); the indirect path is predominantly innervated by collaterals of projections that descend through the pyramidal tract, and whose corticocortical collaterals are not reciprocal (Lei et al. 2004; Morishima and Kawaguchi 2006). As noted earlier (§4.3), projections from interconnected cortical areas systematically converge on striatal FSIs at the single unit level (Ramanathan et al. 2002), and FSIs show a substantial preference for direct path SPNs (Gittis et al. 2010). Thus, the innervation of the direct path is distinguished by systematic patterns of reciprocal long range connectivity and corresponding striatal convergence, whereas indirect path corticostriatal inputs are predominantly collaterals of descending fibers such as corticopontine motor output, whose cells of origin do not reciprocate with each other, and as reviewed below, show markedly less striatal convergence.
Evidence suggests that the direct path through the BG to thalamus implicates SPNs in the matrix compartment exclusively (Rajakumar et al. 1993), and that the corticostriatal innervation of the matrix is differentiated from that of the striosomes in important ways. While the striosomes and matrix are both broadly targeted by most cortical areas, the striosomes preferentially receive projections from L6 and deep L5, while the matrix is preferentially targeted by superficial L5, and by L2 and L3 (Gerfen 1989; Kincaid and Wilson 1996).
Ascending projections from the densely direct-path-recipient PF thalamic nucleus pervasively and diffusely innervate the matrix compartment of associative striatum, while largely avoiding striosomes; CM projections to sensorimotor striatum are less pervasive but similarly prefer matrix (Sadikot et al. 1992b). The CL and PC nuclei also project densely to the caudate striatum (Kaufman and Rosenquist 1985a). The striatal projections of these intralaminar nuclei appose the dendrites of SPNs, with varying physiological and morphological properties (Lacey et al. 2007), and evidence also suggests that they innervate striatal FSIs (Sidibé and Smith 1999). As proposed earlier (§6.11), thalamostriatal projections may position the striatum to monitor (and therefore optimize and rapidly sequence) the synchronies that its output produces in thalamus, and thus presumptively in cortex, via BG output structures.
Intriguing areal distinctions in cortex have also been identified. In primate, dorsolateral PFC (DLPFC) targets matrix densely and broadly, largely avoiding striosomes, while orbitofrontal and anterior cingulate cortex preferentially target striosomes (Eblen and Graybiel 1995). Matrix appears specialized to project to the pallidal segments and the SNr, while striosomes appear specialized to project to midbrain dopamine centers such as the substantia nigra compacta part (SNc), to whose densocellular zone they are reported to be reciprocally linked (Jiménez-Castellanos and Graybiel 1989; Crittenden et al. 2016).
Activity in matrix reflects immediate prior reward, suggesting reward-guided generation of ongoing cognition and behavior, while that in striosomes reflects anticipated outcomes, and is less reflective of immediate prior reward, suggesting implication in the generation of signals that drive or mediate motivation and the expression of plasticity (Bloem et al. 2017).
Striosomes strongly influence the SNc and VTA through a pallidohabenular circuit (Rajakumar et al. 1993; Herkenham and Nauta 1979; Hikosaka 2010; Hong and Hikosaka 2008; Balcita-Pedicino et al. 2011), while dopaminergic projections from the midbrain preferentially target striatal matrix (Graybiel et al. 1987). The involvement of striosome circuitry in motivational processing, and of dopamine in modulating responses to afferent activity, is reviewed later (§9). In particular, their roles in modulating the dynamics of superficial cortical microcircuits in PFC (Yang and Seamans 1996; Towers and Hestrin 2008), introduced earlier (§6.8), are crucial.
The classic technique for differentiating striosomes from matrix is to stain the striatum to visualize distribution of the enzyme acetylcholinesterase (AChE) (Graybiel and Ragsdale 1978), rendering the striosomes as pale poorly stained patches. Serotonergic projections to striatum also preferentially innervate the matrix compartment (Lavoie and Parent 1990). As reviewed in detail later (§9), dopamine, ACh, and serotonin are potent modulators of oscillatory neuronal responsiveness. Thus, differential prominence of these neurotransmitters in the matrix compartment suggests specialization for the relay of oscillatory activity.
Most striatal ACh arises from an intrinsic population of interneurons comprising 2-3% of striatal neurons (Contant et al. 1996), which is believed to be identical to the electrophysiologically identified tonically active neurons (TANs) of the striatum (Aosaki et al. 1995). These neurons discharge tonically at 2-10 Hz in the absence of sensorimotor activity, and are differentially localized to the matrix, particularly to the matrix border regions adjoining striosomes (Aosaki et al. 1995).
The PPN, itself profusely targeted by the GPi and SNr (Semba and Fibiger 1992; Grofova and Zhou 1998; Parent et al. 2001), provides an additional, extrinsic, supply of ACh to the striatum, and this too preferentially targets the matrix compartment (Wall et al. 2013). Moreover, the striatally projecting neurons of the midline and intralaminar thalamus are targeted by the PPN (Erro et al. 1999), and as noted earlier (§3.5), preferentially target the TAN population, participating intimately in goal-directed learning (Bradfield et al. 2013). FSIs, noted above for their selective and robust innervation of direct path SPNs and their putative high fidelity relaying of oscillatory activity, are extensively modulated by cholinergic inputs (Koós and Tepper 2002). Thus, the matrix compartment of the striatum is distinguished by participation in multiple, coordinated cholinergic circuits.
According to the BGMS model, the direct path of the BG establishes task-appropriate long range effective connections, while the indirect path largely serves to damp or desynchronize competing activity, to further secure the selected connections. Wall et al. (2013) identified instructive differences between afferents to these two intermingled populations of SPNs in mouse: The direct path was found to receive significantly heavier projections from primary somatosensory, ventral orbitofrontal, cingulate, frontal association, prelimbic, perirhinal, and entorhinal cortex, and to receive essentially the entire striatal projections from the amygdalar nuclei, STN, and DRN. The indirect path was found to receive a significantly heavier projection from primary motor cortex. Preferential targeting of the direct path by primary somatosensory, and of indirect path by primary motor, comports with a model in which the direct path establishes connections and facilitates actions consistent with context and task requirements, while the indirect path inhibits completed, competing, ineffective, and irrelevant activity and functional connectivity.
Direct path SPNs show higher activation thresholds and more extensive dendritic processes (~25% more dendrites) than indirect path SPNs, suggesting greater integration through the direct path (Gertler et al. 2008). When synchronized cortical activity is confined to a single focus in primary motor cortex, the consequent striatal activation strongly prefers the indirect path (Berretta et al. 1997). This disparity is a natural consequence of the indirect path preference of the corticostriatal projection originating in primary motor cortex, but might also be explained in part by a preferential responsiveness in the direct path to conditions of multi-areal activity, suggested by the role proposed in the BGMS model implicating it in the induction of selective synchronies between distant areas that typically already harbor activity.
The information borne by the intratelencephalic corticostriatal projection appears to be distinct from that borne by the corticostriatal collaterals of the corticopontine projection from the same area. Turner and DeLong (2000) showed that in primate primary motor activity, corticopontine neurons consistently show activity associated with movement execution and, particularly, the muscular contractile command stream, whereas activity in intratelencephalic neurons is often independent of muscle activity, is exquisitely context- and feature-dependent, and is usually confined to a particular aspect of current conditions (sensory context, movement preparation, or movement underway). They suggested that these patterns of direct path input to the striatum are a sparse code, of the sort demonstrated in temporal and visual cortex (Rolls and Tovee 1995; Vinje and Gallant 2000).
Wright et al. (1999, 2001) showed in rat that intratelencephalic corticostriatal afferents from primary sensory areas have diffuse, convergent, and bilateral terminal patterns, implicitly raising opportunities for information integration. In contrast, they showed that corticopontine collateral input is ipsilateral, and preserves topographic specificity and organization, terminating in discrete varicosities without convergence, with thicker and faster axons. Moreover, they showed that the intratelencephalic and corticopontine projections enter the striatum almost at right angles to each other, which appears to further cultivate information integration.
Earlier studies identified the differential pattern of corticostriatal arborizations, finding that those of the intratelencephalic collaterals in the striatum are ~1.5 mm in diameter, with sporadic branching and varicosities, while the corticopontine collateral arborizations are dense, focused within a volume with longest dimension ~500 µm, and do not cross boundaries of adjacent striosomes (Cowan and Wilson 1994; Kincaid and Wilson 1996). Another investigation found that the corticopontine projection originates chiefly in lower L5, while the intratelencephalic projection originates chiefly in upper L5 and in L3 (with L3 predominating slightly in sensory cortex), and that the striatal terminal boutons of the former are roughly twice the size of the terminals of the latter (Reiner et al. 2003). More recent studies have confirmed that intratelencephalic afferents exhibit a higher prevalence of numerous and widely distributed terminals than do corticopontine afferents (Hooks et al. 2018; Morita et al. 2019).
Preferential projection by classes of corticostriatal neurons is a matter of tendencies, not rules. Intratelencephalic corticostriatal axons prefer direct path SPNs by a 4:1 ratio, while corticopontine collateral axons prefer indirect path SPNs by a 2.5:1 ratio (Lei et al. 2004). Recent findings using genetically manipulated mice have shown that the cytological and hodological compartmentation of the striatum into striosomes and matrix is not crisp, with both striosomal and matriceal SPNs receiving both limbic and sensorimotor inputs, and projections to SNc arising from both striosomal and matriceal SPNs (Smith et al. 2016). Earlier studies demonstrated similar minor projections of sensorimotor cortex to striosomes, and revealed sparse projections from striosomal neurons to the pallidal segments (Flaherty and Graybiel 1993). Motivational specificity and contextualization are apparent in striatal matrix activity (Donahue et al. 2018‡), and this intermodal convergence-divergence may be related.
The canonical marker for direct and indirect path SPNs is expression of dopamine receptors from the D1 and D2 receptor families, respectively (Gerfen and Surmeier 2011), but SPNs express DA receptors from the opposing family at low levels (Smith and Kieval 2000), and BG microcircuits intermingle the effects of DA receptors from both families (Gerfen and Surmeier 2011). Indeed, the axons of individual SPNs in primate frequently branch to both direct and indirect path targets (Parent et al. 1995; Levesque and Parent 2005). Moreover, in the ventral pallidum, neurons with projection patterns characteristic of the GPi/SNr and the GPe are closely intermingled, receiving projections from direct and indirect path SPNs (Groenewegen et al. 1993; Smith and Kieval 2000).
Voluntary behavior is preceded by simultaneous activation of both direct and indirect path SPNs (Cui et al. 2013; Donahue et al. 2018‡). This coactivation, while typically antagonistic, is not symmetric (Oldenburg and Sabatini 2015), and evidence suggests that these asymmetric dynamics are central to the sequential elaboration of precisely timed motor commands (Markowitz et al. 2018). Moreover, there is evidence that concurrent and asymmetric activity of SPNs in the direct, indirect, and striosomal paths collectively represents all aspects, phases, and structure of a task, with SPNs tiling the task space with activity to form a continuous representation of the task in all its particulars (Weglage et al. 2020‡; Arcizet and Krauzlis 2018). These dynamics are consistent with the proposition that “matrisomes” consisting of closely intermingled direct and indirect path SPNs, with presumptively overlapping dendritic processes, facilitate coordination of direct and indirect path output (Flaherty and Graybiel 1993).
Beyond these complexities in the physiology and dynamics of the indirect, direct, and striosomal pathways, there are many additional pathways that bypass and supplement them. The hyperdirect path from frontal cortex to the subthalamic nucleus (Nambu et al. 2002) is implicated in stopping actions (Schmidt et al. 2013), while the pallidostriatal projection is thought to be involved in the cancellation of stopped actions (Mallet et al. 2016). There is also evidence of pathways directly linking the cerebral cortex to the globus pallidus (Milardi et al. 2015; Smith and Wichmann 2015), and the globus pallidus to the cerebral cortex (Van Der Kooy and Kolb 1985; Karube et al. 2019‡; Zheng and Monti 2019‡).
The presumptive function of these various cross-channel, cross-receptor, and bypass paths is to enrich the range of dynamics and pool of information available to the implicated individual neurons, by which they might more rapidly or appropriately respond to ever-changing context.
Beyond their GABAergic projections to thalamic relay and association nuclei, the BG are positioned to modulate cortical and thalamic activity through their dopaminergic projections to frontal cortex and associated nuclei of the thalamus, and through projections to the basal forebrain (particularly the nucleus basalis of Meynert, NBM), the pedunculopontine and laterodorsal tegmental nuclei (PPN and LDT), the dorsal and median raphe nuclei (DRN and MRN) at the pontine level of the brainstem, and the thalamic reticular nucleus.
Dopamine (DA) is a key modulatory neurotransmitter intrinsic to the BG, where it raises the excitability of direct path SPNs by activating their D1-class receptors, and reduces the excitability of indirect path SPNs by activating their D2-class receptors (Gerfen and Surmeier 2011). For reviews, see for example Schultz (1998), Bromberg-Martin et al. (2010), and Yetnikoff et al. (2014). Roles for DA in the control of oscillatory activity, in and beyond the BG, have been described that bear directly on the BGMS model, and are reviewed here.
Despite comprising less than one percent of neurons, cholinergic cells perform crucial roles in, and indeed beyond, the nervous system (Woolf and Butcher 2011). They are proposed to play a key role in orienting attention (Sarter and Bruno 1999), in induction of vigilance and fast sleep rhythms (Steriade 2004), in induction of plasticity (Rasmusson 2000), and in the formation of memories (Hasselmo 2006).
Serotonin (5-hydroxytryptamine, 5-HT) is implicated in regulation of sleep and wakefulness (Pace-Schott and Hobson 2002; Monti 2011), cognitive and behavioral flexibility (Clarke et al. 2006), and signaling of reward magnitude (Daw et al. 2002; Nakamura et al. 2008).
These roles of the DA, ACh, and 5-HT systems, and the TRN, are evidently closely related to each other. Indeed, the supplies of DA, ACh, and 5-HT, are closely coupled, as detailed below.
The effects of DA are complex. Through broad projections to BG and amygdalar nuclei, frontal cortex, and associated thalamic nuclei, the release of DA arising from BG-controlled neurons in the ventral midbrain (chiefly SNc, VTA, and the retrorubral field, RRF) and other areas has been proposed to have a crucial role in motivational control, by signaling reward, surprise, novelty, even aversiveness, and in general, saliency (Schultz et al. 1997; Bromberg-Martin et al. 2010; Ioanas et al. 2020‡). Midbrain DA projections have systematic topography (Fallon 1988), and evidence suggests separable correlates in subpopulations of DA neurons for distinct functions, and for various separable aspects of reward prediction error, such as timing vs. magnitude (Lau et al. 2017). Activation of DA projections from the VTA to the basal amygdala has been shown in mice to be associated with the formation of fear memories (Tang et al. 2019‡), which are archetypically aversive. There is evidence that distinct clusters of DA neurons in the ventral midbrain are specialized for an assortment of reinforcement roles, particularly motivational reward and motor invigoration (Saunders et al. 2018), and that distinct BG circuits bear reinforcement signals for distinct functional domains, only some of which entail plainly motivational signaling (Pascucci et al. 2017).
DA projections to the striatum have been shown in mice to be functionally heterogeneous and selective, exhibiting topographic structure, with activation in wave-like spatiotemporal sweeps across regions of functionally related striatum, showing particular and stereotyped heterogeneity along the mediolateral axis (Hamid et al. 2019‡). These spatiotemporal waves showed a strong relationship between propagation direction and instrumental agency: as learning progressed, a task entailing strong instrumental contingency showed progressively more well-defined mediolateral propagation, while a simpler Pavlovian variant of the task showed lateromedial propagation, consistent with established roles for dorsal medial and dorsal lateral striatal functional specialization, and suggesting a crucial role in the dynamics of credit assignment (Hamid et al. 2019‡).
DA has been proposed to signal disparities between expected and actual outcomes, dipping phasically upon disappointment and rising phasically upon surprising reward, driving reinforcement learning mechanisms (Schultz 1998, 2013). In fact, evidence suggests that DA is crucial in signaling prediction errors per se, with or without reward associations (Sharpe et al. 2017).
DA release has been proposed to signal the expected value of work, in order to encourage continuation of efforts expected to culminate in a rewarding outcome, and discourage continuation of other efforts (Hamid et al. 2015). Indeed this neuroeconomic function has been ascribed to the BG as an ensemble (Goldberg and Bergman 2011). As noted earlier (§8.3), striosomes appear specialized to control ventral midbrain DA centers; medial PFC control of striosomes, and striosomal control of ventral midbrain DA, have been implicated in cost-benefit decision making (Friedman et al. 2015; Crittenden et al. 2016). Activity in striosomes, compared to that in striatal matrix, has been shown to preferentially encode reward-predicting cues in particular, and anticipated outcomes in general (Bloem et al. 2017).
Surprising sensory events can evoke prominent, short-latency DA bursts, regardless of reward association, in 60-90% of DA neurons throughout the full extent of the SNc and VTA, apparently constituting an alerting response serving to marshal attention; these bursts seem to correlate with the degree to which the stimulus captures attention by surprise, they diminish with predictability and familiarity, and they are fairly nonselective, triggered by sensory surprises that superficially resemble motivationally significant stimuli (Bromberg-Martin et al. 2010). This comports with the many studies that have found that the BG are integral to orientation of attention, and generation of responses, to motivationally relevant sensory stimuli (e.g. van Schouwenburg et al. 2010b; Cools et al. 2004; Leventhal et al. 2012).
In vitro studies on PFC pyramidal neurons have found that DA raises their excitability (Penit-Soria et al. 1987; Shi et al. 1997; Yang and Seamans 1996). Similarly, in the thalamic MD nucleus, DA has been shown in vitro to raise sensitivity to afferent activity (Lavin and Grace 1998). DA release in the MD largely derives from direct appositions arising from the VTA; indeed neurons in the VA and VL nuclei are also directly targeted by the midbrain DA centers (VTA, SNc, and RRF), as are the midline nuclei (Sánchez-González et al. 2005). D2 receptors are found throughout the associative thalamus (Rieck et al. 2004), and while DA terminals only sparsely synapse on neurons in the intralaminar thalamus (Sánchez-González et al. 2005), D2 receptors in the CM, PF, PC, and CL nuclei are particularly dense (Rieck et al. 2004), suggesting a large role there for volume-conducted DA action, with correspondingly less spatiotemporal specificity.
Following observations of treated and untreated parkinsonian primates, human and non-human, it has been proposed that DA has a decisive role in the regulation of global beta synchrony in BG, with increases in DA providing for narrowly focused striatal responses to cortical beta activity and consequent facilitation of action, while decreases in DA promote broad propagation of cortical beta, concomitant global beta synchrony, and the retarding or arresting of action (Jenkinson and Brown 2011; Magill et al. 2001). As noted earlier (§4.15), the DA-depleted striatum is characterized by the spontaneous and pervasive formation of synchronized clusters of SPNs (Humphries et al. 2009).
A pattern of broad beta synchrony, focally disrupted in association with performance of rewarded tasks, has been found in healthy (non-parkinsonian) monkeys (Courtemanche et al. 2003). These patterns appear to be DA-dependent: In an experiment in which global DA levels were manipulated to ~500% and <0.2% of their natural baseline, the low-DA condition was accompanied by pervasive synchrony with locally prevailing LFP, while the high-DA condition showed widespread focal desynchronization from prevailing LFP in primary motor cortex and dorsolateral striatum (Costa et al. 2006). DA manipulation was not found to affect overall cortical firing rates, underscoring the primacy of synchrony (and not rate) in these dynamics. The pattern of the hyperdopaminergic condition resembles the “desynchronization” of focally synchronized gamma oscillations in activated thalamocortical ensembles (Steriade et al. 1996), which according to the BGMS model often involve synchronized oscillations propagating focally through the BG.
Recent evidence suggests that DA acts to shift the size of responding SPN ensembles, rather than the rate of discharge of individual SPNs; acute DA blockade halves and triples the number of responding direct and indirect path SPNs respectively, generating a strong imbalance in favor of the indirect pathway, and significantly impairing spontaneous locomotion (Maltese et al. 2019‡). Earlier (§4.17), I suggested that SPNs cast votes for decisions, which are then tallied by downstream structures. According to this narrative, the effect of DA on the striatum can be viewed as expanding or contracting the pool of votes. The pool of support for striatal decisions, which in this view occur when ensembles of SPNs activate in response to synchronized inputs, is thus dynamically broadened or narrowed according to upward and downward modulations of DA supply, respectively.
Injection of DA into PFC has been seen to induce a spontaneous increase in synchrony between PFC and hippocampal LFPs, and to starkly alter the dynamics of PFC pyramidal neurons; activity shifts from in-phase with reciprocally associated interneurons (suggesting interneuronal inhibition) to opposite phase (suggesting interneuronal augmentation) (Benchenane et al. 2010). These effects of DA injection on PFC-hippocampal synchrony and PFC pyramidal neuron dynamics mimicked those seen without DA injection, in a well-trained behavioral task (Y maze navigation), at the choice point (the fork). DA released upon well-predicted reward, by inducing synchronization of PFC-hippocampal cell assemblies, might assure that effective behaviors are committed to long term memory, while ineffective ones are not (Benchenane et al. 2011). Naturally, counterproductive behaviors must also be remembered as such, implicating DA release associated with general saliency (Bromberg-Martin et al. 2010).
As noted earlier (§6.8), DA in PFC has been found to attenuate receptivity to inputs on L1 apical dendrites (Yang and Seamans 1996), and to depress GABAergic lateral interactions among L2/L3 interneurons (Towers and Hestrin 2008), reducing the spatiotemporal coherence of oscillation there. As DA level rises, PFC neurons may thus become progressively less affected by superficial inputs from the BG-recipient thalamus and corticocortical feedback paths, so that effective behaviors are protected from disruption and distractions, and in particular, from induction of empirically extraneous functional connectivity. Indeed, DA release in PFC is suggested to stabilize working memory items there (Gruber et al. 2006). The effect of DA release on cortex may extend well beyond directly DA-recipient frontal cortex: an integrative theory has been proposed by van Schouwenburg et al. (2010a) and Bloemendaal et al. (2015) that DA release in PFC induces it to influence interconnected posterior cortex to stabilize goal-relevant representations and protect them from distractions, even while DA release in the BG promotes flexible adaptive responses to new information. Consistent with these accounts, evidence from resting state fMRI studies in DA-manipulated humans suggests that local activity and large scale functional networks are stabilized and reinforced by systemic DA elevation, while systemic DA depletion results in elevated variability of local activity, and the dissolution of large scale networks, particularly impacting between-module connectivity while largely sparing within-module connectivity (Shafiei et al. 2019).
The ACh supply for the cortex and thalamus arises from the basal forebrain, particularly the NBM, and from the PPN and LDT nuclei in the brainstem reticular activating system. Comprehensive direct cholinergic projections from the NBM to cerebral cortex (Mesulam et al. 1983; Mesulam 2004) are posited to modulate the predisposition of the targeted areas to robust afferent-driven oscillation, with fine spatiotemporal specificity (Muñoz and Rudy 2014). Each individual neuron in the NBM projects to a single small area of cortex confined to a diameter of 1-1.5 mm, prompting the proposal that the cholinergic population of the NBM is arranged to give arbitrary addressability of small areas of cortex, permitting activation of complex constellations subserving specific functions (Price and Stern 1983). fMRI evidence in humans suggests involvement of the NBM in the general orchestration of large scale cortical network dynamics, implicating both cholinergic and non-cholinergic projections (including coreleased glutamate and GABA) (Markello et al. 2018). The NBM's projections to TRN further position it to exert a wide-ranging influence over corticothalamic activity (Levey et al. 1987). BG control of the NBM is detailed below (§9.11), including profuse innervation of all sectors by the ventral striatum (Mesulam and Mufson 1984; Grove et al. 1986; Haber et al. 1990; Haber 1987).
The PPN and LDT have wide-ranging subcortical cholinergic projections, comprehensively innervating the thalamus, including its reticular nucleus (Hallanger et al. 1987; Satoh and Fibiger 1986; Steriade et al. 1988; Paré et al. 1988; Lavoie and Parent 1994). PPN targeting of the thalamus includes its primary sensory nuclei—the dorsolateral geniculate (DLG), medial geniculate (MG), and the ventrobasal complex (ventral posterolateral (VPL) and ventral posteromedial (VPM)) (Hallanger et al. 1987). It additionally projects densely to the NBM and nearly all BG structures (Lavoie and Parent 1994).
Underscoring their functional significance, these cholinergic supply centers have prominent roles in disease processes. PPN lesions result in akinesia, and PPN degeneration is associated with PD (Pahapill and Lozano 2000). Alzheimer's disease is associated with attrition of the magnocellular cholinergic population in the NBM, typically to less than 30% of normal (Arendt et al. 1983). In Sz, the concentration of choline acetyltransferase in PPN and LDT is markedly lower than normal, while the concentration of nicotinamide-adenine dinucleotide phosphate (NADPH) diaphorase appears to be roughly twice normal (Karson et al. 1996; German et al. 1999). Indeed, systemic cholinergic abnormality may be a frequent correlate of Sz, and atypical antipsychotics such as clozapine and olanzapine have a high affinity for muscarinic receptors (Raedler et al. 2006; Scarr and Dean 2008).
If the tonic supply of ACh to a cortical locus is interrupted, neurons there become dramatically less sensitive to their excitatory afferents, and correspondingly more prone to tonic synchrony with their neighbors; ACh modulates the propensity of these neurons to track high frequency afferent oscillation and generate corresponding efferent oscillation, particularly in the beta and gamma bands (Rodriguez et al. 2004). Phasic increase in ACh supply to an area, when coupled with afferent activity, induces profound plasticity within tens of minutes, persistently elevating the propensity of the targeted area to synchronize with afferent high frequency oscillation and consequently desynchronize with neighboring tonic oscillation (Rodriguez et al. 2004).
Most of the brainstem diffuse modulatory systems may act on cortex indirectly through the NBM ACh and raphe 5-HT systems; cortical electrocorticographic (ECoG) activation can be completely abolished by concurrent blockade of ACh and 5-HT (Dringenberg and Vanderwolf 1997, 1998). Rats subjected to this concurrent blockade, and exhibiting complete loss of ECoG activation, nonetheless engage in active locomotion, with normal posture and open eyes; however their behavior is disorganized and aimless like that of decorticated rats, including repeated, unhesitating walking plunges over precipices, and insensate behavior in swim-to-platform tests (Vanderwolf 1992).
In cortex, ACh is modulatory, neither excitatory nor nonselectively disinhibitory; its presynaptic release does not by itself induce postsynaptic activity (Sillito and Kemp 1983). When coupled with excitatory afferent activity, ACh has a dramatic facilitatory effect on most cortical neurons, while maintaining or narrowing their respective receptive fields; tonic activity (discharges attributable to background afferent activity) is also reduced, so the overall effect is a marked increase in signal/noise ratio (Sillito and Kemp 1983).
The effect of ACh on cortical interneurons is more diverse, with fast spiking inhibitory (FSI) interneurons in L5 hyperpolarized via muscarinic receptors, disinhibiting the L5 pyramidal neurons they target, while low threshold spiking (LTS) inhibitory interneurons are excited via nicotinic receptors, raising inhibitory output to their more superficial targets in L1-L3 (Xiang et al. 1998).
Cholinergic hyperpolarization of cortical FSIs may relax the coincidence detection window for perisomatic inputs to pyramidal neurons (Pouille and Scanziani 2001), effectively increasing their receptive field, even while the direct effect of ACh on them is a narrowing of their receptive fields as described above. Moreover, the coherent lateral spread of oscillatory activity in L2/L3 (Tamás et al. 2000) may be depressed by ACh hyperpolarization of FSIs (as by DA (Towers and Hestrin 2008)), spatially focusing activity in cortex. In toto, these effects appear to stabilize working memory attractor networks (Qi et al. 2019‡).
It has been shown in behaving rats that short latency ACh release, through effects mediated by a diversity of receptor types, is crucial to the generation and synchronization of performance-correlated oscillation in PFC (Howe et al. 2017). In task trials in which the animal detected a sensory cue, significantly elevated PFC ACh levels were detected within 1.5 s of cue presentation, and remained elevated until reward delivery. Gamma oscillation in the same area, measured by LFP, was found to be significantly elevated, at ~90 Hz from ~200-400 ms after cue presentation, then at ~50 Hz from ~400-1300 ms after the cue. Local infusion of an M1 muscarinic antagonist attenuated these gamma responses in trials in which the animal detected the cue, and was associated with a trend toward more missed cues. Infusion of a nicotinic antagonist attenuated the initial high gamma response to detected cues, and similarly had no effect on oscillatory power in trials in which the animal missed the cue. Detected cues, but not missed cues, were associated with significant cross-frequency coupling of the 50 Hz gamma response, to local theta oscillation detected by LFP. This coupling was abolished by infusion of the M1 antagonist, and was attenuated by the nicotinic antagonist.
The effects of ACh on thalamic neurons have been found to be similar to those in cortex, facilitating responsiveness of excitatory neurons to afferent activity via M1 and M3 muscarinic receptors, as well as via nicotinic receptors, and having an opposite effect on inhibitory interneurons, where it induces hyperpolarization via M2 receptors, indirectly facilitating responsiveness (Parent and Descarries 2008; Steriade 2004). The cholinergic projection to PF (representing intralaminar nuclei) densely terminates in exclusively direct synapses (Parent and Descarries 2008), and PPN/LDT stimulation in the anesthetized cat, causing cholinergic activation of the thalamus, produces sustained, synchronized high frequency oscillation in intralaminar neurons and reciprocally connected cortical neurons, resembling patterns seen in the waking and REM sleep states (Steriade et al. 1996).
The terminal pattern of the cholinergic projection to the dorsal lateral geniculate nucleus (DLG, representing primary sensory nuclei) is almost entirely extrasynaptic (Parent and Descarries 2008), and this relatively diffuse pattern is likely to have markedly less spatiotemporal specificity than synaptic paths, so the diffuse ACh innervation of DLG comports with the expectation (according to the “binding by synchrony” hypothesis, briefly discussed later (§10.2)) that modulatory inputs to early sensory areas are arranged to not disrupt the fine time structure of activity therein. ACh inputs to the TRN are both synaptic (Parent and Descarries 2008) and extrasynaptic (Pita-Almenar et al. 2014), and are reported to hyperpolarize TRN neurons through M2 muscarinic receptors, disinhibiting their targets in the thalamus (Steriade 2004; Lam and Sherman 2010).
ACh has a variety of effects on striatum, through a variety of receptors: it can directly induce SPN depolarization and spontaneous firing, and in particular, facilitate the excitability of NMDA (glutamate) receptors on SPNs, while simultaneously reducing glutamate and GABA release; corticostriatal long term potentiation (LTP) in SPNs is also dependent on ACh activation of M1 muscarinic receptors (Calabresi et al. 1998, 2000). As noted earlier (§3.5), when intrinsic cholinergic interneurons in the striatum are subjected to synchronous spike volleys, their cholinergic action on dopaminergic axons promotes intrinsic DA release in the striatum (Threlfell et al. 2012).
Early experiments entailing injection of cholinergic agents into striatum, pallidal segments, and STN, showed dysregulatory effects that generally appeared to be pathological activations (DeLong and Georgopoulos 2011).
It has been proposed that the PPN, briefly discussed earlier (§8.4), is so intimate with the BG as to constitute an inextricable component thereof (Mena-Segovia et al. 2004). The GPi, VP, and SNr strongly and systematically project high velocity axon collaterals to it (Semba and Fibiger 1992; Grofova and Zhou 1998; Haber et al. 1985; Parent et al. 2001; Harnois and Filion 1982), and cholinergic and glutamatergic cells in the PPN in turn profusely target dopaminergic cells in the SNc, with at least some of the PPN cells that target SNc receiving projections from SNr (Grofova and Zhou 1998). However, BG regulation of the PPN cholinergic supply to thalamus is complex, apparently largely indirect, and yet to be fully elucidated. BG projections to PPN have been reported to preferentially target non-cholinergic cells (Mena-Segovia and Bolam 2009), and the BG may reciprocate preferentially with the rostral sector of the PPN, while it is the caudal PPN that projects to the thalamus and tectum (Martinez-Gonzalez et al. 2011). However, it has also been reported that the GPi projects throughout PPN, most prominently to the central PPN (Shink et al. 1997), which in turn projects to the NBM (Lavoie and Parent 1994). Moreover, caudal PPN is targeted by the DRN, which itself is targeted by the BG, though the effect of 5-HT on the PPN is complex and unresolved (Vertes 1991; Steininger et al. 1997; Martinez-Gonzalez et al. 2011).
The ventral striatum projects profusely to all sectors of the NBM (Mesulam and Mufson 1984; Grove et al. 1986; Haber et al. 1990; Haber 1987), and the NBM receives substantial projections from the SNc and VTA, targeting cholinergic neurons (Záborszky and Cullinan 1996; Gaykema and Záborszky 1997). At least some VS afferents to NBM terminate directly on corticopetal cholinergic neurons; GABA input to these neurons is posited to dampen excitability, resulting in corresponding inattention in their cortical targets (Sarter and Bruno 1999). The GPe, like the NBM, but much less profusely, has direct cholinergic projections to cerebral cortex (Eid and Parent 2015), and both coexpress GABA in these projections (Saunders et al. 2015a, 2015b). And the GPe, like the NBM, projects directly to the TRN. The NBM may be an inextricable component of an extended BG system, as has been suggested of other areas of the substantia innominata (Heimer et al. 1997). Indeed a model has been proposed that integrates ACh projections from the NBM, the GPe, and the VP, with BG loop circuitry (Záborszky et al. 1991, Fig. 6).
Noradrenaline (NA) originating in the locus coeruleus (LC) of the pontine tegmentum is implicated in the direct modulation of arousal throughout the forebrain; the LC responds to noxious, novel, and other highly salient stimuli, toward which attention is to be oriented, with low latency phasic responses time-locked to the stimulus (Berridge 2008; Sara and Bouret 2012). These phasic responses are posited to reset network connectivity to facilitate assembly of a new network oriented to the salient stimulus, and there is evidence that NA arising from LC has a more general role in set shifting, crucially implicating the reciprocal connectivity of LC with PFC (Sara and Bouret 2012).
However, the striatum is not an LC target (Aston-Jones and Cohen 2005), and descending inputs to the LC have been found to be highly restricted, excluding most BG and all thalamic structures; activation of LC by afferent activity has been found to be either generalized to its entirety, or generalized to an entire sensory domain; perhaps most tellingly, output from the LC has been found to be non-specific, with efferent populations in LC distributed throughout its extent, and only modest and partial segregation according to target structure (neocortex, thalamus, cerebellum, etc.) (Aston-Jones et al. 1986; Waterhouse et al. 1993; Loughlin et al. 1986).
Thus, while the LC is integral to the regulation of oscillatory activity and functional connectivity in the thalamocortical system, it seems clear that the LC is nonspecific in its mechanisms. It also seems clear that it is not substantially integrated into BG circuitry, notwithstanding evidence of a sparse projection from the ventral pallidum to rostral LC (Groenewegen et al. 1993). It seems likely that stimulus-related network formation facilitated by LC reset signals entails broad synchronies to which the striatum responds after the fact.
5-HT supply to the telencephalon arises from the MRN and DRN, which project strongly to the midline, intralaminar, and mediodorsal thalamic nuclei, much of the BG, and to the entirety of cerebral cortex and the medial temporal lobe (Lavoie and Parent 1990; Vertes 1991; Vertes et al. 1999; Baumgarten and Grozdanovic 2000). Raphe projections exhibit complex specificity, with the DRN projecting to cortex with various topographies, while the MRN projects to cortex more diffusely (Wilson and Molliver 1991).
5-HT has an effect on its cortical targets much like that of ACh, facilitating responses to afferents, yielding ECoG desynchronization (Neuman and Zebrowska 1992), though the effect on individual neurons is complex, with most cells depolarized via 5-HT2 receptors but some hyperpolarized via other receptors (Davies et al. 1987).
5-HT2A receptors are present on the apical dendrites of L5 pyramidal neurons, so 5-HT release facilitates responsiveness (Carter et al. 2005) precisely where it is inhibited by DA and ACh release. This effect apparently counteracts the posited focusing and stabilizing effects of DA and ACh described above; indeed almost all known hallucinogenic drugs act through this channel, and activation of 5-HT2A receptors is necessary and sufficient for their hallucinogenic effects (Glennon et al. 1984; González-Maeso et al. 2007; Fiorella et al. 1995; but see Maqueda et al. 2015).
The notion arising from the BGMS model is that 5-HT2A agonists (even including, rarely, SSRIs for treatment of never-before-hallucinating patients (Bourgeois et al. 1998; Waltereit et al. 2013)) open cortical columns to induction of effective connections via spike-timing-dependent gain control by corticocortical feedback and BG-thalamocortical output, and hallucinogens thereby induce spurious information flow and associations that would not normally reach the implicated pyramidal somata. Evidence suggests that psychedelic facilitation of spurious effective connections is not uniform, but rather entails abnormal enhancement of connectivity in sensory and somatomotor areas, simultaneous with abnormal attenuation of connectivity in associative areas, including the default mode network (Preller et al. 2018). This bears a striking resemblance to the large scale network dysconnectivity characteristic of Sz (Ji et al. 2019; Giraldo-Chica et al. 2018).
Though the dysconnectivity of Sz may principally or frequently be rooted in GABAergic and dopaminergic dysfunction (discussed at greater length later (§12.5)), there is also a suggestion of 5-HT dysfunction (Geyer and Vollenweider 2008). Atypical antipsychotics such as clozapine, risperidone, and olanzapine show much higher affinity for 5-HT2 receptors, which they usually occupy almost completely, than for the D2 receptors targeted by earlier antipsychotics such as haloperidol (Kapur et al. 1999). Beyond this, common direct BG involvement is plausible. 5-HT2C receptors in the striatum, activated by hallucinogens (Fiorella et al. 1995), have been found to excite striatal FSIs (Blomeley and Bracci 2009), and direct striatal involvement in Sz has been posited (Graybiel 1997; Simpson et al. 2010; Wang et al. 2015).
All parts of the BG are innervated serotonergically by the raphe nuclei, with heterogeneous density within and between the organs of the BG, and highest density in the SN and GP (Lavoie and Parent 1990). The median and dorsal raphe nuclei (MRN and DRN) are targeted by the VP, SNr, and VTA (Peyron et al. 1997; Gervasoni et al. 2000; Levine and Jacobs 1992; Groenewegen et al. 1993). Coupling with BG DA centers and DA control structures is extensive. The VTA projects to the DRN and MRN; the DRN and MRN also project to DA cells in the SNc and VTA, and raphe projections to the SNr appear to be directed to the dendrites of DA neurons (Baumgarten and Grozdanovic 2000). The lateral habenula (LHb) projects strongly to all parts of the DRN (Peyron et al. 1997) and to the MRN (Herkenham and Nauta 1979), while the MRN projects massively throughout the extent of LHb (Vertes et al. 1999) and the DRN shows light but distinct targeting of LHb (Vertes 1991). The LHb is integral to BG DA circuitry—it is reciprocally linked with the VTA, directly and via the rostromedial tegmental nucleus (RMTg) (Herkenham and Nauta 1979; Hikosaka 2010; Balcita-Pedicino et al. 2011), and is profusely innervated by GPi and VP (Parent et al. 2001; Hong and Hikosaka 2008; Shabel et al. 2012; Groenewegen et al. 1993).
The MRN and DRN project densely to the PPN and LDT, and the DRN projects densely to the substantia innominata (including NBM, in primates) (Vertes 1991; Vertes et al. 1999; Steininger et al. 1997). The substantia innominata in turn projects to the DRN (Peyron et al. 1997), and PPN and LDT project to MRN and DRN (Semba and Fibiger 1992). The central 5-HT and ACh systems are thus directly and reciprocally coupled.
As noted above (§9.6), the NBM projection to cortex is comprehensive and topographically organized. Single loci in NBM project jointly and specifically to interconnected areas of cortex, particularly frontal and posterior areas (Pearson et al. 1983; Ghashghaei and Barbas 2001; Záborszky et al. 2015). These loci of joint projection appear to entail distinct intermingled populations, with only a tiny minority (~3%) of cells collateralizing to both frontal and posterior areas (Záborszky et al. 2015), suggesting combinatorial flexibility. The raphe nuclei, particularly the DRN, are reported to exhibit similar organization, with small groups of dorsal raphe cells projecting to widely distributed, anatomically interconnected neocortical foci (Wilson and Molliver 1991; Molliver 1987). Evidently, these patterns of divergence are much like those of the thalamocortical projection, described earlier (§1.6).
Direct and dense projections from PFC and other frontal cortical association areas to the NBM (Mesulam and Mufson 1984), PPN, and LDT (Semba and Fibiger 1992) thence to cortex and thalamus is a putative mechanism for sustained attention and inattention (Sarter et al. 2001; Záborszky et al. 1997). Indeed, PFC inactivation completely abolishes sensory-evoked ACh release in the sensory thalamus, and significantly reduces tonic ACh release in sensory cortex (Rasmusson et al. 2007). PFC projections to the DRN (Gonçalves et al. 2009) and LC (Jodoj et al. 1998; Aston-Jones and Cohen 2005) are thought to have similar and related functions. Because PFC is thoroughly and densely targeted by BG output via the thalamus and the midbrain DA centers, and projects directly and strongly to all BG input structures, PFC control of cholinergic and serotonergic centers implies BG influence on them, and suggests further coordination of output from these modulatory centers with BG output.
The TRN, through GABAergic projections to other thalamic nuclei, is thought to act in a modulatory role, influencing activity and oscillations in the entire thalamus and cortex, particularly corticocortical functional connectivity (Pinault 2004). A crucial role in the generation of spindles during sleep is recognized (Contreras et al. 1997). Prefrontal projections to TRN are thought to play a prominent role in orientation of attention and suppression of distractors (Zikopoulos and Barbas 2006; Guillery et al. 1998), and dysfunction of the TRN, resulting in deficits in these and related functions, has been associated with Sz (Ferrarelli and Tononi 2011; Pinault 2011).
The GPe projects to the full rostrocaudal extent of the TRN (Hazrati and Parent 1991; Shammah-Lagnado et al. 1996), and this projection has been directly implicated in attentional control (Nakajima et al. 2019). Given evidence that some circuits through the TRN are open loops implicating more than one cortical area (Brown et al. 2019‡), this projection is positioned to directly modulate inter-areal cortical signaling. BG inputs to the TRN also target cells that project to the intralaminar thalamus (Kayahara and Nakano 1998), and experiments in vitro suggest that DA release in GPe inhibits its inputs to TRN (Gasca-Martinez et al. 2010), suggesting that these inputs conform to functions identified for the indirect path, dampening or disconnecting activity. Additionally, glutamatergic nigroreticular projections have been demonstrated arising from striatum-recipient cells throughout the SN, both from the pars reticulata and the pars compacta, with roughly half of these fibers also found to release DA (Antal et al. 2014).
The interposition of the thalamic reticular nucleus in collaterals of L6 corticothalamic projections (Deschênes et al. 1994) is posited to produce nonlinearity, such that low frequency activity has a suppressive influence on thalamus via the TRN, while higher frequency activity is stimulative (Crandall et al. 2015). Modulation of the TRN (by the BG and PFC, in particular) might alter this dynamic, providing for adjustment of the threshold above which cortical activity stimulates activity in BG-recipient thalamus, and below which it is suppressive. This would gate the action of the BG on cortex, by controlling the supply of activity available for modulation at the implicated thalamocortical neurons. The BG and PFC are arranged to control this gate by adjusting the ACh supply to TRN, reducing or abolishing the suppressive influence of the TRN on corticothalamic targets (Lam and Sherman 2010).
The BG have been proposed to function in perceptual decision making in a fashion analogous to their function in behavioral decision making (Ding and Gold 2013). As reviewed earlier (§6.1), BG direct path output is arranged to influence activity not only in frontal cortex, but in posterior areas, including posterior sensory areas. Pathways described earlier (§9) by which the BG modulate central DA, ACh, and 5-HT supplies, and the TRN, imply a broad modulatory influence of the BG on sensory processing.
Motor control, long associated with the BG, has an inherent intimacy with attention, which entails selective perception; for example, common mechanisms and networks have been identified underlying attention and oculomotor control, both within and beyond the BG (Corbetta et al. 1998; Hikosaka et al. 2000). There is evidence that striatal activations continually track visuospatial attentional orientation, even in the absence of saccades and other overt actions (Arcizet and Krauzlis 2018).
When isochronous rhythmic visual events are presented to monkeys, in a task requiring saccades when oddballs are encountered in the sequence, firing of some striatal neurons is strongly entrained to the stimulus rhythm, and missing-stimulus oddballs evoke even stronger isochronous responses from those neurons, suggesting that perception of such phenomena involves cortico-basal ganglia ensembles (Kameda et al. 2019). Indeed, structural asymmetries of the globus pallidus correlate with alpha oscillatory power asymmetries in the visual cortex (Mazzetti et al. 2019), and activity in the BG-recipient central thalamus shows significant attention-related, performance-correlated upward shifts in power spectra (Schiff et al. 2013).
Primary sensory areas of the thalamus, and thalamic sensory areas in intimate topographic registration with the primary areas, are apparently avoided by direct path output (Percheron et al. 1996; Parent et al. 2001). This arrangement is an expected corollary of the proposed “binding by synchrony” mechanism (von der Malsburg 1999; Singer and Gray 1995; Womelsdorf et al. 2007; Jia et al. 2013; Barth and MacDonald 1996; Siegel et al. 2008). Rigid BG-induced spike timing disruption of sensory processing pipelines at the thalamic level would derange the spatiotemporally precise registration by which ensembles of neurons representing a stimulus are proposed to be coherently bound together, and to be differentiated from neurons in the same area that are active but not associated with the stimulus. In fact, there is evidence for binding by synchrony in sensory nuclei of the thalamus; corticothalamic projections bearing synchronized oscillations associated with a visual stimulus entrain thalamocortical activity associated with that stimulus, increasing the effective neuronal gain for associated features (Sillito et al. 1994).
While the BG direct path output apparently avoids sensory thalamic nuclei, it does not avoid sensory areas at the cortical level. As noted earlier (§6.1), the BG-recipient intralaminar nuclei (CL, PC, CM, and PF) have been shown to project to visual, auditory, and somatosensory cortex (Van der Werf et al. 2002; Scannell 1999), and the MD, VA, and VL nuclei have been shown in primates to project to visual association cortex and the angular gyrus area of parietal cortex (Middleton and Strick 1996; Clower et al. 2005; Tigges et al. 1983). In the rostral intralaminar thalamus, the PC and CL nuclei show distinct intimacy with sensory areas, reaching all visual areas but the primary receptive fields; these projections show no apparent topographic pattern, but are accompanied by heavy projections to densely interconnected areas such as the frontal eye fields and posterior parietal association areas 5 and 7, with some axons found to collateralize multi-areally, e.g. to visual area 20a and areas 5 and 7 (Kaufman and Rosenquist 1985a; Van der Werf et al. 2002).
As noted earlier (§7.10), the caudal intralaminar nuclei in cat project to secondary and some associative auditory cortex, but avoid the primary, posterior, ventroposterior, and temporal auditory fields (Scannell 1999). This extensive lacuna suggests that the early stages of auditory processing are particularly sensitive to disruption of spike patterns. This might be attributable to the unique orientation of auditory perception to environmental phenomena (sounds) that are typically oscillatory and momentary, so that the crucial phenomenological attributes of stimuli can only be represented in early processing stages with neural spikes that are precisely locked in time to occurrence of those attributes. This representation may take the form of periodic codes, patterned to precisely reflect stimulus periodicity, which by coherent propagation and integration can enable precise sound localization (Brown and Curto 2019‡). Similar periodic coding principles might prevail in the grid cells of the hippocampal system (Brown and Curto 2019‡), which is similarly free of direct BG inputs, as outlined later (§13.2.4).
The projection systems associated with the GPe, SN pars lateralis (SNl), PPN, LDT, NBM, and DRN extensively target sensory areas, including those in thalamus. As noted earlier (§9.6), the PPN targets the primary sensory nuclei of the thalamus (Hallanger et al. 1987). The caudal GPe projects directly to the auditory and visual sensory sectors of the caudal TRN, to auditory cortex, the inferior colliculus, and through the SNl further influences visual and auditory processing via the latter's projections to the superior and inferior colliculi (Shammah-Lagnado et al. 1996; Moriizumi and Hattori 1991; Yasui et al. 1991). Projections of GABAergic cells in the NBM to TRN target the vision-specific portion of the latter, while cholinergic cells in NBM project to corresponding visual cortex (Bickford et al. 1994). While paths through the DRN from BG output structures to sensory cortex have yet to be directly demonstrated, the DRN comprehensively innervates cortex (Vertes 1991) and, as reviewed earlier (§9.15), is reciprocally coupled to the BG.
The superior colliculus is a key center for sensory (particularly visuospatial) processing: it is implicated not only in ocular saccades but in covert (i.e., non-motoric) orienting of attention (Robinson and Kertzman 1995), supplies powerful inputs to thalamic MD and pulvinar nuclei (Wurtz et al. 2005; Stepniewska et al. 2000) thence to visuocognitive cortex (Berman and Wurtz 2010; Lyon et al. 2010), and is under the direct influence of the SNr (Hikosaka and Wurtz 1983). The SC also projects to the thalamic intralaminar nuclei, and through them, the striatum, where it has strong attentional effects (Herman et al. 2019‡). The SC and BG thus influence each other strongly and reciprocally.
It is quite intriguing that BG output avoids the pulvinar, but extensively innervates the SC, given that the SC extensively innervates the pulvinar. Perhaps this relates to the proposed imperative to avoid deranging fine timing information in thalamocortical sensory modules. Nigrotectal terminals, while GABAergic, mostly appose medium or small dendrites (Behan et al. 1987); enveloping perisomatic GABAergic appositions like those of the nigrothalamic projection (Bodor et al. 2008) are present in the same tectal population, but arise elsewhere (Behan et al. 1987).
Nakajima et al. (2019) have demonstrated in mice that the BG are crucial mediators of corticothalamic regulation of inattention to distracting stimuli, building on earlier work substantiating roles for the TRN (Halassa et al. 2014; Wimmer et al. 2015) and PFC (Rikhye et al. 2018; Schmitt et al. 2017) in sensory attention mechanisms. In particular, they found that the prelimbic region of PFC induces inattention to distracting visual stimuli via a pathway through the dorsal caudal striatum, where projections from PFC and visual cortex converge, to the caudal GPe, thence to the visual sector of the TRN, thence to the primary visual thalamus (lateral geniculate body, LGB). They found that neglect of auditory distractors was mediated by a similar path ending in the medial geniculate body (MGB), and moreover, that in subtasks requiring auditory discrimination, auditory signal to noise ratio was executively enhanced via this PFC-BG-TRN-MGB path.
In healthy humans, speech production involves significant inhibition of responsiveness in auditory cortex to the sounds of self-produced speech; this inhibition is deficient in Parkinson's disease (Railo et al. 2019‡). This suggests that the BG are a key component of a mechanism whereby the organism avoids distraction by the perceptual correlates of successfully produced intentional behaviors. In short, the paths through the dorsal striatum to GPe, thence to TRN and thalamus, from the SNr to the SC, thence to thalamus, and from the ventral striatum to NBM, thence to cortex, may be crucial elements of a system that continually transforms behavioral output into selective, anticipatory inattention.
Corticostriatal input from primary motor cortex has been found to preferentially flow to the GPe (Wall et al. 2013), and corticostriatal input flowing to GPe is predominantly collaterals of axons destined for the pyramidal tract, bearing activity tightly correlated with executed motor commands (Lei et al. 2004; Morishima and Kawaguchi 2006). Collaterals of these axons also target the proximal dendrites of projection neurons in the intralaminar thalamic nuclei (Deschênes et al. 1998), from which these signals are relayed to all components of the BG. A key role posited for the signals carried by collaterals of motor output is as an “efference copy” or “corollary discharge”, primarily serving to contextualize sensory input, as suggested by projections from motor cortex to somatosensory cortex (DeFelipe et al. 1986). These signals may also serve to inform a system that differentiates between self-generated and other-generated percepts, attending the latter while disregarding the former (Crapse and Sommer 2008).
The BG might continually compute a dynamic template imparted upon the thalamus via the TRN and SC, and upon the cortex via the NBM, so that sensory input that is the expected result of motor output, and is therefore cognitively extraneous and distracting, is functionally disconnected. Sparsity in direct path corticostriatal input, but not in indirect path input (Turner and DeLong 2000), comports with particular involvement of the indirect path in mediating this continual expectation-driven inattention. GPe facilitation of the TRN might act particularly by raising the threshold for corticothalamic inputs to transition from inhibitory to excitatory effects on their thalamic targets (Crandall et al. 2015).
The BG influence the SC directly (Hikosaka and Wurtz 1983), and if an effect of this influence is to induce neglect of expected percepts, then unexpected percepts will be salient. This comports with evidence that the SC, upon encountering unexpected sensory events, can reconfigure sensorimotor orientation in the thalamus via the zona incerta (Watson et al. 2015).
These arrangements suggest a combined dynamic in which the BG learn to minimize surprise, centrally implicating dopamine signaling (Schultz 1998, 2013), as noted earlier (§9.2). Indeed, some highly abstract models of cognition imply that the minimization of surprise is an organizing principle for the brain as a whole (Friston 2010), given evidence that feedback projections bear predictions, while feedforward projections bear “newsworthy” prediction errors (Friston 2018; Rao and Ballard 1999).
It has also been suggested that BG-mediated selection of an action jointly activates areas implicated in processing the expected perceptual correlates of that action (Colder 2015; but see Urbain and Deschênes 2007). The combined dynamic might consist of activation and effective connection of the executive and perceptual areas implicated in the action, culminating in execution of the action, whereupon an efference copy of the corticofugal motor output follows the paths through the BG to the TRN and NBM described here, in addition to corticocortical paths. By this narrative, if the action has the expected result, TRN and NBM outputs mask out the associated sensory inputs, presumably with a crucial role for collaterals of motor cortex output projecting to somatosensory areas of cortex and thalamus. If the results of the action deviate from expectations, then sensory inputs associated with the deviation are not masked out, but act as bottom-up drivers with particular salience due to the anticipatory recruitment of the associated cortical perceptual areas. The disparity is thus efficiently signaled, facilitating remediation.
That the BG systematically adjust sensory perceptions to track expectations and minimize surprise is suggested by the finding that the ventral striatum and ventral pallidum mediate prepulse inhibition of the acoustic startle response (mild auditory stimulus presented 30-500ms before startling stimulus) (Kodsi and Swerdlow 1995). This prepulse inhibition is deficient in many diseases associated with the BG, including OCD, Huntington's disease, and GTS (Swerdlow and Geyer 1998), is attenuated in Sz (Swerdlow and Geyer 1998; Quednow et al. 2008; Geyer and Vollenweider 2008), and is altered by hallucinogenic drugs (Vollenweider et al. 2007; Geyer and Vollenweider 2008). Anticipatory inhibition of the cortical response to self-generated speech is deficient not only in PD (Railo et al. 2019‡), but also in Sz and bipolar disorder (Ford et al. 2013).
There is evidence that corollary discharge underlying self-other differentiation is generally dysfunctional in Sz (Ford et al. 2001). Even basic coordination of motor output with sensory input is affected: smooth tracking of moving objects with the eyes is consistently impaired in Sz patients and their close relatives (Levy et al. 1994). This might be explained by dysfunction of the mechanisms of anticipatory inattention, if internally caused and therefore predictable sensory events are given spurious salience, prompting inappropriate actions (e.g., saccades). Deficient performance in Sz on stimulus-antisaccade tasks might be similarly rooted in dysfunction of the mechanisms of executive inhibition (Fukushima et al. 1988).
The accurate differentiation of self-generated from other-generated effects is crucial in the cognitive representation of agency (intentional action), and its dysfunction is likely intrinsic to Sz (van der Weiden et al. 2015; Ford et al. 2007). Evidence of projections in higher primates from the MD, VA, and VL nuclei to the angular gyrus (Tigges et al. 1983) is also suggestive, as this cortical area has been shown in humans to have a role in awareness of action consequences, and in particular, in the detection of disparities between intentions and results (Farrer et al. 2008), suggesting extensive BG involvement in the dynamics of agency. Agency in itself seems to influence the perception of results relative to the actions that produced them: Subjective intentionality significantly shortens the reported delay between action and results, compared to delays reported in involuntary action scenarios (such as an experimenter tugging an appendage with a fabric loop), and this shortening is lessened when the sense of agency in the action is disrupted by hypnosis (Lush et al. 2017) or coercion (Caspar et al. 2016).
The BGMS model implicates the BG in the activation and coordination of large scale cortical networks spanning and pervading the sensory, motor, cognitive, and motivational domains. Evidently, general cognitive coordination is uniquely challenging in terms of combinatorial tractability. BG physiology reviewed earlier (§4) suggests that they are suited for such a role; here I explore this proposition more directly and in greater depth.
The connectedness of the cerebral cortex—the proportion of combinatorially possible direct long range links that are anatomically actualized—is quite high, at least 66% in macaques (Markov et al. 2014) and as much as 97% in mouse (Gămănuţ et al. 2018). 130-140 distinct cortical areas have been identified in the macaque (Markov et al. 2014), implying at least 5,500 (66% × (130 × 129) × ½) bidirectionally or unidirectionally interconnected pairs in each hemisphere and as many as 25,000 such pairs overall. More notionally, these figures imply at least 1078 (2130 × 2) distinct areal combinations, some substantial fraction of which might be both anatomically connected and usefully selected for momentary multi-areal effective connectivity. In humans, the number of distinct cortical areas is even larger, estimated at 180 (Glasser et al. 2016), implying more than 40,000 linkages and 10108 distinct areal combinations. The sheer scale of this network is also apparent in the estimated populations, with roughly 8 × 109 neurons and 6.6 × 1013 synapses in the human cerebral cortex, connected by roughly 10 million kilometers of axons (Murre and Sturdy 1995).
Implicit to the BGMS model is a proposal that the mammalian brain tames these combinatorial and population explosions with a mechanism combining spatiotemporally precise but mesoscopic connectivity selection with the stupendous topological flexibility of the BG. That the BG are in fact central to this facility is strongly implied by evidence that they are among the most connected regions of the brain (van den Heuvel and Sporns 2011), that they participate in a particularly wide variety of large scale synchronized networks (Keitel and Gross 2016), and that their dysfunction is directly associated with a graded contraction of the number of distinct areal combinations activated in the course of cognition (Sorrentino et al. 2019‡). The corticostriatal projection, with its prodigious convergence and divergence (Flaherty and Graybiel 1994; Hintiryan et al. 2016), a total synapse population in humans of roughly 1012 (Kreczmanski et al. 2007; Kincaid et al. 1998; Zheng and Wilson 2002), and uncorrelated postsynaptic activity (Wilson 2013), is quantitatively suited to play a key role grappling with the immense dimensionality of the neocortex.
Moreover, the unusual diversity of conduction delays through the BG, reviewed earlier (§3.3), may arrange for a population of “polychronous groups” of neurons that exceeds not only the number of neurons in the system, but perhaps even the number of synapses, providing for immense combinatorial coverage and dimensionality, and correspondingly stupendous representational capacities (Izhikevich 2006). In this view, each projection neuron in the striatum transiently participates in a vast array of contextually activated assemblies, each of which relates particular spatiotemporally dispersed inputs to corresponding learned spatiotemporally focused outputs, projecting to pallidal/nigral projection neurons, which in turn focally target the thalamus (and other areas). This arrangement is functionally homologous to an arrangement in the cerebellum, discussed at greater length later (§13.1), that is thought to provide for recognition of roughly 1082 distinct spatiotemporal patterns in the mouse (Sultan and Heck 2003). It is also similar in some important respects to arrangements of massive phased arrays of independent antennas, used and proposed for transmission and reception of signals in extremely flexible, high capacity, parallel, dynamically configurable communication links (Rusek et al. 2013) and, particularly, radar systems (Fuhrmann et al. 2010).
Open circuits through the BG and thalamus, originating in one region and projecting to another one distant from the first, have long been appreciated (Joel and Weiner 1994). A comprehensive corticostriatal projection, and massive convergence through the BG, arrange for selections to reflect global, system-level information, including large scale dynamic network topology. In the BGMS model, this global information informs cortical routing decisions, and indeed allows for establishment of efficient routes even between topologically remote areas, via connectivity hubs.
Within the global workspace model of cognition (Dehaene and Naccache 2001; Baars 2005; Dehaene and Changeux 2011; Baars et al. 2013), the BG might arbitrate ephemeral access to and by specialized processors, and more generally, “dynamically mobilize” cortical areas for effective connection within the long range distributed network of conscious cognition. Equivalently, in the dynamic core model (Tononi and Edelman 1998), the BG might determine from moment to moment which corticothalamic modules are functionally well-connected.
Hybrid metaheuristics, a combinatorial optimization technique, involves such arrangements (Blum et al. 2011): densely and broadly connected areas may constitute a generic problem-solving (metaheuristic) mechanism, while more specialized and less widely connected areas are selectively integrated with the generic mechanism, when their respective domains of expertise are relevant to problems for which conscious intervention has occurred. The BG figure prominently, because many of the cortical areas with the highest anatomical and functional connectedness—areas including the superior and lateral prefrontal, anterior cingulate, and medial orbitofrontal (Cole et al. 2010; van den Heuvel and Sporns 2011; Harriger et al. 2012; Elston 2000)—are particularly dense targets of BG output (Middleton and Strick 2002; Ullman 2006; Akkal et al. 2007). Indeed, the striatum itself has been found to contain the most connected brain regions by some measures (van den Heuvel and Sporns 2011), and to exhibit continual activation throughout the duration of a task, with shifting loci of activation that “tile” the span of the task, tracking context (Arcizet and Krauzlis 2018; Weglage et al. 2020‡).
Proponents of the global workspace theory of consciousness contemplate “auto-catalytic” organization of long range functional networks in cortex (Dehaene and Naccache 2001) to avoid a homuncular infinite regress (Dehaene and Changeux 2011). The BGMS model implies that these functional networks are self-organized by a coalition of cortical and subcortical mechanisms, with the BG in particular crucial to the selection and recruitment of cortical networks, and to the inhibition and isolation of areas not implicated in a selected network. The combined system of the PFC and BG has been proposed expressly to constitute a mechanistically complete explanation for coherent cognition, avoiding implications of a cognitive homunculus (Hazy et al. 2006). Curiously, the sensorimotor striatum contains many fragmentary sensorimotor homunculi in various configurations (Flaherty and Graybiel 1994), implying that the associative striatum contains many fragmentary cognitive maps, which through the looping architecture of the BG, might be said to regress indefinitely, if not infinitely (see more below, regarding perturbative iteration).
The arrangement of the BG to influence cortical activity mesoscopically, without driving activity directly, has inspired the view that they dynamically modulate state attractors, shaping the evolution of cortical activity (Djurfeldt et al. 2001). This view is particularly appealing in light of evidence and simulations that suggest that the cerebral cortex intrinsically maintains critical dynamics, supporting continually evolving dynamical activity (van Vreeswijk and Sompolinsky 1996; Haider et al. 2006; Okun and Lampl 2008; Ahmadian and Miller 2019‡; Ma et al. 2019; Haider et al. 2012; Rubin et al. 2015). Criticality in networks has been shown to provide for optimal controllability of those networks: slight gain reductions in such networks shift them toward modular isolation and preservation of state, while slight gain increases shift them toward connectedness and integration (Li et al. 2019). This has direct implications for the relationship of the BG to cortex, because they are positioned to adjust the gain of thalamocortical and corticocortical circuits.
Critical dynamics characterize the cortex in the awake but resting state, while attention focused on a task induces broad subcriticality, reducing susceptibility to distractors (Fagerholm et al. 2015). Task-specific output from connectivity hubs in cortex may be crucial in maintaining task-appropriate effective connectivity despite the widespread subcriticality that accompanies task-focused cognition (Senden et al. 2018), and there is ample evidence that the BG are integral to hub circuitry and dynamics (e.g. Vatansever et al. 2016; Averbeck et al. 2014; Assem et al. 2019‡). Moreover, BG output to brainstem and basal forebrain neurotransmitter centers, detailed earlier (§9), position the BG to orchestrate shifts in balance between criticality (while in resting states) and subcriticality (while engrossed in tasks), supplying task-specific facilitation through the direct path during task-focused, subcritical episodes.
Decisions may consist of dynamic and plastic reconfigurations of attractors in cortex that stabilize selected action representations against perturbations. Modeling of frontocortical neural activity underlying stimulus-response behavior in mice suggests that contextually appropriate response representations are actively stabilized against distractors during delay periods, with distance between attractor basins (associated with alternative actions) rising with the strength of the stimulus, and attractor basin depth increasing with learning (Finkelstein et al. 2019‡). Physiological evidence from human and non-human primates (fMRI and electrophysiologically measured spike rates, respectively) suggests that task performance entails global “quenching” of variability and long-range activity correlations, suppressing spontaneous activity, while associated computational modeling suggests that it entails stabilization of task-related attractors (Ito et al. 2019‡). These dynamics directly implicate the striatum. Modeling of the combined system of PFC and striatum in monkeys suggests they are central to stimulus-response learning, and that learning entails increasing the distance between, or the attractor basin depth of, the representations of alternative actions (Márton et al. 2019‡).
It has been proposed, under the rubric of “integrated information theory”, that consciousness in its essence is a series of selected states, each an ephemeral complex of informational relationships, within an internally well-connected system with massive dimensionality and the power to discriminate among the myriad possible states as wholes—in mammals, the system of the cerebral cortex and thalamus (Tononi 2004). The physical aspects of the brain directly implicated in consciousness are proposed to be those with maximal cause-effect power (Tononi et al. 2016). By these criteria, the global integration of cortical states implicit to the massive convergence of the BG, and predominant BG control of spike patterns in many of their thalamocortical targets, signify a central role in consciousness. Indeed, the integrated information measure Φ, evaluated for activity in a network spanning deep layers of lateral intraparietal cortex, caudate striatum, and intralaminar thalamus, is a particularly reliable indicator of state of consciousness (anesthesia, NREM sleep, resting wakefulness, or anesthesia interrupted by thalamic stimulation) (Afrasiabi et al. 2020‡).
Conscious cognition, and prefrontal cortex, are thought to be crucial for high level supervision, particularly the goal-motivated resolution of problems not resolved at lower (more local) levels (Dehaene and Naccache 2001; Miller and Cohen 2001). Evidence supports this proposition, and implicates the BG in these dynamics. It has been shown in rats that the learning of new goal-directed actions, adapted to changing scenarios, depends on the connection from PFC to striatum (Hart et al. 2018). In primates, the anterior cingulate cortex (ACC) and DLPFC are thought to subserve detection and mitigation of conflicts, through an iterative looping arrangement (Carter and van Veen 2007). As noted above, these areas densely reciprocate with the BG, with the ACC particularly targeting striosomes (Eblen and Graybiel 1995). As a dense target of mesencephalic DA, ACC has been proposed to form a loop with the BG subserving conflict management (Holroyd and Coles 2002). Evidence from electrophysiology in humans suggests that a key role of the ACC in conflict management is the selective amplification of useful task-relevant information in widely distributed functional constellations (Ebitz et al. 2020‡), which comports neatly with heavy projection of ACC to the striatum—to striosomes, through which cognitive salience might be selectively dopaminergically modulated, and to matrix, through which rhythms associated with the selection might be circulated broadly to cortex.
In terms of goal-motivated and iterative problem solving, a striking corollary of the BGMS model is that the BG both recognize and generate large scale patterns of synchrony in cortex, so that synchrony-oriented information processing in the BG is not just integrative, but recurrent. Iteration can provide for the formulation of solutions by a process of perturbative adjustments to representations (Lourenço et al. 2003), constituting a metaheuristic algorithm. In this vein, closed circuits in cortex and BG have been proposed to underlie the dynamical emergence of valuations and decisions, with structural hierarchy and hidden layers enabling adaptation to varying timescales (Hunt and Hayden 2017).
Structural and functional recurrence in the PFC and BG in particular have been suggested to arrange for the progressive integration of evidence to drive decisions (Bogacz and Gurney 2007; Caballero et al. 2018; Yartsev et al. 2018), a facility which is deficient (relatively inefficient and inflexible) in Parkinson's Disease (O’Callaghan et al. 2017).
Another consequence of these looping arrangements is that they provide for neural network layering (depth) that is notionally infinite. In general, network recurrence allows for limited computational resources to be recycled and repurposed over successive iterations, allowing for situationally appropriate tradeoff of speed versus accuracy (Spoerer et al. 2019‡).
Optimization without a priori expertise can benefit from random perturbations, whereby systems can escape impoundment in local optima to find broader or global optima (Lourenço et al. 2003; Palmer and O’Shea 2015; Palmer 2020‡). A related supposition is that the noise inherent to neural activity in cortex facilitates cognitive generalization, tending to make stimulus categorization maximally inclusive (Hu et al. 2019‡). Remarkably, artificial random perturbation of cortical microcircuits in PFC has been shown to significantly and broadly enhance cognitive performance (Sheffield et al. 2020‡).
It is thus interesting that pallidal and nigral
projection neurons appear to be arranged to continually inject noise into the
thalamocortical system, through their tonic, rapid, independently rhythmic
discharges. While this noise may improve signal fidelity in important
as suggested earlier (§5.2), it may also act to randomly perturb BG-thalamocortical
state, via thalamocortical, thalamostriatal,
Until a representation is robustly stabilized, random perturbations may contribute crucially to an organism's search for useful responses to environmental challenges and opportunities. Indeed, it has been suggested that brain dynamics are generally characterized by alternation between two modes, one chaotic, noisy, energetically diffuse, and prone to creativity, the other stable, deterministic, and energetically focused (Palmer and O’Shea 2015; Palmer 2020‡). If so, the BG are well-positioned to control these modes: as suggested above, BG influences over central dopamine, acetylcholine, and serotonin supplies position the BG to switch between broadly critical and broadly subcritical modes, which are clearly akin to the chaotic and focused modes described by Palmer and O’Shea (2015).
BG-controlled dopamine release in PFC is thought to stabilize working memory (WM) items, represented as attractor networks, against distractors and noise, simultaneous with its release in the BG enhancing targeted output to the implicated cortical loci (Gruber et al. 2006). This proposal comports neatly with findings, discussed earlier (§6.8), that dopamine increases overall PFC pyramidal neuron responsiveness to afferent activity, but reduces the responsiveness of their L1 apical dendrites (Yang and Seamans 1996), and depresses GABAergic lateral interactions in L2/L3 interneurons (Towers and Hestrin 2008). WM impairment in Sz has long been recognized as a cardinal symptom (Lee and Park 2005), and may be explained in large part by dysfunctions of DA regulation, excitatory-inhibitory balance, functional connectivity, and apical dendrite excitability (Uhlhaas 2013; van den Heuvel et al. 2013; Braver and Cohen 1999; Grace 2016; Goldman-Rakic 1999; Geyer and Vollenweider 2008; Dandash et al. 2017; Braun et al. 2019‡).
BG intimacy with the thalamic MD and VA nuclei underscores a multifarious role for the BG in managing working memory (Frank et al. 2001; McNab and Klingberg 2008; Chatham and Badre 2015; Kalivas et al. 2001; Haber and Calzavara 2009; Watanabe and Funahashi 2012; Mitchell and Chakraborty 2013; Xiao and Barbas 2004; Parnaudeau et al. 2013). The BG are thought to gate the establishment of items in WM and their inclusion in subsequent cognition, and to eject them from WM when they are no longer relevant to current context and goals, enabling reallocation of WM resources. Recent results show that the densely BG-recipient MD nucleus in particular can operate in precisely this fashion, sustaining context-dependent WM-related activity in PFC during delay periods (Bolkan et al. 2017), and regulating rule-contingent functional connectivity in PFC (Schmitt et al. 2017).
Constellations of PFC neurons activate persistently in sensorimotor stimulus-response scenarios (Haller et al. 2018). In self-paced tasks requiring complex and flexible responses to stimuli, such as antonym production, loci in human PFC were found to generate high gamma oscillations that immediately followed those in sensory cortex associated with the stimulus, and continued until, and during, high gamma generation in motor cortex associated with the response. Moreover, this PFC high gamma was essential for effective response production. These patterns of activity suggest that the PFC is crucial in the flexible organization of large scale functional networks, integrating relevant information and selecting appropriate responses (Haller et al. 2018).
Prefrontal gamma power rises as a function of WM load (Roux et al. 2012), and alignment of the fine phase angle of spikes in an oscillating cell subpopulation in PFC, relative to prevailing oscillation in the wider population, may serve to delimit and orthogonalize items represented by that subpopulation, readily explaining item capacity limitations (Siegel et al. 2009; Lisman and Idiart 1995). In an arrangement analogous to spike-timing-dependent selection among conflicting sensory inputs (Fries et al. 2002), BG-mediated selection among current WM items might entail synchronization of thalamocortical spiking with the corticocortical spiking associated with the selected item. BG-facilitated gamma bursts, securing effective connections, may thereby be produced, reintegrating the WM item into ongoing cognition. Indeed, gamma bursts in PFC accompany (putatively, induce) both the establishment of items in WM, and the subsequent activation of those items for inclusion in ongoing cognition (Lundqvist et al. 2016). The opposing prevalence of gamma and beta in WM neuron constellations has been implicated directly in top-down control of WM (Lundqvist et al. 2018a).
While WM entails persistent activity (Curtis and D'Esposito 2003; Wang 2001; Goldman-Rakic 1995), WM items may be further stabilized by, or even briefly exist only as, ephemeral synaptic potentiation and associated ephemeral attractor states in cortex (Lundqvist et al. 2018b, 2016; Miller et al. 2018; Rose et al. 2016). Some models of PFC-BG function in WM in fact necessitate such an intracellular state maintenance mechanism (Frank et al. 2001).
It seems clear that the frontal cortex and the BG are functionally coextensive components of a single inextricable system (Frank et al. 2001; Miller and Cohen 2001; Calzavara et al. 2007; Miller and Buschman 2007). Their functional delineation is thus fraught with nuance and ambiguity. One interpretation is that the BG learn associations quickly, at a lower level of generality, and frontal cortex learns associations more slowly, at a higher level of generality and stability, trained by the BG, so that frontal cortex eventually takes the lead in responding to stimuli (Miller and Buschman 2007; Antzoulatos and Miller 2011). While it is intuitively obvious that representations reflecting a larger number of examples will tend to be more general and abstract, it is not clear that the BG and frontal cortex differ crucially on this count. Even after over-training of a task, at least in some scenarios activity in the BG still leads that in frontal cortex (Antzoulatos and Miller 2014), and there is strong evidence that at least in some forms of motor learning, cortex is crucial for initial acquisition but not necessary for performance subsequent to consolidation (Kawai et al. 2015).
According to Frank et al. (2001), the frontal cortex represents information with persistent patterns of activation, while the BG fire selectively, usually impulsively, and only coincident with substantial afferent activity, to induce updates to those persistent patterns. This view is supported by evidence that habit formation is accompanied by the emergence of sharp start and stop signals in the BG (Jog et al. 1999; Barnes et al. 2005; Jin and Costa 2010; Smith and Graybiel 2013). And it comports with the view adopted within the BGMS model, that thalamocortical feedback activity can facilitate transient cortical gamma bursting, establishing effective connections through synchronization (Larkum 2013; Womelsdorf et al. 2014). As noted above, brief gamma bursts in PFC have particularly been proposed to impulsively shift brain state to selectively integrate and manipulate working memory items (Lundqvist et al. 2018a).
The BGMS model, and the model of Frank et al. (2001), furthermore comport with the sparse pattern of task-specific activity exhibited by direct path corticostriatal neurons (Turner and DeLong 2000), noted earlier (§8.6), because the corticostriatal activation associated with a particular behavioral and environmental conjunction is thereby inherently spatiotemporally limited. Cortical inputs to the indirect path arise mainly from a different population (Wall et al. 2013) that is not at all sparse in its activation patterns, consistent with an inhibitory function.
There are hints to the functional delineation of frontal cortex and BG in the results of several fMRI studies. A study by Cools et al. (2004) arranged to separate the metabolic correlates of shifts in the task relevance of objects (effectively, polymodal sensory stimuli), from those of transiently operative abstract rules. The BG and PFC were both found to be integral to the former type of preparatory shift, allocating attention and responsiveness to the task-appropriate stimuli, but not to the latter, which was only associated with activity in PFC, subsequently biasing striatal responses appropriately. Nonetheless, evidence from another fMRI study suggests that the striatum may continually model the utility of abstract rules as such, activating in response to category prediction errors, as distinguished from reward prediction errors (Ballard et al. 2018).
Another study showed activation of the BG when unexpected sensory stimuli prompted reorienting of attention, but not in preparatory orienting and maintenance of attention (Shulman et al. 2009). However, a similar study identified a role for the ventral BG in mediating shifts of preparatory attention in response to unattended (but evidently detected) sensory cues, inducing appropriate changes in frontal-posterior functional connectivity (van Schouwenburg et al. 2010b).
Exploration of the physiological correlates of cognitive and attentional flexibility has substantiated a role for the BG (Leber et al. 2008; van Schouwenburg et al. 2010b, 2012), leading van Schouwenburg et al. (2012, 2014) to propose that frontal cortex, particularly DLPFC, controls striatal function through dense topographically organized projections, and that these projections are crucial to cognitive flexibility. Magnetoencephalographic studies of the parkinsonian brain strongly support the proposition that the BG are directly implicated in cognitive flexibility: as the disease progresses, the variety of large scale functional constellations contracts correspondingly, with the striatum, pallidum, and thalamus, all significantly affected (Sorrentino et al. 2019‡). And using fMRI, a correlation has been shown between cognitive flexibility and resting functional connectivity of hub areas with the BG (Vatansever et al. 2016). Convergence of broad PFC projections in the head of the primate caudate nucleus, and activation of this area in cognitive tasks irrespective of the implicated domains, suggest that the BG themselves contain hub areas that process information generically (Averbeck et al. 2014; Assem et al. 2019‡).
In adolescent development, the flexibility-related pathologies characteristic of OCD and attention deficit hyperactivity disorder (ADHD) are associated particularly with deficient myelination of projections from PFC to striatum (Ziegler et al. 2019). It may be particularly relevant that in humans, but not in other species, frontal and striatal regions continue to develop well into adulthood (Sowell et al. 1999).
The decision to pursue curiosity, despite risks inherent to that pursuit, is particularly associated with striatal activation (Lau et al. 2018‡), implicating the anterior cingulate cortex, central dorsal striatum, anterior globus pallidus, and adjacent ventral pallidum (White et al. 2019‡). Similarly, the ventral striatum and pallidum are particularly activated when an individual perseveres in self-motivated attempts at extremely difficult tasks (Murayama et al. 2019‡). Multitasking, which entails both flexibility and perseverence, is most strongly associated with elevated information flow via the putamen to the inferior parietal sulcus and to a frontal area just anterior to the SMA; limits in the capacity of this path may underlie limits in the human capacity for multitasking (Garner et al. 2019‡; Garner and Dux 2015).
Even in the absence of PFC, the BG can maintain flexibility and curiosity. Lesion experiments in primates have demonstrated that flexible and contextually appropriate, “intellectual” behavior and curiosity are retained even in prefrontally decorticated animals, albeit with a surfeit of reactivity, provided the BG are preserved (Mettler 1945). This same series of experiments demonstrated that removal of striatal tissue produces hyperactivity and incuriosity, noting that “Animals lacking the striatum always display a certain fatuous, expressionless facies from which the eyes stare vacantly and with morbid intentness.” Subsequent bilateral pallidotomy in these animals produced hypokinesia eventually “indistinguishable from periods of sleep”
Evidence from the pathological brain supports the proposition that the BG are functionally coextensive with frontal cortex, implicating them extensively in action, awareness, and cognition. Throughout this paper I have noted implication of the BG system in schizophrenia, and related Parkinson's disease to various cognitive, perceptual, and behavioral syndromes. The BG have been implicated in GTS and OCD, as noted earlier (§7.10). Below, I briefly explore additional evidence of BG/frontal functional coextensiveness from clinical and lesion studies, and more thoroughly explore BG involvement in Sz.
Lesions of the BG in humans, particularly of the caudate portion of the striatum, lead to cognitive and behavioral deficits commonly associated with frontocortical lesions—frequently, abulia (loss of mental and motor initiative), disinhibition, memory dysfunction, and speech disturbances including, rarely, aphasia (Bhatia and Marsden 1994). Similar deficits occur with BG-recipient thalamic infarcts involving the MD and VA nuclei (Stuss et al. 1988) and intralaminar nuclei (Castaigne et al. 1981; Van der Werf et al. 1999), with intralaminar infarcts also consistently associated with profound attention deficits. Accidental bilateral destruction of the GPi, incidental to treatment for Parkinson's disease, has resulted in akinetic mutism (Hassler 1982).
Severe disability following brain injuries is consistently associated with selective cell loss in central thalamic nuclei, and as noted earlier (§7.3), permanent vegetative state (PVS) is associated with loss spanning the rostrocaudal extent of the intralaminar nuclei and MD nucleus (Schiff 2010). PVS is invariably accompanied by diffuse subcortical white matter damage, and is usually accompanied by widespread or severe thalamic damage, but often presents with no apparent structural abnormalities in the cerebral cortex or brainstem (Adams et al. 2000).
PVS is also associated with significant impairment of backward connectivity from frontal to temporal cortex, relative to minimally conscious patients and normal controls (Boly et al. 2011), directly implicating the most densely BG-recipient areas and layers of cortex. Similarly, anesthesia-induced unconsciousness is associated with disruption of backward connectivity from frontal to parietal cortex (Ku et al. 2011) and, as noted earlier (§7.3), with inactivation of the intralaminar nuclei (Alkire et al. 2008). Moreover, artificial activation of the intralaminar thalamus can restore wakefulness, and restore fronto-parietal functional connectivity in forward and feedback directions (Redinbaugh et al. 2020). Evidence suggests that it is decoupling of modulatory feedback paths, disrupting the vertical integrator function of neocortical pyramidal neurons, that is the ultimate mechanism of a variety of anesthetics acting by a variety of proximal mechanisms (Suzuki and Larkum 2020).
In some patients exhibiting akinetic mutism and other severe deficits associated with the minimally conscious state, administration of the GABAA agonist zolpidem has been found to reliably induce substantial but transient recovery, apparently by restoring normal function and oscillatory structure in frontal cortex, striatum, and thalamus (Brefel-Courbon et al. 2007; Schiff 2010). Similar transient recoveries in other patients exhibiting similar symptoms with BG involvement have been reported in response to administration of the DA agonist levodopa (McAuley et al. 1999; Berger and Vilensky 2014).
Perhaps the most remarkable discovery to emerge from various studies of the physiology of reduced or lost consciousness, is that the intralaminar thalamic nuclei, comprising a very small area indeed, are quite indispensable for consciousness (Bogen 1995; Baars 1995; Van der Werf et al. 2002). These are also the thalamic nuclei most intimate with the BG, bearing implications amply explored earlier (§7).
Many diseases are associated with corticostriatal abnormalities (Shepherd 2013), and Sz in particular has been proposed to be fundamentally a dysfunction of cortico-striatal loops, particularly implicating DLPFC and its striatal targets (Robbins 1990; Simpson et al. 2010). It is associated with significant anatomical attenuation of the DLPFC-VS projection, observed in both patients and their asymptomatic siblings (de Leeuw et al. 2015), simultaneous with abnormally elevated functional connectivity in the ventral frontostriatal system, and abnormally attenuated functional connectivity in dorsal frontostriatal systems, both of which are correlated with severity of symptoms, and are likewise apparent in both patients and their asymptomatic first-degree relatives (Fornito et al. 2013).
More generally, Sz is associated with deficient BG-mediated disengagement of the default mode network during directed task performance, simultaneous with striatal hyperactivity (Wang et al. 2015). Similarly, fMRI evidence demonstrates that Sz patients exhibit a characteristic pattern of significant differences in dynamical functional connectivity responses to sensory stimuli, with greater than normal connectivity established for some long range pairs, and less than normal for others (Sakoğlu et al. 2010). Consistent with those results, EEG evidence suggests perceptual deficiencies in Sz follow in part from dysfunction of top-down attention mechanisms, entailing abnormally low attention-related sensory gain, while bottom-up sensory processing is spared (Berkovitch et al. 2018).
Comparisons of spatial WM task performance by patients with frontocortical lesions, PD, and Sz, reveal related and often severe deficits (Pantelis et al. 1997). Sz patients show particularly severe deficits in set-shifting (Jazbec et al. 2007), and significantly attenuated WM capacity (Silver et al. 2003). If, as discussed earlier (§11.7), WM items are delimited by finely graded phase distinctions (Siegel et al. 2009), then the narrowness and accuracy of temporal discrimination imposes a limit on addressable item capacity. In Sz this selectivity is reduced by dysfunction of GABA-dependent cortical coincidence window mechanisms (Lewis et al. 2005; Gonzalez-Burgos et al. 2015), affecting the dynamics of corticocortical, BGMS, and non-BG-recipient transthalamic paths in similar measure. These deficient dynamics may lead to the localized structural deterioration characteristic of Sz: PFC and the thalamic mediodorsal nucleus are integral to WM (Bolkan et al. 2017; Schmitt et al. 2017), and atrophy of the circuitry linking these areas is significant in Sz (Giraldo-Chica et al. 2018).
Sz is associated with pervasive and progressive compromise of cerebral white matter integrity (Lim et al. 1999; Mori et al. 2007), decreased dendritic spine density (Glantz and Lewis 2000; Elston 2000), and pyramidal cell body atrophy (Rajkowska et al. 1998), particularly impacting long range links associated with connectivity hub areas of cortex (van den Heuvel et al. 2013; Collin et al. 2014; Crossley et al. 2014), which have been shown by mathematical argument to be particularly fragile (Gollo et al. 2018). Correspondingly, Sz is characterized by functional hypoconnectivity in the associative networks that implicate hub areas, simultaneous with functional hyperconnectivity in sensorimotor networks (Ji et al. 2019; Giraldo-Chica et al. 2018).
While abnormal hub area anatomical connectivity is most pronounced in individuals affected directly by the disease, the unaffected siblings of Sz patients also show significant attenuation of these links, relative to normal controls, even while connectivity in non-hub areas is unaffected in siblings, and is not significantly affected in Sz (Collin et al. 2014; de Leeuw et al. 2015). These patterns imply a large genetic component to the disease, and an etiology that implicates mechanisms of connectivity that are specific to hub areas, which as noted earlier (§11.3) include areas that are particularly dense targets of BG output.
Indeed, hereditary vulnerability to Sz may be a general hallmark of Homo sapiens distinguishing the species from other vertebrate taxa: many long range structural connections unique to the human cerebrum are also among those most affected by Sz, and may evidence a differentiating optimization for integrative long range connectivity via cortical hubs, resulting in a unique vulnerability to Sz, and also to other uniquely human pathologies such as autism, bipolar disorder, and Alzheimer's disease (van den Heuvel et al. 2019; Crossley et al. 2014).
Cerebral disintegration in Sz may be rooted in GABAergic dysfunction, and consequent pervasive oscillatory deficits (Lewis et al. 2005; Ferrarelli and Tononi 2011; Gonzalez-Burgos et al. 2015; Uhlhaas and Singer 2010; Marissal et al. 2018). A recent genome-wide association study (GWAS) suggests that executive function, which as noted above is particularly compromised in Sz, is particularly associated with genes expressed in GABA pathways (Hatoum et al. 2019‡). GABA dysfunction can disrupt the synchronies to which the cortex and striatum respond, the mechanisms whereby the BG effect selections and modulate spike timing in their targets, and the time alignments between corticocortical and trans-thalamic spike volleys that are necessary for BGMS and other subcortically mediated synchronization mechanisms. Moreover, corticostriatal projections to striosomal SPNs, as to matriceal SPNs, synapse sparsely, with high thresholds for discharge, resulting in similar sensitivity to input synchronies (Kincaid et al. 1998; Zheng and Wilson 2002). As noted earlier (§11.5), striosomes and PFC are arranged in recurrent dopaminergic loops. Thus synchronal abnormalities in inputs to striatum likely produce dopaminergic dysregulation, and associated dynamical dysfunction and pathological expressions of plasticity, constituting a key mechanism for pathological progression.
Abnormal neural noise in Sz, which is a particularly strong marker for the condition, might also be rooted in abnormally elevated inhibitory conductances, themselves an adaptation to the inhibitory (GABAergic) interneuron degeneration characteristic of the condition (Peterson et al. 2017‡).
While therapies targeting GABA have thus far produced modest and mixed results, continued development may lead to effective prophylactic and genuinely curative drug treatments, with the potential to alleviate the negative and cognitive symptoms that have heretofore robustly resisted treatment (Carpenter et al. 1999; Gonzalez-Burgos et al. 2015; Keefe et al. 2007).
The etiology of Sz, and even the epoch of its emergence as a disease in Homo sapiens, are notoriously obscure and controversial (Tandon et al. 2008). Sz patients exhibit a variety of seemingly contradictory symptoms, classified generally as positive, negative, and cognitive, with each subject exhibiting an idiosyncratic syndrome (Kay et al. 1987; Simpson et al. 2010). The explanation for this variety and obscurity is readily apparent, if the irreducible etiology of Sz is dysfunction of highly distributed and heterogeneous synchronization mechanisms centered on the thalamus—particularly BGMS, but also, largely analogous mechanisms implicating the cerebellum and hippocampus, described later (§13). A particular initiating syndrome within a particular mechanism or component thereof would likely result in Sz with a distinct symptomatology, but the cascading effects of the initiating syndrome disrupt the large scale dynamics of the brain in fundamentally similar ways, allowing etiologically distinct syndromes to be meaningfully grouped together under the rubric of “schizophrenias”.
Whatever its root causes, schizophrenia implicates most components of the BGMS system. Prominent are syndromes of the PFC, striatum, frontostriatal connectivity, and DA signaling, as noted above, and of the intralaminar nuclei and their connections to PFC, cortical FSI function, the TRN, GABA signaling generally, the PPN and LDT, the cholinergic system generally, and the 5-HT system, noted earlier. Also implicated are left-lateralized GP hyperactivity (Early et al. 1987), cytological and neurochemical anomalies in the BG-recipient and associative thalamus more broadly (Cronenwett and Csernansky 2010), aberrant functional connectivity of thalamus with cortex generally (Cheng et al. 2015), and abnormalities in the gross anatomy of the basal ganglia (Mamah et al. 2007). If any of these components is disrupted, the capacity for BGMS to appropriately establish and dissolve effective connections in cortex, and regulate the dynamics of existing connections, is disrupted in some fashion.
It is an old saw that madness and genius have much in common, even while evidence consistently demonstrates an inverse relationship between measures of intelligence and measures of psychopathology such as schizotypy (DeYoung et al. 2012). The “Openness/Intellect” trait in the “Big Five” personality model, and the concept of apophenia (the perception of patterns or connections where none exist), suggest a resolution of this paradox: schizotypy and genius both entail the perception of patterns and connections that are unusual and unfamiliar to others, the former as apophenia (confusion about reality), the latter as penetrating insight into reality (DeYoung et al. 2012; DeYoung 2015).
The Openness/Intellect trait is founded in cognitive exploration (DeYoung 2015), which is particularly associated with striatal activation (Lau et al. 2018‡) and BG networks (White et al. 2019‡). Generally, as noted earlier (§11.9), the striatum and its links with hub areas, particularly in frontal cortex, are integral to cognitive flexibility (Leber et al. 2008; van Schouwenburg et al. 2014; Vatansever et al. 2016; Mettler 1945).
Frontal cortex and striatum, according to genome-wide association study results, are particularly implicated in the expression of extremely high IQ (Coleman et al. 2019), and as reviewed above (§12.3), Sz is marked by extensive disruption of the functional relationship and structural connectivity between frontal cortex and striatum, with a strong hereditary component.
While Parkinson's disease is marked by a graded contraction of cognitive flexibility (Sorrentino et al. 2019‡), it frequently involves complex and extended visual hallucinations marked by mind-wandering (a mental state in which thoughts are unguided and unconstrained), and this mind-wandering is associated with pathological coupling of hub areas with visual areas of cortex (Walpola et al. 2020). Psychosis in PD often involves pareidolia, a type of apophenia in which meaningless stimuli, such as clouds, are falsely recognized as meaningful patterns, such as cats; as PD progresses, these mild symptoms may grow in severity, culminating in psychotic delusions (ffytche et al. 2017).
In the BGMS model, the striatum is the crux of flexible functional connectivity decisions. Given continuous semantic spaces in neocortex (Rao et al. 1999; Huth et al. 2012; Simmons and Barsalou 2003; Rajalingham and DiCarlo 2019; Lettieri et al. 2019; Zhang et al. 2019a‡), this implicates the striatum directly and uniquely in the formation of unusual connections characteristic of apophenia (as in Sz and PD) and penetrating insight (as in genius). This also suggests a dichotomy that is perhaps surprising, between intelligence and genius, with conventional measures of the former less responsive to distinctions in BG function, and the latter more closely associated with BG function, and less captured by conventional measures of intelligence.
Brains may intrinsically be subject to a tradeoff between propensity for creativity and madness, on the one hand, and stolid stability, on the other. When cortical representations minimize distance in semantic space between alternatives, creativity might be heightened, because perturbative exploration then more readily covers the semantic space. Perturbative exploration is largely grounded in the conjunction of homeostatic criticality (Haider et al. 2006; Ma et al. 2019; Ahmadian and Miller 2019‡), which inherently positions a universe of alternatives adjacently in phase space, and neural noise, which in itself is thought to be an indispensable ingredient in the creative exploration of semantic space (Palmer and O’Shea 2015). Indeed, evidence suggests a direct correlation between the critical regime and fluid intelligence (Ezaki et al. 2020). However, with smaller distances between alternatives, and greater noise, there is a greater burden on regulatory and coordinative systems, both those intrinsic to cortex, and subcortical systems centered on the thalamus, particularly the BG and cerebellum. When these systems function deficiently, madness may result.
In Sz, spike coincidence windows are pathologically enlarged (Lewis et al. 2005; Gonzalez-Burgos et al. 2015), resulting in relatively indiscriminate signal gating, while noise is pathologically elevated (Winterer and Weinberger 2003; Peterson et al. 2017‡), presumptively causing and/or evidencing representational instability. These conditions plausibly push regulatory and coordinative systems to the breaking point, and because they are arranged in closed loops with cortex, the breakage feeds back on itself. Through plasticity mechanisms, the resulting confusion is likely to be progressively embodied in the physiology underpinning semantic maps and relations, as reviewed above (§12.5).
In principle, the brain could guard against this cascade of confusion by systematically increasing the semantic spatial distance between alternatives, for example by regulating cortical microcircuits to be subcritical, effectively reducing neural noise, at the cost of a proportional reduction in creativity and flexibility. Indeed, in the normal brain, sleep deprivation is accompanied by compensatory regulatory departure from criticality (Meisel et al. 2017), and graded cognitive deficits (Banks and Dinges 2007).
Clearly, in Sz, regulatory compensation fails; analysis of network dynamics suggests that in Sz, top-down control of cognition entails higher than normal energy expenditure (is pathologically effortful), while small perturbations have an unusually large impact on network state (signifying pathological instability) (Braun et al. 2019‡). That Sz is still somewhat common in Homo sapiens (roughly 4.5 per 1000 (Tandon et al. 2008)), despite its devastating symptoms, attests to the irreducible evolutionary advantages of homeostatic criticality and endemic noise, and the creativity and flexibility they engender.
Proverb comprehension was the basis of some early diagnostic tests for Sz, and while in clinical practice these tests have proved unreliable, a more recent study demonstrated a strong correlation between performance on a proverb comprehension task and performance on a theory of mind task, and much better performance on the proverb comprehension task among normal controls than among Sz patients (Brüne and Bodenstein 2005). Proverbs are metaphors, and the successful comprehension of a metaphor entails the recognition of certain abstract semantic relations, simultaneous with the suppression of other, concrete, semantic relations. If these semantic relations are realized physiologically as long range effective connections, then impairment of the supervisory control of effective connectivity would manifest as impaired comprehension of metaphors.
The impression that emerges from the various behavioral and physiological anomalies characteristic of Sz, is of a brain that is continually and indiscriminately surprised. Percepts that are normally anticipated or familiar, and ideas that are normally dismissed as absurd, are not appropriately neglected, but instead are given spurious salience, with associated hyperdopaminergia (Kapur 2003; Bromberg-Martin et al. 2010). Subjective duration, causality, sequentiality, and simultaneity, are abnormal and distorted (Martin et al. 2013; Schmidt et al. 2011; Ciullo et al. 2016).
These phenomena can all follow from deficiencies in the mechanisms of representation, and these deficiencies plausibly follow from dysfunction in spike-timing-dependent gain mechanisms, particularly implicating GABA. Representational deficiencies readily lead to senseless surprise and ideation, and associated maladaptive attentional focus and expressions of plasticity. With expectations and impressions that are fundamentally untrustworthy (cognitive symptoms), paranoia and bizarre behavior (positive symptoms) and indiscriminate withdrawal (negative symptoms) naturally follow.
By this narrative, treatment that restores the trustworthiness of expectations and impressions, producing remission of cognitive symptoms, will naturally lead to remission of positive and negative symptoms. And indeed, just these sorts of results are apparent in experimental non-pharmacological therapies, consisting only of working memory exercises: cognitive and negative symptoms show significant and sustained improvement (Ramsay et al. 2017; Cella et al. 2017), and perhaps most remarkably, there is evidence that these exercises restore some of the functional connectivity between thalamus and PFC that is characteristically lost in Sz (Ramsay et al. 2017).
The cerebellum is an ancient structure, present at the base of vertebrate phylogeny (Bell 2002), and its pervasive involvement in precision motor learning and sensory-motor coordination was established generations ago (Ito 2002). The cerebellum learns associatively, and in particular, learns predictive relations underlying forward control in stimulus-response behaviors (Giovannucci et al. 2017) and dynamic proprioception (Weeks et al. 2017). The extreme regularity of its physiology, and an absence of intrinsic excitatory feedback paths, have long inspired mechanistic, computational models of its function (Heck 2016; Cheron et al. 2016).
Across a wide variety of mammalian taxa, there is a consistent ratio of 3-4 cerebellar neurons for each neocortical neuron, suggesting a close relationship between these structures (Herculano-Houzel 2010). The cerebellum is reciprocally linked with cerebral cortex and the BG (Bostan and Strick 2010; Bostan et al. 2013; Bostan and Strick 2018; Milardi et al. 2016), and its closed loops with cortex (via the thalamus) resemble those of the BG (Schmahmann and Pandya 1997; Strick et al. 2009; Glickstein et al. 1985). Oscillatory synchronies between the cerebellum and the cerebrum are recognized and proposed to be functionally significant (Courtemanche et al. 2013; Courtemanche and Lamarre 2005; Cheron et al. 2016).
Like the BG, the cerebellum exhibits fractured and repeated somatotopy and modular divergence-convergence (Manni and Petrosini 2004; Apps and Garwicz 2000; Flaherty and Graybiel 1994), with inputs from widely distributed neocortical areas converging in various combinations (Brodal and Bjaalie 1997; Kincaid et al. 1998).
The cerebellum seems to be functionally nearly coterminous with the BG, including extensive and varied cognitive and other non-motor roles (Strick et al. 2009; Schmahmann and Pandya 1997) and generic (“multi-demand”) roles (Assem et al. 2019‡). Roles have been identified for the cerebellum in fear memory formation and expression (Frontera et al. 2020‡), rhythmic perception (Kameda et al. 2019), and spatial attention (Craig et al. 2019‡) and inattention, including inattention to self-generated percepts (Kilteni and Ehrsson 2020), a role suggested earlier (§10.6) for the BG.
Learning to remove predictable features from sensory inflow is thought to be a fixture of cerebellar function throughout vertebrate phylogeny (Bell 2002; Ito 2001). While dopamine fluctuations signal prediction errors and induce plastic adaptations in the BG and wider forebrain (Schultz et al. 1997; Sharpe et al. 2017), in the cerebellum, powerful and highly specific inputs from the inferior olivary nuclei of the brainstem constitute prediction error signals; these signals induce plastic adaptations, if they are not dynamically inhibited by accurate predictions relayed from the Purkinje projection cells to the olivary nuclei via the deep cerebellar nuclei (Schweighofer et al. 2013; Ito 2001). These plastic adaptations, centered in the cerebellar cortex, result in the generation of refined motor output (Ito 2002; Giovannucci et al. 2017).
Each olivocerebellar axon branches to form about 7 “climbing fibers” in the cerebellum (Fujita and Sugihara 2013), each of which strongly and repeatedly apposes a single Purkinje cell with a 1:1 ratio (Reeber et al. 2013). The olivocerebellar system exhibits exquisitely precise topographic relations (Reeber et al. 2013), and its inputs span the central nervous system, from the lumbar spine, through various nuclei of the brainstem, the deep cerebellar nuclei, and superior colliculus, to layer 5 of the frontal and parietal neocortex, but notably exclude occipital and temporal cortex, the thalamus, and the BG in their entirety with the sole inclusion of the ventral tegmental area (Swenson and Castro 1983a, 1983b).
The pontocerebellar mossy fiber system, constituting the sole extrinsic input to the granule cells of the cerebellar cortex, relays neocortical input from frontal motor and eye field areas, parietal including posterior areas, the entire cingulate gyrus, and extra-striate occipital visual areas, with minor inputs from polysensory and auditory association areas in temporal cortex, the parahippocampal gyrus, and dorsolateral and some medial PFC; the corticopontine projection is topographically precise, while the pontocerebellar projection is convergent-divergent, providing opportunities for integration (Brodal and Bjaalie 1997; Glickstein et al. 1985; Schmahmann and Pandya 1997).
fMRI, electrophysiological, and tracer/degeneration studies demonstrate that cerebellar output targets the motor, association, sensory, and intralaminar nuclei of the thalamus, and the superior colliculus, and through them, widely distributed areas within motor, sensory, and association cortex, including skeletomotor, oculomotor, prefrontal, somatosensory, parietal, insular, temporal including primary auditory, and occipital including primary visual, with extremely high temporal fidelity (Schmahmann and Pandya 1997; Sultan et al. 2012; Strick et al. 2009; Kalil 1981; Katoh et al. 2000). This expansive, temporally precise, largely reciprocal influence of the cerebellum on motor, sensory, and association cortex is consistent with a central role in forward modeling of the anticipated sensory correlates of actions (Sultan et al. 2012; Blakemore et al. 1998; Lindner et al. 2006; Synofzik et al. 2008).
However, as noted earlier (§6.3), integration of the cerebellum into neocortical circuitry systematically differs from the arrangements that characterize the BG. Movement can be evoked by electrical stimulation of zones in the motor thalamus receiving input from the cerebellum (Vitek et al. 1996; Buford et al. 1996). Cerebellum-recipient neurons are consistently within the parvalbumin-positive “core” population, associated with specific and narrowly circumscribed topographic projections (Jones 2001; Kuramoto et al. 2009). And projections from these areas terminate in L2-L5, including L4 proper (Kuramoto et al. 2009; García-Cabezas and Barbas 2014; Clascá et al. 2012). Thus, as discussed earlier (§6.2), the laminar targets of paths through the cerebellum resemble corticocortical feedforward paths; in contrast, BG paths resemble corticocortical feedback paths.
However, the cerebellum is not necessary for coherent thought and behavior—these are preserved with manageable and finite deficits even in cases of complete cerebellar agenesis (Yu et al. 2015). In primates, the white to gray matter ratio is lower in cerebellum than in neocortex (in chimpanzee, 0.24 and 0.64 respectively), and across mammalian taxa the scaling exponent of that ratio is significantly lower in cerebellum than in neocortex (1.13 and 1.28 respectively) (Bush and Allman 2003). While roughly 80% of projection fibers in neocortex have terminals elsewhere in neocortex, the cerebellar cortex does not form links with itself (Heck and Sultan 2002). The main projection of the cerebellar cortex to the deep cerebellar nuclei is not directly reciprocated—indeed, cerebellar cortex forms no excitatory closed loops with the deep cerebellar nuclei or otherwise (Cheron et al. 2016). These features suggest that the cerebellum intrinsically performs little or none of the intrinsically recurrent and self-associative processing characteristic of the corticothalamic system in particular and the forebrain in general. Unusually low variability of the dynamic functional connectivity of cerebellar loci, and uniquely high similarity of structural to functional connectivity in the posterior cerebellum (Fernandez-Iriondo et al. 2020‡), suggest a prevalence of relatively narrow, precise functions for cerebellar modules, with little dynamic flexibility.
Nonetheless, that the cerebellum and BG have similar topological relationships to cortex suggests that they may affect cortical activity by similar mechanisms. The physiological minutiae of the cerebellum bear a striking resemblance to key aspects of BG physiology and function:
The GABAergic Purkinje cells of the cerebellar cortex have high intrinsic firing rates of up to 90 Hz, and they modestly converge (~50:1) on projection neurons in the deep cerebellar nuclei, which are accelerated and entrained when that input is synchronous (Heck et al. 2013). The deep cerebellar nuclei receive collaterals of the excitatory mossy fibers that innervate the granule cells (Shinoda et al. 1992), and precisely follow stimuli to frequencies above 600 Hz, with little fatigue (Sultan et al. 2012), in an arrangement similar to that of the projection cells in intralaminar thalamic nuclei, reviewed earlier (§7.5).
Each Purkinje cell's flat fan-like dendritic tree receives excitatory inputs from ~175,000 weak appositions by thin, unmyelinated, slow-conducting (~0.5 m/s) “parallel fibers” traveling perpendicular to the Purkinje dendritic fans (Heck and Sultan 2002). In humans, the granule cells that originate these parallel fibers number about 1011 (Andersen et al. 1992), which comprises the vast majority of neurons in the brain as a whole (Azevedo et al. 2009). In mouse, each mossy fiber ends in roughly 150 small “rosettes”, each of which apposes roughly 21 granule cells, so that each input fiber to the cerebellum is distributed to the cerebellar cortex through over 1000 parallel fibers (Sultan and Heck 2003), vastly expanding the dimensionality of the input pattern representation. The astronomically large population of granule cells, and the associated divergence-convergence of cerebellar circuitry, suggest enormous combinatorial power (Marr 1969; Sultan and Heck 2003).
Purkinje cells, like striatal SPNs, respond only to synchronized input (Sultan and Heck 2003), and have “Up” and “Down” states, with transitions triggered by impulsive input currents (Loewenstein et al. 2005). Similar to the slow and varied corticostriatal fiber population, and with a similar range of conduction delays, parallel fibers can function to align a constellation of non-coincident afferent spike volleys originating in diverse locations, so that their arrival at a Purkinje cell is precisely coincident, evoking a response; other Purkinje cells, at which these inputs are not coincident, are unresponsive (Braitenberg et al. 1997; Braitenberg 1961; Heck 1993, 1995; Heck et al. 2001; Heck and Sultan 2002; Sultan and Heck 2003). Braitenberg et al. (1997), noted earlier (§1.9), term this the “tidal wave” mechanism.
Similar to the BG, and with important implications discussed earlier (§7) regarding BG circuitry, the cerebellum targets both superficial and deep neocortical layers, via separate deep cerebellar nuclei; in particular, the fastigial nucleus targets superficial layers as part of a putative diffuse activating system (Steriade 1995).
These arrangements suggest that the cerebellum may affect cortical activity much as the BG do in BGMS. Indeed, this proposal has been previously suggested (Courtemanche et al. 2013), is implied by evidence of task-specific preparatory and reactive oscillatory synchronies between functionally related cerebellar and neocortical areas (Courtemanche and Lamarre 2005), and is strongly suggested by evidence that the cerebellum is necessary for normal activity-related gamma synchrony between sensory and motor cortex (Popa et al. 2013). It was recently reported that Purkinje cells explicitly represent the oscillatory phase difference between medial PFC and hippocampus, crucially implicating conduction delays in cerebellar parallel fibers (McAfee et al. 2019), and indeed evidence directly supports the proposition that the cerebellum modulates gamma coherence between these two cortical areas (Liu et al. 2020‡).
As discussed earlier (§3.2) regarding the BG, and evidently with equal relevance to the cerebellum, recurrent paths with variable delays allow for selective reinforcement of particular oscillatory frequencies. Because each large scale functional constellation exhibits a characteristic profile of dominant frequencies (Keitel and Gross 2016; Becker and Hervais-Adelman 2019‡), this also suggests that the cerebellum, like the BG, can dynamically activate and stabilize specific large scale networks appropriate to context. Indeed, transcranial magnetic stimulation (TMS) of the cerebellum at theta and beta frequencies has been shown to facilitate performance of tasks whose functional networks are known to involve oscillations at frequencies that match the TMS, with double dissociation of the effects (Dave et al. 2020‡). Remarkably, individual Purkinje cells can learn to respond to temporally simple inputs with delayed, temporally complex, multiphasic outputs (Johansson et al. 2014; Majoral et al. 2020), which—like intrinsic oscillation in the BG—might allow the cerebellum to generate contextually appropriate oscillatory output even in the absence of oscillatory input.
The hippocampal system is thought to function as a persistent associative memory repository of first resort, capturing patterns of cortical network activation representing significant associations as they occur, in close coordination with prefrontal cortex (Eichenbaum and Cohen 2014; Battaglia et al. 2011; Rolls 2010; Squire et al. 2004; Damasio 1989; Meyer and Damasio 2009). Subsequent to initial encoded storage in the hippocampal formation, these associations are for a limited time available for retrieval (reactivation of the original cortical pattern), both to contribute to mental activity during wakefulness when relevant, and for migration to less labile (and more capacious) areas outside the hippocampus (Frankland and Bontempi 2005; Folkerts et al. 2018). This process of migration is believed to occur mostly or entirely during sleep (Wilson and McNaughton 1994; Battaglia et al. 2011; Rasch and Born 2013; Schapiro et al. 2019).
The special facilities of the hippocampal formation follow in part from its unique plasticity (Martin and Morris 2002; Deng et al. 2010; Snyder et al. 2005; Shors et al. 2001; Cameron and Mckay 2001; Hastings and Gould 1999) and its exceptional capacity for long-range functional connectedness (Lavenex and Amaral 2000; Mišić et al. 2014; Grandjean et al. 2017).
Note that, in this brief treatment, the term “hippocampal formation” refers to the collection of medial temporal lobe areas that are functionally and spatially contiguous with the hippocampus proper, namely the dentate gyrus, hippocampus, subiculum, presubiculum, parasubiculum, and entorhinal, perirhinal, and parahippocampal cortices (Lavenex and Amaral 2000). The “hippocampal system” comprises the hippocampal formation, the thalamic midline and anterior nuclear groups, the mammillary bodies, the septal nuclei and diagonal band of Broca, the circuitry interconnecting these loci, particularly the fornix, and the connections of these areas to the rest of the brain.
Bilateral lesions destroying or disabling the hippocampal formation are associated with severe anterograde amnesia and graded retrograde amnesia, but spare intellectual, attentional, and most working memory capacities, motor skill learning, and semantic and other non-episodic memory, and are not associated with any apparent progressive deterioration, neither of the initially unaffected mental faculties, nor of brain physiology outside that directly affected by the initial lesions (Schmolck et al. 2002; Corkin 2002; Annese et al. 2014; but see Schapiro et al. 2019). This pattern of deficits shows that the function of the hippocampal formation is highly specialized. Moreover, that function is not contingent on conscious engagement (Henke 2010).
Yet memory processing by the hippocampal formation entails sensitivity to and activation of widely distributed networks, close integration with PFC, profuse projections to the ventral BG (Brog et al. 1993), profuse innervation by midbrain DA centers (Gasbarri et al. 1996), and consolidation processes implicating widely synchronized thalamocortical signaling, all of which it shares with the highly generalized BGMS system described here. Moreover, long-term memory deficits in general, and hippocampal system dysfunction in particular, have been multifariously implicated in Sz as vulnerability indicators and primary symptoms (Holthausen et al. 2003; Harrison 2004; Seidman et al. 2003; Sigurdsson et al. 2010). Thus, though the functional role of the hippocampal system is circumscribed, many of its operating principles and physiological underpinnings are shared with the BG-thalamocortical system.
Oscillatory activity, and sensitivity to phase, have been amply demonstrated in the hippocampal system and in long term memory processing (Fell and Axmacher 2011; Colgin 2011; Tort et al. 2008; Fernandez et al. 2013). Information flow in the hippocampal system is systematically organized around theta oscillation, with encoding of new incoming information at antiphase with retrieval of past information (Siegle and Wilson 2014; Hasselmo et al. 2002; Wilson et al. 2015). Rhythmic coordination of hippocampus and striatum has been demonstrated during learning (DeCoteau et al. 2007), and hippocampus and PFC exhibit increasingly synchronized oscillation as rules are acquired in a task framework, with peak coherence at the moment of decision; during subsequent sleep, hippocampal cell assemblies that participated in the coherent oscillation during performance are preferentially replayed (Benchenane et al. 2010).
It seems plausible, even likely, that reactivation of connectivity patterns by the hippocampal system entails a BGMS-related mechanism dependent on relays through the thalamus, particularly implicating the midline and anterior nuclear groups. Components of the hippocampal system project to all of the midline nuclei, which in turn project to superficial and deep layers of most cortical areas, and to the ventral striatum (Van der Werf et al. 2002). Notably, BG direct path and hippocampal system inputs are mutually exclusive in the midline and intralaminar nuclei, each nucleus innervated by one or the other, but not both (Van der Werf et al. 2002). The midline nuclei, like the intralaminars, are well-positioned to control cortical synchronies and associated effective connectivity (Saalmann 2014).
The anterior nuclear group is densely and reciprocally linked with the hippocampal formation, and projects extensively to neocortex, particularly to secondary motor, prefrontal, cingulate, retrosplenial, and some visual and temporal areas, but does not project to the BG (Jankowski et al. 2013), though many of these cortical targets are also targeted by BG-recipient thalamus.
While inputs to the BG-recipient thalamus arise from the entire cortex, cortical inputs to the midline and anterior nuclei are highly restricted, confined almost entirely to the hippocampal system, despite projections from these nuclei encompassing nearly the entire cortex (Van der Werf et al. 2002; Jankowski et al. 2013). Thus, whereas BGMS is proposed to attend the control of arbitrary corticocortical connectivity, necessitating elaborate selection by the BG, involvement by the hippocampal system in the initial reactivation of a memory might entail only signals from the hippocampal formation to neocortex, corticocortically and via transthalamic paths through the midline and anterior nuclei.
Direct projections from the hippocampal formation to neocortex also parallel those of BG-recipient thalamus: evidence from genetically manipulated mice suggests that perirhinal projections to neocortical layer 1 gate the formation of memories, with a crucial role for selectively facilitated bursting of L5 pyramidal neurons, both in the formation and the retrieval of behavioral memories (Doron et al. 2019‡), much as projections from BG-recipient thalamus to superficial cortex influence the formation and persistence of functional connections, as discussed earlier (§6.5).
The central role of theta oscillation in hippocampal system dynamics might in part reflect a relative inflexibility of trans-hippocampal path delays, so that adequate time alignment can only be attained at lower working frequencies. This implies that the terminal patterns of hippocampal system projections to neocortex disfavor the interneurons that realize narrow coincidence windows there, though this is yet to be demonstrated empirically.
The obvious suggestion is that the midline nuclei and anterior nuclear group function within the hippocampal system the way the intralaminar nuclei and MD, VA, VL, and VM nuclei function within the system described by the BGMS model, with both systems operating chiefly by the spike-timing-dependent mechanisms endemic to the cerebral cortex, thalamus, and striatum. The PFC and ventral striatum, jointly targeted by the hippocampal formation and by thalamic and other nuclei in both systems, are then positioned to coordinate activity in these two vast and largely separate systems, particularly by incorporating motivation and behavioral relevance into the control of memory formation and activation. BG influence on hippocampal system activity is implied by projections from the BG-recipient PC and PF nuclei to perirhinal, entorhinal, prelimbic, and parahippocampal cortices (Van der Werf et al. 2002). Additionally, the paraventricular and reuniens nuclei of the midline group receive dense projections from midbrain DA centers (Van der Werf et al. 2002), whereby the BG presumptively align memory dynamics with motivational context, while the central medial nucleus of the intralaminar group receives hippocampal system inputs (uniquely among nuclei classified as “intralaminar”), and projects densely to the dorsal striatum (Van der Werf et al. 2002), suggesting episodic memory contextualization of dorsal BG inputs.
A notable architectural distinction between these two systems is that hippocampal formation input to thalamus is excitatory, like neocortical and cerebellar inputs to thalamus, whereas BG input is GABAergic. Thus the selection and timing of signals to be dispersed by the midline and anterior nuclei is presumptively determined before those signals arrive in thalamus, consistent with the much narrower collection of inputs compared to that of BG-recipient thalamus. However, TRN inputs to these nuclei (Jankowski et al. 2013; Kolmac and Mitrofanis 1997; Van der Werf et al. 2002), and pallidal collateral inputs to the anterior nuclear group (Parent et al. 2001) provide paths whereby PFC and the BG might influence memory processes at the thalamic level (Pinault 2004; Zikopoulos and Barbas 2006; Guillery et al. 1998; Hazrati and Parent 1991; Shammah-Lagnado et al. 1996; Antal et al. 2014).
Just as the PFC and BG may orchestrate BG-thalamocortical neglect of expected percepts (discussed earlier (§10.6)), PFC has been suggested to orchestrate neglect by the hippocampal formation of previously stored episodic information, inhibiting redundant memorization (Frankland and Bontempi 2005). Indeed, the suppression of contextually inappropriate memory activation has been proposed as a general role for the PFC in its relationship to the hippocampus (Eichenbaum 2017). PFC activity patterns have also been shown to represent associations between a stimulus and context, an action triggered by that stimulus and context, and an unexpectedly rewarded outcome associated with that action, allowing for accurate credit assignment and the formation of memories that subsequently guide behavior toward reward (Asaad et al. 2017). Novelty and surprise are associated with an increase in theta synchrony between PFC and the hippocampus, supporting learning of unexpected information (Gruber et al. 2018). Moreover, the hippocampal formation is itself sensitive to familiarity (Squire et al. 2004), and through its projections to PFC and the ventral striatum, may promote neglect of familiar perceptual minutiae that would otherwise be distracting. Sz is marked by deficiencies in these capabilities, and corresponding hippocampal abnormalities (Jessen et al. 2003; Weiss et al. 2004).
The most studied and attested roles of the hippocampal system involve memory formation and recall, but the mechanism whereby it is understood to do this — rapid plasticity that establishes long range links between cortical areas — might be quite general. Evidence of hippocampal involvement in the performance and consolidation of skilled motor behavior, even without crucial involvement in initial acquisition of the skill (Schapiro et al. 2019; Burman 2019, 2018‡), suggests such a generality, and indeed suggests that the hippocampal system may orchestrate consolidation of long range links that do not themselves directly transit the hippocampal system. In terms of BGMS, it may not make any fundamental difference whether two areas are linked by direct, appropriately potentiated corticocortical projections, or by temporary routes through the hippocampal system. These two classes of long range linkage inevitably coexist according to the consolidation and reconsolidation theories of hippocampal function, and might indeed act synergistically. With hippocampal function centered on the rapid establishment of long range anatomical connections amenable to reactivation, BG function entailing the activation of long range anatomical connections, and PFC integral to the circuitry of both, it seems inevitable that these two systems are unified in their function. This proposition does, however, raise important questions about conduction delays associated with trans-hippocampal paths, compared to those of the corresponding corticocortical paths that are thought to be the ultimate destination of the relations migrated by consolidation.
The zona incerta (ZI) is an agglomeration of cytologically heterogeneous diencephalic nuclei below the thalamus, adjacent to the TRN and STN, connected with many of the areas and populations involved in BGMS (Mitrofanis 2005; Ricardo 1981; Shammah-Lagnado et al. 1985). Multiple, overlapping somatotopic maps are found throughout the ZI, maintaining largely parallel segregated circuits between the neocortex, thalamus, superior colliculus, brainstem, and spinal cord (Nicolelis et al. 1992; Power et al. 1999), but with no apparent topographic structure in projections to intralaminar thalamus (Power et al. 1999). Like the striatum, the ZI is extensively innervated by cortical layer 5 (Mitrofanis and Mikuletic 1999). Like the GPi and SNr, many of its neurons contain parvalbumin (Trageser et al. 2006), and it has extensive GABAergic projections to thalamus, with giant terminals apposing the proximal dendrites of projection neurons in association nuclei (Barthó et al. 2002; Power et al. 1999).
The physiology of the ZI is unlike that of the BG and TRN in several important respects, such that its operating principles are clearly distinct. Unlike either the BG or the TRN, the ZI has prominent direct projections to cerebral cortex; these projections are GABAergic, predominantly appose outer L1, are topographically organized, are densest in somatosensory cortex, and also project sparsely to visual cortex (Lin et al. 1997; but see Saunders et al. 2015a). While BG projections to intralaminar nuclei preferentially appose smaller and more distal dendrites, ZI projections to the intralaminar thalamus appose larger and more proximal dendrites (Barthó et al. 2002). The tonic discharge rate of ZI neurons, averaging 2-4/s (Périer et al. 2000; Trageser et al. 2006), is a small fraction of that of GPi/SNr projection neurons.
Like the BG and thalamus, the ZI is targeted by cholinergic projections from the PPN and LDT (Trageser et al. 2006), and like the cortex, it is targeted by the basal forebrain (Kolmac and Mitrofanis 1999), but the effect of ACh on ZI is to silence it (Trageser et al. 2006). Thus the ZI response to ACh resembles that of the TRN (Steriade 2004; Lam and Sherman 2010). However, whereas the TRN receives collaterals of L6 axons and not of L5 axons, and has a reciprocal relationship with the rest of the thalamus, as noted above the ZI receives L5 collaterals, and it does not receive thalamic inputs (Barthó et al. 2002).
Through its widespread projections to thalamus, the ZI has been suggested to synchronize oscillations in large populations of projection cells, acting as a relay whereby signals from its afferents can selectively facilitate transmission of sensory signals by those projection cells (Barthó et al. 2002, 2007). Experimental and clinical results in PD show that hyperactivity and hypersynchrony in the ZI are associated with dyskinesia and bradyphrenia just as they are in the GPi and SNr (Merello et al. 2006; Périer et al. 2000, 2002), and indeed that deep brain stimulation (DBS) in ZI may be a more effective technique for alleviating medically refractory PD than DBS in STN (Plaha et al. 2006).
An intriguing proposal is that rhythmic GABAergic input to the sensorimotor and intralaminar thalamus from ZI relays activity from attentional orientation centers such as the superior colliculus, disrupting BG-related activity in the thalamus and replacing it with selective receptivity to unexpected sensory inputs deemed salient by attentional orientation centers (Watson et al. 2015). This fits well with the proposition that the general function of the ZI is to gate sensory receptivity (Trageser and Keller 2004; Trageser et al. 2006; Lavallée et al. 2005; Urbain and Deschênes 2007), and like BGMS, is a proposal that selections can be made in the thalamus by GABA-mediated spike-timing-dependent gain control. And with its GABAergic projections to upper L1, the ZI is positioned to adjust spike-timing-dependent gain in its targets with particular rapidity and thoroughness.
The claustrum may be functionally similar to the ZI and, by extension, to the BG, but with its own peculiarities. It has long been a subject of notoriously inconclusive study (Edelstein and Denaro 2004). Its function is murky, and like the ZI, it is something of a chimera, combining physiological and functional attributes of the cerebral cortex, the striatum, the thalamus, and the basolateral amygdala (Swanson and Petrovich 1998). Like the BG, the claustrum appears arranged to synchronize the cortical areas with which it is connected, but unlike the BG, it appears to do so only occasionally. In particular, unlike pallidal projection neurons, and like ZI projection neurons, claustral projection neurons have a low tonic firing rate, 0-10 spikes/s in awake animals (Edelstein and Denaro 2004 p.5).
The claustrum contributes to the recruitment of a generally well-adapted neural response to novel and unexpected situations (Badiani et al. 1998; Remedios et al. 2014), and to emotionally freighted stimuli (Redouté et al. 2000), by orienting attention toward immediate, external sensory specifics. Evidence also suggests its involvement is necessary for optimal performance in learning and behavioral scenarios that are cognitively demanding (White et al. 2018‡), with significant activation particularly at the onset of a demanding task phase (Krimmel et al. 2019).
Similar to the ZI, the claustrum reciprocates with topographic cortical maps for all of the exteroceptive senses; the implicated claustral neurons have quite large receptive fields, with some incidence of polymodality (Sherk 1986). The entire claustrum is modulated by afferents communicating situational saliency (exceptionality) from VTA and SNc, the thalamic reuniens nucleus, the lateral hypothalamus, the locus coeruleus, the dorsal raphe nucleus (Słoniewski et al. 1986), and through some path yet to be fully anatomically elucidated, from a cholinergic source (Salerno et al. 1981; Nieoullon and Dusticier 1980).
Perhaps claustral neurons synchronize oscillations in the cortical areas connected to it, similar to the dynamic described earlier (§6.5) in the thalamocortical projection. For example, activity is relayed from anterior cingulate cortex, through the claustrum, to parietal and visual cortex, and this pathway may be a mechanism for top-down cognitive control (White and Mathur 2018). In the claustrum, however, the effect appears to be well-gated by the ascending modulatory afferents, resulting in the low tonic firing rate noted above. This is a limited role, similar to a general role proposed previously for the claustrum (Crick and Koch 2005), itself similar to the role ascribed to the BG in this paper.
The cortical areas with which the claustrum has been established to reciprocate, and which are therefore most likely to be subject to synchronization in the manner of thalamocortical circuits, are (1) various topographically mapped unimodal sensory areas for each of the senses (at least the exteroceptive ones) (Sherk 1986), (2) the frontal eye fields and supplementary motor area (SMA) (Sherk 1986), (3) several default mode network loci (orbitofrontal, cingulate) (Sherk 1986), and (4) the hippocampal system (Wilhite et al. 1986). The claustrum has unreciprocated projections to much of the rest of cortex, notably area 46 (DLPFC) (Sherk 1986), which might impart selective receptivity in the recipient areas to sensory and motivational activity that activates the claustrum, selectively boost cognitive activity that is already phase-locked with it, and relatively diminish other activity.
The claustrum has convergent afferents from thalamic midline and intralaminar nuclei (reuniens, CM, PF, PC, and CL) (Van der Werf et al. 2002). When the cholinergic, noradrenergic, dopaminergic, serotonergic, and other diffuse modulatory claustropetal afferents signal situational or anticipated salience (Salerno et al. 1981; Schultz 1998; Matsumoto and Hikosaka 2009; Nakamura et al. 2008), claustral neurons might attempt to synchronize activity in the sensory stream with itself, with oculomotor and skeletomotor activity, with the default mode network, and with the hippocampal system, so that the situation and its sensory correlates are well-attended, consciously integrated, and well-reflected by the memories that are activated and recorded.
In the BGMS model, effective connections are established in cortex in response to striatal decisions. This phenomenon would likely be detectable in the relationships among cortical and striatal LFPs, and would inherently be detectable using large electrode arrays to measure single unit activity in large populations of neurons in cortex and striatum. Specifically, in well-trained tasks, a highly significant relationship of consistent delays should be found between the timing of a spike volley arising in a particular set of cortical loci, the presence and timing of subsequent spike volleys arising in one or more connected striatal loci, and the establishment of functional connections between the first set of cortical loci and a consistent set of other loci as measured by LFP or individual spike activity in the latter.
In some experimentally accessible and reliably reproducible scenarios, the establishment of a long range corticocortical functional connection is predicted to be strongly contingent on the occurrence and precise timing of the striatal activation, as suggested by the results reported by van Schouwenburg et al. (2010b). The BGMS model predicts that if the striatal activation is absent or mistimed, the connection is unlikely to be established, else the connection is likely to be established. While experiments of this type cannot fully validate the BGMS model, done carefully they can decisively invalidate it, or provide highly suggestive evidence for it. Moreover, raw LFP and spike data sets for experiments that have already been performed can likely be reanalyzed to look for this phenomenon.
The central prediction arising from the BGMS model is that relationships of entrainment characterize sparse ensembles of directly connected neurons spanning the entire BG direct path during activation. If the effect of pallidal and nigral output on the thalamus is probed in awake healthy (normal) animals, the prediction is that phasic activation in many cases entrains thalamic activity. Preliminary results reported by Schwab (2016) give evidence of ensemble phasic entrainment of motor thalamus by the GPi, while underscoring that spatiotemporal sparseness and stochasticity in this activation and entrainment greatly complicate characterization at the single unit level.
Notwithstanding methodological hurdles, oscillation in an area of the intralaminar thalamus, measured in a fashion that carefully avoids conflation with activity associated with afferent corticothalamic activity, is predicted to be clearly coherent with phasic oscillatory BG input to that area, characterized by LFP in structurally connected areas of BG output structures.
Another key prediction is that in an over-trained task, phasic pallidal spiking to a particular thalamic target associated with onset of a particular salient context within the task will exhibit, in aggregate, a very stable, narrowly distributed (±<2 ms) delay relative to the first cortical spike volley associated with onset, implicating a stable set of striatal and pallidal/nigral neurons, for environmental conditions and level of arousal similar to those that prevailed during training.
Because the average delay through dorsal trans-GPi paths is roughly one gamma period, while the average dorsal trans-SNr delay is roughly one beta period, BGMS predicts functional prominence of the gamma cycle in a cortical area and scenario when its inputs to BG flow primarily through the dorsal striatum to the GPi (particularly implicating the dorsal sensorimotor striatum), while the beta cycle is expected to dominate when activity flows primarily to the SNr (implicating the associative striatum). Even longer delays, commensurate with the theta cycle, may accompany paths through the ventral striatum and ventral pallidum, due to its intimacy with the medial temporal lobe (briefly reviewed below).
More generally, according to BGMS the dominant oscillatory band of activity in a given cortical locus and context should be the best predictor of the BG paths it activates, and the converse should hold similarly. This principle generalizes beyond thalamocortical-BG circuits, to encompass frequency bands characteristic of parallel and analogous systems, several of which were discussed earlier (§13). These domains and characteristic frequency dynamics can be associated with particular thalamic nuclear complexes; for example, the mediodorsal nucleus is associated with beta synchronies, the pulvinar with alpha synchronies, and the anterior thalamic group with theta synchronies (Ketz et al. 2015).
Cross-frequency coupling has been demonstrated in theta band interactions of the hippocampus with striatum (Tort et al. 2008), and is posited to be a general theme in BG-thalamocortical dynamics (Cannon et al. 2014; Brittain and Brown 2014). Putative cross-frequency BGMS operates by spike-time-dependent gain in cortex no less than in-band BGMS, suggesting the corollary prediction that cross-frequency coupled BG activity in over-trained tasks produces spike volleys in target areas that are spatiotemporally coincident, at a regular frequency ratio, with selected corticocortical spike volleys. Heavy projections from the hippocampal formation to the ventral striatum (Brog et al. 1993) suggest that the in-band relationship may hold in these scenarios, i.e. that the dominant band of the afferent determines the activated BG path at the striatal stage, with divergence to a parallel path in a subsequent stage.
Lag-free long range synchronies in cortex (e.g. Vicente et al. 2008), with narrow pyramidal somatic coincidence windows (Pouille and Scanziani 2001; Volgushev et al. 1998), exist simultaneous with finite long range corticocortical delays (e.g. Gregoriou et al. 2009; Nowak and Bullier 1997). Exactly how does this work, at the level of cortical microcircuitry? How do the discharge and conduction delays of thalamocortical neurons and fibers compare as a function of nuclear origin? In particular, how do the delays of paths through the intralaminar and midline nuclei compare to those through other thalamic nuclei?
Gamma synchrony accompanies effective corticospinal activation (Schoffelen et al. 2005; Fries 2005), and the effects of single-pulse TMS stimulation in motor cortex on corticospinal activation depends on the phase of cortical beta at the moment of stimulation (Torrecillos et al. 2020). The pedunculopontine nucleus is integral to the control of voluntary movements (Tsang et al. 2010), and is profusely targeted by the GPi (Parent et al. 2001). Does this relationship entail BGMS? The same question applies to BG targeting of the superior colliculus, with regard to its attentional orientation and oculomotor functions.
Do the amygdala, hypothalamus, and other subcortical structures beyond those reviewed earlier (§13), use a BGMS-like mechanism to influence thalamocortical activity? The amygdala in particular has been construed as parallel to the ventral BG (Olmos and Heimer 1999), and indeed its central and medial nuclei are proposed to be continuous and homologous with the BG (Swanson and Petrovich 1998). Moreover, projections from the amygdala to PFC have been shown to convey signals that bias decision making (Burgos-Robles et al. 2017), similar to the role ascribed earlier (§6.2) to the BG.
The BGMS model implies an elaborate physiological arrangement of coordinated modularity, spanning all developmental levels, and many distinct neurotransmitter systems. How is this orchestrated? Rules governing the self-organization of projecting fiber populations and appositions must play a large part (e.g. Wedeen et al. 2012; Sanes and Yamagata 2009; de Wit and Ghosh 2015). But clearly, developmental exuberance, and activity-driven, correlation-sensitive plasticity must play a very large role. Exactly how are these development and plasticity mechanisms arranged to route and terminate long range fiber bundles appropriately, and optimize the timing of the stimulus-response functions of the BG as an ensemble?
How strong and broad is the BG influence on the cholinergic and serotonergic supplies to thalamus, cortex, and striatum? Are the BG arranged for bipolar control of these supplies, as they are for dopamine? And what are the topographies and microcircuitry of the BG inputs to TRN, NBM, PPN, LDT, DRN, MRN, and SC, by reference to the topographies and microcircuitry of their respective projections to and from thalamus and cortex?
The cytology and microcircuitry of the striatum are crucial to BGMS, and to basal ganglia dynamics in general.
How do the cortico-FSI and cortico-SPN projections to striatum differ in
cytological, laminar, and areal origin, in patterns of
preference, apposition, and topography/
What are the functions of striatal neuron types beyond the SPNs, FSIs, and ACh interneurons, particularly as they relate to BGMS? In particular, what are the roles of somatostatin-positive LTS interneurons, and of calretinin-positive interneurons, which have yet to be classified physiologically (Kreitzer 2009)? Uniquely human adult neurogenesis of striatal calretinin interneurons (Ernst et al. 2014) is intriguing — what is the functional significance of this?
Is oscillatory phase preserved in the paths through STN, GPe, NBM, SNl, and PPN/LDT, and if so, do they entrain their targets? Do the BG, through some paths, entrain targets to antiphase, to quickly and decisively abolish connections? Is this one of the functions of FSIs that target indirect path SPNs? Indeed, is this one of the functions of GPe input to striatum (which preferentially innervates FSIs) and TRN, and of STN to GPi/SNr? Such arrangements seem plausible, but the evidence is as yet tenuous—albeit tantalizing (e.g. Schmidt et al. 2013).
The maximum conduction velocity in human corpus callosum is anomalously slow (Caminiti et al. 2009). Schizophrenia is also typified by abnormalities of the corpus callosum, and of interhemispheric coordination (Foong 2000; Whitford et al. 2010; Hoptman et al. 2012). Are interhemispheric dynamics in humans special, from a BGMS perspective, or otherwise?
“Consciousness” is a notoriously slippery concept, even chimerical in many accounts. These narrative tribulations seem to be evidence of the fundamental qualities of conscious cognition. Where consciousness is broached above, most prominent are flexibility, integration and the breaching of modularity, intervention when modular strategies are flummoxed, and perhaps most unsettling, arrangements of notionally infinite recurrence.
Accordingly, consciousness is here construed to be an evolving pattern of relations, featuring many degrees of freedom (high dimensionality), arbitrary information combination, self-acting and self-affecting selections (decisions), and self-referential iteration (self-causation). More exhaustively, it is a mental mechanism that features the partly overlapping attributes and facilities of uniqueness, representation, genericness, ephemeral specificity, arbitrary associativity, intentionality, attention, perception, episodic continuity, and action. Evidently, conscious actions can be inwardly directed (chiefly, cognitive transformations, recollections, memorizations, and decisions, all relative to current patterns of activity) or outwardly directed (behaviors). Reportability and self-awareness, frequently attributed specially to human consciousness, are (by the present narrative) corollary. “Uniqueness” here signifies that there is only one consciousness in the normal waking brain (largely a corollary of its arbitrary associativity, and the physiology underlying that facility), and that the representations and core mechanisms of conscious cognition lack any architectural modularity dividing them into perceptive, cognitive, and active domains, but rather that these domains are all directly and irreducibly implicated by the same physiological substrate (as seen, for example, in the thalamic intralaminar nuclei, noted earlier (§7.9)). For a discussion of the capacities at issue, and of the inextricable entanglement within consciousness of the attributes and facilities attributed to it above, see Engle (2002).
Within the network of cortical connectivity hubs, particular areas and networks have been identified that are associated with facilities attributed above to consciousness. For example, Vincent et al. (2008) propose a “frontoparietal control network” comprising lateral PFC, anterior cingulate cortex, and the inferior parietal lobule, topographically and topologically separate from the hippocampal network and “dorsal attention” network. PFC and posterior parietal cortex (PPC) in particular have been implicated in theories of the physiological basis of fluid intelligence (Jung and Haier 2007). Hub networks are uniquely developed in humans, and genes identified as uniquely divergent in humans are highly expressed in these networks, are associated with individual variation in the function of these networks, and are associated with intelligence and schizophrenia (Wei et al. 2019).
The integrated system of the PFC and striatum, suggested to be central to cognitive flexibility (Leber et al. 2008; van Schouwenburg et al. 2010b, 2012, 2014; Hazy et al. 2006), includes many connectivity hubs and resting state network nodes (Cole et al. 2010; van den Heuvel and Sporns 2011; Harriger et al. 2012; van den Heuvel et al. 2009; Elston 2000). Indeed, correlated activity has been demonstrated between cortical resting state network nodes and loci distributed widely in the striatum (Di Martino et al. 2008; Vatansever et al. 2016; Wang et al. 2017‡).
Posterior cortex, and its relationship with striatum, are also clearly implicated in consciousness. As noted noted earlier (§11.4), LFP activity in a network spanning deep layers of lateral intraparietal cortex, caudate striatum, and intralaminar thalamus, evaluated with the integrated information measure Φ, reliably indexes state of consciousness (Afrasiabi et al. 2020‡). Functionally connected triads, in which posterior parietal cortex drives frontal cortex and the BG, are implicated in motor performance (Hwang et al. 2019), and the evidence from Afrasiabi et al. (2020‡), and related results from Redinbaugh et al. (2020), suggest such triads may indeed be necessary for consciousness.
The modularity of an individual's resting state cortical networks, measured by fMRI, has been found to be predictive of task performance as a function of task complexity, with high modularity subjects exhibiting relatively high performance on simple tasks, and low modularity subjects exhibiting relatively high performance on complex ones (Yue et al. 2017). In general, the establishment of effective connections between dissociable networks, heralding a collapse of their mutual modularity, is associated with conscious awareness (Godwin et al. 2015), and particularly entails functional integration of task-specific networks with the resting state network (Fukushima et al. 2018), while pharmacologically induced loss of consciousness is associated with the pervasive breakdown of effective connectivity in cortex (Ferrarelli et al. 2010b).
It has been proposed that PFC is organized into a spatially graded hierarchy, with the highest and most abstract representations located anteriorly, and the lowest and least abstract located posteriorly (Christoff and Gabrieli 2000; Badre and D'Esposito 2009). In the least abstract of these areas, numerous visuotopic maps (for example) have been identified (Silver and Kastner 2009), whereas the frontopolar cortex is concerned with highly abstract consideration and reconciliation of conflicting goals, and the management of competing and alternating cognitive sets (Mansouri et al. 2017). Functional specialization of the dorsomedial and dorsolateral PFC has been proposed, with the former monitoring performance, and the latter guiding it; links between these areas exhibit mutual preferences according to position along the anterior-posterior axis (Taren et al. 2011). Within the DLPFC, subdivisions are apparent from their functional correlates and network connectivity—an anterior-ventral subregion is associated with attention and action inhibition processes, and is intimate with anterior cingulate cortex, while a posterior-dorsal one is associated with action execution and working memory, and is intimate with PPC (Cieslik et al. 2013).
In general, the primate cortex is characterized by gradients in average connectivity distance, with neurons in primary sensory and motor areas typified by the shortest average connection distances, while areas most remote from primary areas are typified by the most distant connections, particularly implicating lateral and medial frontoparietal cortex (Margulies et al. 2016; Oligschläger et al. 2019).
Goldman-Rakic (1988) described pervasive, systematic interdigitation of parallel circuits linking association cortex. Consistent with this account, several of the widely distributed networks identified in humans—resting state, frontoparietal control, and dorsal attention—have been shown to consist of at least two distinct parallel networks with similar gross structure, but spanning distinct interdigitated subregions, and with intriguing topological distinctions; e.g., one of the identified resting state subnetworks is intimate with the hippocampal system, while another is devoid of such intimacy (Braga and Buckner 2017; Braga et al. 2019).
Clearly this topographic and topological structure has consequences for mental architecture. Indeed, the microstructural characterization of projections between hub areas is among the most promising subjects for future investigation.
Perhaps anterior and medial PFC, the posterior medial and parietal cortex intimate with it, and the basal ganglia areas intimate with them, contain areas in which domain-specific functional topography is a purely transitory consequence of their effective connections from moment to moment, given highly abstract topographies along lines of hierarchy and other generic aspects of cognition. Evidence suggests as much. Using fMRI in humans, Assem et al. (2019‡) show that much of the frontoparietal control network, with adjacent areas within the default mode and dorsal attention networks, the head of the caudate nucleus, and substantial sectors of the cerebellum, are activated in cognitively demanding tasks regardless of the specific domains implicated by the task, suggesting the role of these areas in cognition is generic, rather than domain-specific. Using tracers in monkeys, Averbeck et al. (2014) show that the head of the caudate nucleus is a zone of convergence for projections from all parts of the PFC, which they suggest is evidence that the striatum also contains topological hubs that perform computations across diverse domains.
Additional fMRI evidence supports the proposition that hub areas are engaged in cognition generically. Evidence suggests that functional representations in hub areas are dynamically oriented to task demands (Vromen et al. 2018‡), and that abstract relations in verbal comprehension are represented generically in hub areas (Zhang et al. 2019a‡). This versatility seems to follow from the capacity of individual neurons, and indeed individual synapses, to participate in a vast array of distinct ephemeral assemblies of neurons with contextually appropriate conduction delays (Rigotti et al. 2013; Izhikevich 2006), even simultaneously (Naud and Sprekeler 2018; Caruso et al. 2018; Bernardi et al. 2018‡).
The proposition that the functional topography of hub areas is highly abstract and context-dependent also comports with the view of van den Heuvel et al. (2012) that a core network of connectivity hubs (a “rich club”) serves as a common, and therefore contentious, communication “backbone” subject to “greedy routing” strategies by more locally connected (and specialized) areas. Indeed, task-related activity in functionally connected PFC and PPC can be very similar, with almost identical tuning and time courses, throughout the performance of a task implicating working memory (Chafee and Goldman-Rakic 1998), and beta band synchronies between these areas reflect only behaviorally relevant representations, with PFC neurons synchronized to PPC beta oscillation only if selective for contextually relevant categories (Antzoulatos and Miller 2016). The view that cortical areas with high abstraction and long range connectivity function as thoroughfares also follows from findings, noted earlier (§1.4), that frontal-posterior LFP synchrony accompanies attentional orientation, whether by top-down or bottom-up processes (Buschman and Miller 2007). Unsurprisingly, hub areas are disproportionately implicated in systemic brain disease and mental illness (Crossley et al. 2014).
Activity in particularly abstract areas of the PFC might arrange itself to impart nearly arbitrary patterns to the striatum, inducing highly flexible transformations by the BG of cortical activity and effective connectivity, and resolving backbone contention through selections, largely by the BGMS mechanism. This may largely underlie selective, task-related output from hub areas despite unselective inputs (Senden et al. 2018, 2017). By this narrative, the BG make available an enormous repertoire of neural gestures, that activity in hub areas can use to gate and operate on network activity, particularly activity within hub areas themselves. The corticostriatal projections from hub areas, and the input-output relations of the targeted areas of striatum, are then the essential substrate for cognitive flexibility, as suggested by van Schouwenburg et al. (2014). Indeed, as noted earlier (§11.9), cognitive flexibility is associated with functional connectivity of hub areas to BG (Vatansever et al. 2016). Moreover, significant dysfunction in this relationship, including both deficient cortical control of striatum, and deficient striatal control of cortex, has been shown in Sz (Wang et al. 2015).
One interpretation of these arrangements is that consciousness in its essence is an enormously flexible and agile mechanism for relating causes (stimuli) to effects (resulting thoughts and behaviors). By this narrative, conscious contemplation occurs when these ephemeral cause-effect relations are chained together, each stirring the next into existence, so that consciousness depends intrinsically on the physical causality of brain activity. This comports neatly with the view, noted earlier (§11.4), that the “physical substrate of consciousness” exhibits maximal cause-effect power (Tononi et al. 2016).
As suggested earlier (§11.7), if working memory items are patterns of activation in PFC, each characterized by a distinct phase angle (Siegel et al. 2009), then—through corticocortical feedback projections and BGMS—the PFC might establish and dissolve effective connections implicating a particular working memory item (effectively, a thought) with little or no interference from, or indeed to, latent items, except when a latent item is selected for integration with an active item. DA under striosomal control, and ACh and 5-HT under PFC and BG control, are also crucial parts of this tool set, dynamically tuning the receptivity, contrast, focus, selectivity, and stability, of cortical signal paths.
Because inputs to the BG encompass the entire cortex, the BG can respond to activity in areas that are not functionally connected (not synchronized) with activity in conscious areas, and might act to synchronize the latter with the former, or the former with the latter. In this way, subconscious activity might be boosted into consciousness by BG selections. Indeed this likely describes any BG-mediated reorientation of attention in response to a sensory stimulus, particularly implicating the thalamic intralaminar nuclei.
The propositions that the highest levels of PFC lack persistent domain-specific functional topographies, control the BG with great flexibility, and are directly implicated in conscious cognition, relate to the dynamics of skill learning and performance. Experience-driven skill acquisition entails the reorganization of cortical topography in sensory and motor areas (Buonomano and Merzenich 1998; Kleim et al. 2004). Topographic reorganization seems to necessitate hub areas without fixed functional topography, in order to maintain function while accommodating the shifting semantic correlates of the neurons comprising the implicated map. In principle, long range connections linking shifting maps to generic hubs might enable sensible integration into cognition at every stage of topographic reorganization.
The orchestration of topographic plasticity, and functional continuity during that process, likely implicate not only highly abstract areas of neocortex, but also the hippocampal formation, which is extremely labile and exceptionally well-connected (as briefly reviewed earlier (§13.2.1)), and has direct and transthalamic links to secondary motor cortex (Jankowski et al. 2013; Van der Werf et al. 2002). Perhaps the involvement of the hippocampal system in the domain of spatial navigation is just a special case of a general competence “navigating” the similarly 2 dimensional graded maps of neocortex—that is, the role of the hippocampal system in spatial navigation is actually to physically register associations between spatial locations and semantic objects, represented as specific large scale patterns of strengthened connectivity (mutual excitability) in neocortex, and this facility is readily suited to register associations among such semantic objects, with no particular association with physical space (Eichenbaum and Cohen 2014). This facility seems perfectly suited to recruitment as scaffolding for the reorganization of graded maps in neocortex, including that underlying motor expertise (see e.g. Schapiro et al. 2019; Burman 2019, 2018‡).
Initial performance of a qualitatively new skill depends on the availability and engagement of working memory (Reber and Kotovsky 1997), and is aided by attention to the minutiae of performance (Beilock et al. 2002). Learning the skill does not entail topographic reorganization until late in the process, and initially pivots on activity and plasticity in the BG and cerebellum (Ungerleider et al. 2002), with a crucial role for corticostriatal SPN plasticity (Koralek et al. 2012). Once proficiency is attained, performance can in fact be significantly disrupted by attention (Beilock et al. 2002). If the BG learn precise sensorimotor sequences through practice (Graybiel 1998), and their performance involves finely tuned subcortical loops, then inapt engagement of high-order PFC, supplying disruptive signals to the striatum, is likely to disrupt overall performance. Similar disruption of input patterns to the cerebellum would have similar consequences.
It seems likely that the arrangement of high resolution spatially graded feature maps, with a “rich club” topology of dense high resolution interconnections, hub areas some of which are never plastically committed as feature maps, and dynamic timing-based mesoscopic control of effective connectivity and signaling characteristics by a recurrent, highly convergent-divergent multistage subsystem arranged for self-referential reinforcement learning, is not unique to mammals, or even to vertebrates, but rather is the essential architecture of many evolved conscious systems. Perhaps the resulting information processing is consciousness, in the sense meant by Tegmark (2015) in his proposition that “consciousness is the way information feels when being processed in certain complex ways”.
And if the predicates of consciousness can be realized by a generic architecture, that also suggests that instances of this architecture are likely wherever there are organisms exhibiting complex and flexib