Spikes, synchrony, and attentive learning by laminar thalamocortical circuits

Abstract

This article develops the Synchronous Matching Adaptive Resonance Theory (SMART) neural model to explain how the brain may coordinate multiple levels of thalamocortical and corticocortical processing to rapidly learn, and stably remember, important information about a changing world. The model clarifies how bottom-up and top-down processes work together to realize this goal, notably how processes of learning, expectation, attention, resonance, and synchrony are coordinated. The model hereby clarifies, for the first time, how the following levels of brain organization coexist to realize cognitive processing properties that regulate fast learning and stable memory of brain representations: single-cell properties, such as spiking dynamics, spike-timing-dependent plasticity (STDP), and acetylcholine modulation; detailed laminar thalamic and cortical circuit designs and their interactions; aggregate cell recordings, such as current source densities and local field potentials; and single-cell and large-scale inter-areal oscillations in the gamma and beta frequency domains. In particular, the model predicts how laminar circuits of multiple cortical areas interact with primary and higher-order specific thalamic nuclei and nonspecific thalamic nuclei to carry out attentive visual learning and information processing. The model simulates how synchronization of neuronal spiking occurs within and across brain regions, and triggers STDP. Matches between bottom-up adaptively filtered input patterns and learned top-down expectations cause gamma oscillations that support attention, resonance, learning, and consciousness. Mismatches inhibit learning while causing beta oscillations during reset and hypothesis testing operations that are initiated in the deeper cortical layers. The generality of learned recognition codes is controlled by a vigilance process mediated by acetylcholine.

Introduction

This article proposes how the brain coordinates multiple levels of thalamocortical and corticocortical processing to rapidly learn, and stably remember, important information about the world. The Synchronous Matching Adaptive Resonance Theory (SMART) model that is presented here shows how bottom-up and top-down pathways work together to accomplish this goal by coordinating processes of learning, expectation, attention, resonance, and synchrony. In particular, SMART explains how attentive learning requirements are realized by detailed brain circuits, notably the layered organization of cells in neocortical circuits and how they interact with first-order (e.g., the lateral geniculate nucleus, LGN) and higher-order (e.g., the pulvinar nucleus, PULV; Sherman and Guillery, 2001, Shipp, 2003), and nonspecific thalamic nuclei (van Der Werf et al., 2002).

Corticothalamocortical pathways work in parallel with corticocortical routes (Maunsell and Van Essen, 1983, Salin and Bullier, 1995, Sherman and Guillery, 2002). Specific first-order thalamic nuclei relay sensory information to the cerebral cortex, whereas specific second-order thalamic nuclei receive their main input from layer 5 of lower-order cortical areas and relay this information to higher-order cortical areas (Sherman and Guillery, 2002, Fig. 1a).

The SMART model clarifies how a match at the specific first-order and higher-order thalamic nuclei may induce fast learning and stable memory of neural representations in the thalamocortical system (cf., Gove et al., 1995, Grossberg, 1980, Grossberg, 2003). Such a match may occur, for example, at LGN cells in response to bottom-up driving retinal inputs and top-down modulatory expectations from layer 6 of cortical area V1 (Sillito et al., 1994). At a higher level of brain organization, a match may occur at pulvinar cells in response to driving bottom-up inputs from layer 5 of V1 (Rockland et al., 1999) and top-down modulatory cortical inputs from layer 6 of V2. The model proposes how this bottom-up/top-down matching process can allow bottom-up and top-down feedback loops to cause a persistent resonant state which supports spike synchronization in the gamma frequency range (20–70 Hz). Such an oscillation frequency is fast enough to support spike-timing-dependent plasticity (STDP; Levy and Steward, 1983, Markram et al., 1997, Bi and Poo, 2001), since STDP is maximal when pre-synaptic and post-synaptic cells fire within 10–20 ms of each other (Traub et al., 1998, Wespatat et al., 2004). In contrast, during a mismatch, slower beta frequency (4–20 Hz) oscillations are caused. STDP is disabled at this lower frequency. The model hereby proposes how thalamocortical matching, resonant feedback, synchronous oscillations, and STDP learning may be coordinated, notably how match-sensitive differences in oscillation frequency can enable or disable learning.

The matching process is carried out by a top-down, modulatory on-center, off-surround circuit (Grossberg and Stone, 1986, Carpenter and Grossberg, 1987, Carpenter and Grossberg, 1991, Grossberg, 1995, Grossberg, 1999a) which selects a critical feature pattern of attended features, while inhibiting unattended features. This process clarifies how attention carries out a form of “biased competition” (Desimone and Duncan, 1995, Desimone, 1998). The attended feature patterns are the ones that can be rapidly learned in the adaptive weights of bottom-up adaptive filters and top-down expectations. In the case of a partial mismatch, there may simultaneously be cells at which matching and learning occurs, as well as other cells at which mismatch, inhibition, and suppression of learning occurs. Thus, in describing match vs. mismatch states, one needs to understand that there may be cells at which bottom-up and top-down signals mismatch, even though there is a good enough partial match for a synchronous resonant state to persist long enough for STDP to occur at the matched features.

If a mismatch between bottom-up and top-down signal patterns is large enough, it prevents such a synchronous resonant state from developing. Within the model, resonance is prevented when mismatch causes a rapid reset of ongoing information processing, and triggers a memory search, or hypothesis testing, for uncommitted cells, or an already familiar recognition category that can better match bottom-up data. In particular, such a memory search can enable either a totally new recognition category to be learned, or a learned refinement of the critical features that can activate an already familiar recognition category. Thus, the model proposes that there are cycles of resonance and reset, with resonance supporting learning, and reset driving hypothesis testing that leads away from poorly matched states to better ones.

START clarifies how such a memory search may be controlled by an interaction between specific thalamic nuclei, nonspecific thalamic nuclei, and the cerebral cortex. The SMART model (Fig. 2, Fig. 3) predicts how a big enough mismatch at a specific thalamic nucleus can generate a novelty-sensitive burst of activation at a nonspecific thalamic nucleus. Nonspecific thalamic nuclei, such as the midline and intralaminar nuclei (van Der Werf et al., 2002), as well as “matrix” cells in the specific thalamic nuclei (Jones, 2002), derive their name from the fact that they receive diffuse innervations from the sensory periphery and the reticular formation, and project diffusely to the superficial layers of the cerebral cortex (Fig. 1b).

In particular, the nonspecific thalamic nuclei are predicted to generate reset signals in the form of novelty-sensitive bursts of activation during mismatch episodes. Such a burst is broadcast nonspecifically to the superficial layers of the cerebral cortex, notably layer 1. The nonspecific burst is sensed by dendrites in layer 1 of cortical layer 5 cells. The model explains how the burst leads to a reset event by propagating from layer 1 dendrites via their layer 5 cells to layer 6 and then on to layer 4, shutting down previously active cells there, and thereby enabling a different pattern of activation to take hold in layer 4. This reset event causes a slower beta oscillation frequency in the model. Thus the reset event prevents learning of poorly matched bottom-up and top-down information, both by inhibiting the active learned categorical representations whose top-down expectations led to the mismatch, and also by creating a slower oscillation frequency to which STDP is insensitive. The details of how this works will be described below.

As noted above, the SMART model predicts that the reset event is expressed in the deeper layers of cerebral cortex, such as layers 4 to 6, and may thereby initiate slower beta oscillations in these layers. The more superficial cortical layers (e.g., layers 2/3) may, in contrast, express faster gamma oscillations. The model supports its proposal about how match-sensitive differences in oscillation frequency can enable or disable learning by quantitatively simulating data about single-cell biophysics, pharmacology, and neurophysiology; laminar neuroanatomy; aggregate cell recordings, such as current source densities and local field potentials; large-scale oscillations at beta and gamma frequencies; and functionally links them all to requirements about how to achieve fast attentive learning and stable memory. It is also suggested below how to directly test this prediction.

Many authors have examined synchronous oscillations within and across brain regions as one way in which behaviorally significant brain states are organized (Engel et al., 2001). Aggregate and single-cell recordings from multiple thalamic and cortical levels of mammals have shown high-frequency and low-frequency rhythmic synchronous activity correlated with cognitive, perceptual and behavioral tasks. In addition, large-scale neuronal population models have been proposed to model oscillatory dynamics (Bazhenov et al., 1998, Lumer et al., 1997, Destexhe et al., 1999, Siegel et al., 2000). However, these models do not link brain spikes, oscillations, and self-stabilizing STDP with the brain states that subserve attentive cognitive information processing.

The SMART model fills this gap. It clarifies data about how bottom-up processing and learned tuning of adaptive filters is modulated by top-down attentive learned expectations that embody predictions or hypotheses that focus attention on expected bottom-up stimuli (Salin and Bullier, 1995, Engel et al., 2001, Gao and Suga, 1998, Krupa et al., 1999, Desimone, 1998, Ahissar and Hochstein, 2002, Herrmann et al., 2004). These data support predictions of Adaptive Resonance Theory, or ART (Grossberg, 1980, Grossberg, 2003, Carpenter and Grossberg, 1987, Carpenter and Grossberg, 1991, Carpenter and Grossberg, 1993, Carpenter et al., 1991) that top-down expectations regulate predictive coding and matching and thereby help to focus attention, synchronize and gain-modulate attended feature representations, and trigger fast learning that is dynamically buffered against catastrophic forgetting. The goal of achieving fast stable learning without catastrophic forgetting is often summarized as the stability–plasticity dilemma (Grossberg, 1980). The stability–plasticity dilemma must be solved by every brain system that needs to rapidly, yet stably, learn about the flood of signals that subserves even the most ordinary experiences. If the brain's design is parsimonious, then we should expect to find similar principles operating in all the brain systems that can stably learn an accumulating knowledge base in response to changing conditions throughout life.

ART has predicted that some fundamental properties of human and animal perception and cognition are part of the brain's solution of the stability–plasticity dilemma. In particular, humans are intentional beings who learn expectations about the world and make predictions about what is about to happen. Humans are also attentional beings who focus processing resources upon a restricted amount of incoming information at any time. Why are we both intentional and attentional beings, and are these two types of processes related? The stability–plasticity dilemma and its solution using resonant states provides a unifying framework for understanding these issues.

In particular, ART predicted that there is an intimate connection between the mechanisms that enable us to learn quickly and stably about a changing world, and the mechanisms that enable us to learn expectations about such a world, test hypotheses about it, and focus attention upon information that we find interesting. ART also proposes that, in order to solve the stability–plasticity dilemma, only resonant states can drive rapid new learning, which gave the theory its name.

Fig. 4 illustrates these ART ideas in a simple two-level example whose anatomical, physiological, and pharmacological substrates are clarified by SMART. Here, a bottom-up input pattern, or vector, I activates a pattern X of activity across the feature detectors of the first processing stage F₁. For example, a visual scene may be represented by the features comprising its boundary and surface representations (Cao and Grossberg, 2005, Grossberg, 1994, Grossberg and Yazdanbakhsh, 2005). This feature pattern represents the relative importance of different features in the inputs pattern I. In Fig. 4a, the pattern peaks represent more active feature detector cells, the troughs less activated feature detectors. This feature pattern sends signals S through an adaptive filter to the second level F₂ at which a compressed representation Y (also called a recognition category, or a symbol) is activated in response to the distributed input T. Input T is computed by multiplying the signal vector S by a matrix of adaptive weights that can be altered through learning. The representation Y is compressed by competitive interactions across F₂ that allow only a small subset of its most strongly activated cells to remain active in response to T. The pattern Y in the figure indicates that a small number of category cells may be activated to different degrees. These category cells, in turn, send top-down signals U to F₁ (Fig. 4b). The vector U is converted into the top-down expectation V by being multiplied by another matrix of adaptive weights. When V is received by F₁, a matching process takes place between the input vector I and V which selects that subset X⁎ of F₁ features that were “expected” by the active F₂ category Y. The set of these selected features is the emerging “attentional focus”.

If the top-down expectation is close enough to the bottom-up input pattern, then the pattern X⁎ of attended features reactivates the category Y which, in turn, reactivates X⁎. The network hereby locks into a resonant state through a positive feedback loop that dynamically links, or binds, the attended features across X⁎ with their category, or symbol, Y. Figs. 4c and d shows how such an ART circuit can search for a novel or better marching recognition category if there is not a good enough match.

This match-based learning process is the foundation of the stability of learned memories, both bottom-up categories and top-down expectations, in an ART model. Match-based learning allows memories to change only when input from the external world is close enough to internal expectations, or when something completely new occurs. This feature makes ART systems well suited to problems that require online learning of large and evolving databases. For example, ART systems have been applied in fields ranging from technological solutions in industrial design and manufacturing, to the control of mobile robots, to remote sensing land cover classification (see Carpenter et al., 2005 for a review).

Reconciling distributed and symbolic representations using resonance. ART models also clarify fundamental issues concerning symbol grounding, in addition to intentional and attentional aspects of primate cognition. The individual features at F₁ have no meaning on their own, just like the pixels in a picture are meaningless one-by-one. The bottom-up category, or symbol, in F₂ is sensitive to the global patterning of these features, but it cannot represent the “contents” of the experience, including their conscious qualia, due to the very fact that a category is a compressed, or “symbolic” representation. The bottom-up/top-down resonance between these two types of information converts the pattern of attended features into a coherent context-sensitive state that is linked to its category through feedback. It is this coherent state, that joins together distributed features and symbolic categories, that can enter consciousness. ART predicts that all conscious states are resonant states. In particular, such a resonance binds spatially distributed features into either a synchronous equilibrium or oscillation, until it is dynamically reset. Such synchronous states were predicted in the 1970′s in the articles which introduced ART (see Grossberg, 1999b, Grossberg, 2003 for data reviews). The SMART model simulates finer properties of synchronous oscillations and their reset in the form of gamma and beta oscillations.

Recent ART models, called LAMINART, began to show how ART predictions may be embodied in laminar cortical circuits (Grossberg, 1999a, Grossberg, 2003, Raizada and Grossberg, 2003). These LAMINART models unify properties of visual development, learning, perceptual grouping, attention, and 3D vision. They did not, however, incorporate spiking dynamics, higher-order specific thalamic nuclei and nonspecific thalamic nuclei, control mechanisms for regulating resonance vs. reset, or pharmacological modulation of learning. The SMART model goes beyond ART and LAMINART models by showing how these properties naturally coexist in the LAMINART framework. In particular, SMART explains and simulates how laminar cortical circuits may interact with specific primary and higher-order thalamic nuclei and nonspecific thalamic nuclei to control match vs. mismatch processes that regulate recognition learning and dynamically buffer learned memories against catastrophic forgetting; how spiking dynamics are incorporated into synchronous oscillations whose oscillation frequencies can provide an additional degree of freedom for controlling cognitively-mediated operations such as matching and fast learning; and how acetylcholine-based processes may embody predicted properties of vigilance control that regulate the generality of learned recognition categories in a way that is sensitive to changing environmental statistics, using only locally computed signals in the network. Table 1 illustrates the new ART operations implemented in the SMART circuitry. Fig. 5 depicts anatomical pathways that are predicted to subserve the arousal, reset, search, and vigilance operations.

What is vigilance and why is it needed? It is not enough to just regulate the stability of learned memories in a changing world. Survival requires that a human or animal learn to correctly discriminate, recognize, and predict important objects and events. An effective learner must be sensitive to changing environmental statistics and feedback that determine how specific or general learned knowledge must be to control and predict the external environment. How does the brain determine how specific (concrete) or general (abstract) a learned recognition category should be in a given situation? If matches trigger learning, then a flexible, situationally-sensitive, criterion of matching is needed to control specific vs. general learning. Such a criterion has been called vigilance (Carpenter and Grossberg, 1987, Carpenter and Grossberg, 1991), corresponding to the intuition that higher vigilance enables finer discriminations to be made. In all ART models, including SMART, high vigilance triggers reset and search for a new category when even small mismatches occur, thereby leading to concrete learning. Low vigilance allows even coarse matches to trigger resonance, and to thereby learn abstract categories that respond to many input variations. What is new in SMART is the prediction that neuromodulation by acetylcholine (ACh) may regulate the level of vigilance through time.

The remainder of this section specifies in greater detail how SMART model circuits work. SMART clarifies how retinal inputs activate the thalamus, and from there, the cortex, through two separate pathways, a specific pathway targeting middle cortical layers (LGN core cells to layers 4 and 6^I cells, a subdivision of layer 6, see Table 2), and a nonspecific pathway targeting superficial layers (LGN matrix cells and nonspecific thalamic nucleus to layer 1 of V1). These two pathways are treated separately due to the different functional roles that were outlined in the previous section.

The SMART specific pathway includes both specific first-order and second-order thalamic nuclei projecting to the middle layers of the cerebral cortex (Jones, 2002). Specific thalamic nuclei are often divided into first-order relays, such as the LGN, which receive inputs from the sensory periphery, and second-order relays, which receive their main inputs from the cerebral cortex (Sherman and Guillery, 2002). Although the largest part of the thalamus consists of second-order relays, the most widely studied structures are the first-order thalamic nuclei. As a consequence, thalamic nuclei are usually seen as relay stations of information from the sensory periphery to the cerebral cortex. This picture is misleading. For instance in the LGN, a first-order relay nucleus, the retina contributes only 5–10% of the total afferents (Sherman and Guillery, 2001). The pulvinar (PULV), one of the largest second-order thalamic nuclei, receives only minimal afferents from the sensory periphery. Most of its inputs originate from the cerebral cortex and the superior colliculus (SC). The LGN receives a massive cortical projection from V1 cortical layer 6, and the PULV receives afferents from layers 5 and 6 of several cortical areas (Rockland, 1998, Wang et al., 2002, Shipp, 2003).

Driving vs. modulatory pathways: Round-large vs. round-small terminals. In the primate, synaptic terminals in the thalamus can be roughly subdivided into two classes (Rockland, 1996, Sherman and Guillery, 2001): (a) round-large (RL) synapses, such as retinogeniculate synapses. These synapses are believed to be driving; (b) round-small (RS) terminations, such as the corticothalamic synapses from V1 layer 6 to the LGN. These synapses are believed to be modulatory. Terminations arising from layer 5 to a second-order thalamic nucleus are similar to retinogeniculate RL synapses, or driving, connections, often found in more proximal segments of the dendrites. This dual pattern of connectivity seems to be constant across species (Rouiller and Welker, 2000). A functional correlate of the distinction between RL and RS synapses is that, whereas lesioning a cortical area that innervates the thalamus through layer 6 alone does not change the receptive field property of the thalamic cell, lesioning an area that innervates the thalamus through layer 5 does abolish the receptive field of the cell (in, for example, areas 17, 18 and 19; Sherman and Guillery, 2002, Soares et al., 2004). In addition, the observed receptive fields in the PULV resemble those of complex cells in visual cortex (binocular and direction selective).

In the SMART specific pathway, LGN core cells are driven by bottom-up sensory inputs and excite both layer 4 and layer 6^I (Fig. 2a). Layer 6^I, in turn, contrast-normalizes layer 4 cell activities in response to bottom-up input patterns (Grossberg, 1980, Heeger, 1992, Douglas et al., 1995) via a modulatory on-center, driving off-surround network (Carpenter and Grossberg, 1987, Grossberg, 1980, Grossberg, 2003) whose off-surround is mediated by layer 4 inhibitory interneurons (Grieve and Sillito, 1991). The direct pathway from LGN to layer 4 enables the cortex to fire despite the modulatory nature of the on-center from layer 6^I to 4. The on-center off-surround of the LGN→6^I→4 pathway biases the emergence of orientation sensitivity in layer 4 cells that spike after the arrival of the LGN input within the STDP learning window (see Section 2.1).

Top-down matching, attention, and learning. Top-down feedback pathways coexist with bottom-up pathways in the brain. SMART proposes that top-down feedback from layer 6^II of V1 to the LGN controls attention and plasticity in both the bottom-up adaptive filter pathways from LGN to V1 and in the top-down expectation pathways (Fig. 2b). As in previous ART models, SMART corticothalamic feedback is realized by a top-down, modulatory on-center, driving off-surround circuit whose on-center helps to create an attentional focus that selects, enhances, and synchronizes behaviorally relevant, bottom-up sensory inputs (match), and whose off-surround suppresses inputs that are irrelevant (mismatch).

The processing that goes on between LGN and V1 has homologs in the processing by PULV and V2, and beyond. Bottom-up driving inputs to higher-order specific thalamic nuclei, such as the PULV, arise from layer 5 of V1, as indicated in Fig. 3a (Salin and Bullier, 1995, Callaway, 1998). Top-down feedback from layer 6^II (see Table 2) of V2 to PULV can match the bottom-up input pattern from V1 layer 5 in a manner similar to how top-down feedback from layer 6^II of V1 matches retinal input in the LGN (Fig. 3, Fig. 2, respectively).

Accumulating experimental evidence supports the ART prediction (Carpenter and Grossberg, 1987, Grossberg, 1980, Grossberg, 1999a, Grossberg, 2003) that that top-down attentional signals are mediated by a modulatory on-center, off-surround network. Both V2→V1 feedback (Bullier et al., 1988) and V1→LGN feedback (Sillito et al., 1994) possess this structure. A similar modulatory on-center, off-surround architecture has been observed in feedback interactions from auditory cortex to the medial geniculate nucleus (MGN) and the inferior colliculus (IC) (Zhang et al., 2004, Gao and Suga, 1998). Consistent with the ART prediction of the role of attention in controlling adult plasticity, Gao and Suga (1998) found that acoustic stimuli caused plastic changes in the IC of bats only when the IC received top-down feedback from auditory cortex. Moreover, plasticity is enhanced with behaviorally relevant auditory stimuli, consistent with the ART proposal that top-down feedback allows matched, and therefore attended, critical feature patterns to be learned, while suppressing mismatched, and thus unattended, features. Nicolelis and colleagues have shown that cortical feedback also controls thalamic plasticity in the somatosensory system (Krupa et al., 1999).

ART also predicted that matching synchronizes the firing patterns of cells coding matched stimuli and thereby facilitates fast stable learning (cf., Engel et al., 2001, Fries et al., 2001, Grossberg, 1976, Grossberg, 1980, Grossberg, 1999a, Pollen, 1999, Usrey, 2002). SMART further develops that proposal to include spiking neurons and the role of the higher-order specific and nonspecific thalamic nuclei.

SMART clarifies how the thalamic reticular nucleus (TRN) mediates the off-surround that helps to select thalamic cells during the matching process (Fig. 2b). The TRN forms a shell around the lateral and dorsal portions of the thalamus, lying in the axonal path connecting the specific and nonspecific thalamus and the cortex (Guillery and Harting, 2003). Afferents to the TRN are mainly branches of bottom-up axons from a specific thalamus to its target cortex, or branches of top-down axons from cortical layer 6 to its specific thalamic nucleus. Notably, the TRN does not receive projections from layer 5. The TRN has a rather uniform local structure. TRN cells are GABAergic, and are reciprocally linked both by chemical inhibitory projections and by electrical synapses (Landisman et al., 2002). Top-down inhibitory feedback from the TRN to specific thalamic nuclei helps to balance top-down cortical layer 6 excitatory signals at their shared target cells (Fig. 2, Fig. 3), and thereby enables the excitatory signals to have only a modulatory effect on these cells (Guillery and Harting, 2003) when these are the only active inputs. In addition to projecting to first-order and higher-order specific thalamic nuclei (Guillery and Harting, 2003), the TRN also projects to the nonspecific intralaminar and midline thalamic nuclei (Kolmac and Mitrofanis, 1997); see Fig. 2, Fig. 3. TRN projections to the intralaminar and midline nuclei are more diffuse than the reticular projection to the specific dorsal thalamic nuclei. The TRN is known to influence a number of important brain processes. In particular, it influences the sleep/wake cycle (Steriade et al., 1993), the efficacy of thalamic inputs to the cortex (Nicolelis and Fanselow, 2002, Swadlow et al., 2002), and attention (Sherman and Guillery, 2001). The current article focuses on the latter two processes, which are clearly relevant to the sleep/wake cycle, while suggesting an additional role for the TRN in suppressing unmatched features in recognition and learning.

Both V1 layer 2/3 and PULV inputs are required to fully activate the SMART V2 area. V1 layer 2/3 and PULV can drive V2 layer 4, which in turns activates V2 layers 2/3, 5 and 6^II, whose axons project to the PULV, where V1 input from layer 5 is attentively matched against the layer 6^II feedback (Figs. 3a and c). Layer 5 of V1 excites the matrix cells (see below), whose input is necessary for V2 layer 5 cells to close the intracortical resonant loop in V2 that is capable of driving fast self-stabilizing learning in V2; see Section 1.3.2.

V2 layer 6^II can also influence attentive top-down corticocortical feedback to layer 4 of V1 via layer 1 apical dendrites of layer 5 cells that project to layer 6^I and then to 4 via a modulatory on-center, off-surround circuit (Fig. 3b). See Section 2.2 for simulation results.

In summary, the SMART specific pathway is responsible for attentively matching bottom-up and top-down information in the specific thalamus, and creating attentive and synchronous resonant states that can support fast stable learning of bottom-up oriented filters and top-down oriented modulatory expectations. When the specific pathway interacts with the nonspecific pathway, it can also experience reset and memory search for better-matching filters and expectations, as the following section clarifies.

The thalamic nonspecific pathway includes both the “matrix” cells in the specific thalamic nuclei (Fig. 2a; Jones, 2002) and the nonspecific thalamic nuclei (Figs. 2d–f). Both pathways project to the superficial layers of the cerebral cortex. The term nonspecific, as opposed to specific, thalamic nuclei (both first-order and second-order nuclei), refers to the midline thalamic and the intralaminar nuclei. The term nonspecific derives from three characteristics of these nuclei, namely: (1) their diffuse innervation from pontine, medullary and mesencephalic reticular formation; (2) the signature of their stimulation in the cortical mantle (somnolence for low-frequency stimulation, arousal for high-frequency); and (3) the anatomical observation that they project to cerebral cortex in a fairly uniform fashion (van Der Werf et al., 2002). Most of the nonspecific thalamic nuclei are characterized by a high degree of convergent cortical input, widespread projections to large portions of neocortical layer 1, inhibition from the thalamic reticular nucleus (TRN), and strong neuromodulatory input from several brainstem centers (van Der Werf et al., 2002).

Neuropsychological and neurological evidence has demonstrated the importance of the intralaminar and midline nuclei for cortical functioning (Llinas and Pare, 1991, Llinas et al., 2002). Midline lesions of the thalamus affect general cognition, resulting in lethargy or coma (Facon et al., 1958) or unilateral hemineglect (Heilman et al., 1993), despite the fact that the specific sensory stimuli are relayed to the cortex.

Core vs. matrix cells. Recent studies have shed additional light on the dichotomy between specific and nonspecific thalamic nuclei, showing how the distinction between patterns of cortical termination (superficial vs. deep layers) not only characterizes cells between nuclei, but also cells within specific thalamic nuclei. Cytological studies on thalamocortical relay cells in monkeys conducted with the use of immunoreactivity for the calcium binding proteins parvalbumin and calbindin have shown a “core” of parvalbumin-rich cells projecting to the middle layers of their cortical targets, surrounded by a “matrix” of calbindin-rich cells projecting to the superficial layers (Jones, 2002). This matrix extends to all specific thalamic nuclei irrespective of nuclear borders, and differs from the core also by the nature of its input.

Core cells receive subcortical afferents that are highly ordered topographically, and a similarly ordered pattern is maintained at the site of cortical terminations of core cell axons at layer 4 (Jones 2002). Matrix cells receive subcortical input which tends to terminate in multiple thalamic nuclei, show a less precise stimulus–response relationship, have receptive fields that are not easily definable, and project to superficial cortical layers. For instance, in the medial geniculate complex, core cells receive tonotopically-ordered inputs from the central nucleus of the inferior colliculus, representing the most direct ascending pathway from the cochlea. Matrix cells are instead innervated by a less direct auditory pathway which ascends in the midbrain tegmentum and terminates diffusely in most of the nuclei that form part of the medial geniculate complex. Similar patterns of terminations are repeated in somatosensory and visual sections of the thalamus (ventral posterior complex and dorsal lateral geniculate nucleus, respectively). The results of Jones (2002) suggest that a functional microarticulation similar to the one observed in the specific and nonspecific thalamic nuclei may be mirrored by the core/matrix cell dichotomy in the specific thalamic nuclei. Both the matrix cells and nonspecific thalamic cells in the nonspecific pathway terminate on apical dendrites of layer 5 cells, mirroring the anatomical and functional similarities between matrix cells in specific nuclei and nonspecific thalamic cells (Jones, 2002).

Priming vs. reset. As noted above, in the SMART model, the matrix cells in the nonspecific pathway provide a priming input that allows the cortical hierarchy to fully process a bottom-up input (Fig. 2a). Sudden increments in activity within the nonspecific thalamic nucleus are also responsible for generating reset signals and memory search during predictive mismatch episodes. The model cell population in this nucleus is excited by converging bottom-up input (Fig. 2d), and sends excitatory connections to layer 1 of the cerebral cortex (Jones, 2002, Miller and Benevento, 1979), where its collaterals contact apical dendrites of layer 5 pyramidal cells (Vogt, 1991, Cauller, 1995, Cauller and Connors, 1994, Cauller and Connors, 2001, Larkum et al., 2002, Larkum et al., 2004). The nonspecific thalamus is also inhibited by the thalamic reticular nucleus, or TRN (Fig. 2e), and the balance between bottom-up excitation and TRN inhibition is controlled by the matching process. If a mismatch is large enough, the decrease of excitation in the LGN due to the misalignment of bottom-up and top-down input will decrease LGN firing and, thus, TRN inhibition to the nonspecific thalamus, while the excitatory bottom-up input will remain unchanged. Thus the total excitatory input to the TRN from layer 6^II and LGN is, all else being equal, larger in cases of match, when many LGN cells are excited, than in cases of mismatch, when only a smaller subset or no LGN cells are allowed to fire. The net result is an increase in firing rate in the nonspecific thalamus (see Section 2.3) that causes a spatially diffuse arousal burst to layer 1 of the cortex.

Arousal, reset, and search. How does a spatially diffuse arousal burst from the nonspecific thalamic nucleus selectively reset the cortical codes that caused a mismatch? At the moment when a mismatch occurs, the brain does not know which cortical areas caused the predictive failure (Grossberg, 1980). Despite this lack of information in the nonspecific thalamus, the mismatch there needs to be able to selectively reset active representations throughout the cortical hierarchy. SMART proposes, in accord with known anatomical and physiological data both in vivo and in vitro (Larkum et al., 1999, Larkum and Zhu, 2002), that layer 5 pyramidal cell firing rate is jointly controlled by nonspecific thalamic inputs and specific layer 2/3 inputs, thus explaining how layer 5 cells exhibit two distinct firing modes (Williams and Stuart, 1999): Layer 5 cells that receive layer 2/3 inputs and nonspecific thalamic inputs during a mismatch episode fire in bursts at high rates (see Section 2.6 for experimental and simulation results). Active 2/3 cells represent cortical codes that caused the mismatch. In contrast, single spikes are produced in layer 5 cells when only one of these sources is activated, either during a match, or during a mismatch when layer 2/3 cells are inactive.

As noted above, layer 5 pyramidal cells send driving inputs directly to higher cortices through the thalamus (e.g., the pulvinar; Sherman and Guillery, 2002, Shipp, 2003; see Fig. 3a), indirectly control corticothalamic feedback at their own cortical level through layer 6^II (Fig. 2, Fig. 3), and also control corticocortical feedback to layer 4 at their own cortical level via layer 6^I (Fig. 2, Fig. 3). Layer 5 can hereby generate widespread bursts of synchronized activity throughout the neocortex mediated by driving layer 5 terminations on higher-order thalamic nuclei (including pathological epileptogenic activity; Williams and Stuart, 1999), and selectively reset multiple cortical areas by relaying from the nonspecific thalamus layer 5 bursts to layer 4 via the 6^I→4 pathway. In particular, model layer 6^I cells are predicted to respond to a thalamic mismatch with selective cortical reset and search for a more predictive cortical code in layers 4 and 2/3 (see Section 2.6 for simulation results).

The nonspecific pathway may also help to regulate modality-specific attention during reset episodes (Crick, 1984, Guillery et al., 1998, Montero, 1997, Weese et al., 1999). SMART predicts how cortical areas that experience strong predictive mismatches in a given modality may reduce priming of the cortical area of a competing modality by inhibiting the corresponding nonspecific thalamic nucleus. In particular, Crabtree and Isaac (2002) have shown that nonspecific thalamic nuclei which subserve different modalities are linked by mutually inhibitory interactions. SMART simulates how TRN-mediated (van Der Werf et al., 2002) inhibitory interactions (Crabtree and Isaac, 2002) between nonspecific thalamic nuclei can cause a pause in firing of one nonspecific thalamic nucleus that can transiently down-regulate layer 5 pyramidal cells of the competing cortical area (see simulations in Section 2.9). SMART further predicts that competing specific nuclei, not only nonspecific nuclei as shown by Crabtree and Isaac (2002), might be inhibited by the TRN in cases of strong mismatches, therefore being a possible thalamic substrate for competitive allocation of attention.

Learned generalization, vigilance, and acetylcholine. How is the generality of recognition categories regulated to represent statistical properties of the environment? As noted above, ART predicts that resonance and learning occur when the degree of match between bottom-up and top-down representations is greater than a gain parameter, called vigilance (see Fig. 4). Vigilance can change due to internal factors, such as fatigue, or external factors, such as predictive mismatch or punishment. A baseline vigilance determines how big a mismatch is initially tolerated before cortical representations are reset. When a predictive error causes a mismatch to occur, the vigilance level is predicted to increase just enough to drive a memory search for a new recognition code. This process is called match tracking (Carpenter and Grossberg, 1987, Carpenter and Grossberg, 1991, Carpenter et al., 1992). Match tracking realizes a kind of minimax learning rule; namely, it enables a learning system to minimize predictive error while maximizing generalization. Choosing a low baseline vigilance leads to the learning of general categories and thus a minimum use of memory resources. Match tracking increases this baseline vigilance just enough to learn the most general categories that are consistent with predictive success.

The SMART model predicts that one way to control vigilance may be to modify the excitability of layer 5 cells during mismatch episodes (Fig. 5). Anatomical studies in monkeys, cats and rats have established that the nonspecific thalamus (in particular, the midline and central lateral thalamic nuclei), whose activation is sensitive to the degree of mismatch, projects to the cholinergic nucleus basalis of Meynert (van Der Werf et al., 2002), one of the main sources of cholinergic innervations of the cerebral cortex. The nucleus basalis of Meynert is also influenced by noxious stimulation and cortical control (Zhang et al., 2004). Saar et al. (2001) have shown that ACh release reduces the after-hyperpolarization (AHP) current and increases cell excitability in layer 5 cortical cells (see Section 2.8). In SMART, this increased layer 5 excitability due to predictive mismatch may cause reset via the layer 5-to-6^I-to-4 circuit, even in cases where top-down feedback may earlier have partially matched bottom-up input, which is a key property of vigilance control. The increase of ACh might therefore promote search for finer recognition categories in response to environmental feedback, even when bottom-up and top-down signals have a pretty good match in the nonspecific thalamus based on similarity alone.

Fig. 3c summarizes all of the simulated SMART circuitry. Table 2 summarizes the main anatomical features simulated, their functional interpretation, and supportive experimental literature. The Methods section provides a detailed description of the model equations and parameters.