Engines of the Brain: The Computational "Instruction Set" of Perception and Cognition

Cognition is the action and interaction of multiple brain regions, and these are becoming understood computationally. Simulation and analysis has led to derivation of a set of elemental operations that emerge from individual and combined brain circuits, such that each circuit contributes a particular algorithm, and pairs and larger groups interact to compose further algorithms. We forward the hypothesis that these operations constitute the “instruction set” of the brain, i.e., that these are the most basic mental operations, from which all other behavioral and cognitive operations are constructed, constructing a unified formalism for description of operations ranging from perceptual to cognitive, including vision, language, learning and reasoning. Telencephalic Organization Figure 1 depicts the primary elements of the mammalian forebrain (telencephalon), shared across all mammalian species. Whereas posterior neocortex receives sensory inputs (via dorsal thalamus), anterior neocortex produces motor outputs and, in so doing, interacts closely with the basal ganglia, a more ancient structure that dominates reptilian brains. Both anterior and posterior cortex interact with hippocampal and amygdaloid structures. Mammalian brains scale across several orders of magnitude (e.g., from milligrams to kilograms), yet overwhelmingly retain their structural design characteristics. As the ratio of brain size to body size grows, particular allometric changes occur, defining differences between bigger and smaller brain designs. Figure 1b illustrates the three largest changes: 1) Connection pathways between anterior and posterior cortex (fasciculi) grow large 2) Output pathways from striatal complex change relative size: the recurrent pathway back to cortex via thalamus increases relative to the descending motor pathway 3) Descending output from anterior cortex to motor systems (pyramidal tract) grows large These changes grow disproportionately with increased brain-body ratio, becoming notably outsized in humans. In relatively small-brained mammals such as mice, the primary motor area of neocortex is an adjunct to the striatally driven motor system. Whereas damage to motor cortex in mice causes subtle behavioral motor impairments, damage to motor cortex in humans causes paralysis. In this example of encephalization of function (Jackson, 1925; Ferrier, 1876; Karten, 1991, Aboitiz, 1993) motor operations are increasingly ‘taken over’ by cortex as the size of the pyramidal tract overtakes that of the descending striatal system. The role of the striatal complex in mammals with large brain-body ratios is presumably altered to reflect that its primary inputs and outputs are now anterior neocortex; in other words, it is now primarily a tool or “subroutine” available for query by anterior cortex. Its operations then are most profitably analyzed in light of its dual utility as organizer of complex motor sequences (in small brained mammals) and as informant to anterior cortex (in large brained mammals). Figure 1: Telencephalic organization Basal Gangl ia / Striatal Complex The basal ganglia, the primary brain system in reptiles, is a collection of disparate but interacting structures. Figure 2 schematically illustrates the primary components included in the modeling efforts described herein. Distinct components of the basal ganglia exhibit different, apparently specialized designs: matrisomes (matrix), striosomes (patch; which exist as small ‘islands’ embedded throughout the surrounding matrix regions), globus pallidus, pars interna and externa (pallidum), subthalamic nucleus (STN), tonically active cholinergic neurons (TANs), and substantia nigra pars compacta (SNc). These are connected via a set of varied neurotransmitter pathways including GABA, glutamate (Glu), dopamine (DA), acetylcholine (ACh), and Substance P (Sp) among others, each of which may affect multiple receptor subtypes. Figure 2. Basal ganglia / striatal complex The two pathways from cortex through the matrix components of the striatal complex involve different subpopulations of cells in matrisomes (matrix): i) MSN1 neurons project to globus pallidus pars interna (GPi), which in turn project to ventral thalamus and back to cortex; ii) MSN2 neurons project to globus pallidus pars externa (GPe), which in turn projects to GPi (and thence to thalamus and cortex). MSN and GP projections are GABAergic, inhibiting their targets. Thus cortical glutamatergic activation of MSN1 cells causes inhibition of GPi cells, which otherwise inhibit thalamic and brainstem targets; hence MSN1 cell activation disinhibits, or enhances, cortical and brainstem activity. In contrast, an extra GABAergic link is intercalated in the pathway from MSN2 neurons to the output stages of matrix; thus activation of MSN2 neurons decreases activation of cortex and of brainstem nuclei. The pathways from MSN1 and MSN2 neurons are thus termed “go” and “stop,” respectively, for their opposing effects on their ultimate motor and cortical targets. Coördinated operation over time of these pathways can yield a complex combination of activated (go) and withheld (stop) motor responses (e.g., to stand, walk, throw), or correspondingly complex “thought” (cortical) responses. Two primary afferents to striosomes are cortical and ascending inputs. The former are the same as the inputs to matrix (despite the schematized depiction in the figure, patch components are distributed through, and colocalized with, matrix). Ascending inputs to patch denote “reward” and “punishment” information and have been shown to upand down-regulate dopamine from SNc (as well as other dopaminergic sites) in response to external stimuli carrying innate or learned valences (e.g., water to a thirsty organism). A cortically triggered action, followed by an ascending DA reward signal from SNc to patch, selectively enhances active cortical glutamatergic synapses on both matrix and patch targets. Patch output back to SNc then inhibits DA response, so that increased cortical activation of patch (via enhanced synaptic contacts) will come to limit the DA input from SNc. On any given trial, then, the size of the DA signal from SNc reflects the size of the actual ascending DA input (i.e., the reward signal) that occurred over previous trials. Thus with repeated experience, adaptive changes occur in both matrix and patch: initially-random matrix responses to a cortical input become increasingly selected for responses that produce reward, and initial naïve striosomal responses will become increasingly good “predictors” of the size of the reward expected to ensue as a result of the action. Tonically active cholinergic neurons (TANs) represent a small fraction (< 5%) of the number of cells in the striatal complex yet densely contact cells throughout matrix; thus they likely play a modulatory role rather than conveying specific information. The GABAergic inhibition of these cells by patch will come to increase for those patch responses that lead to reward, since in these instances the cortical drivers of these patch responses become synaptically enhanced. Thus in those circumstances where cortical inputs lead to expected reward, TANs will tend to have less excitatory effect on matrix. Since the TAN afferents to matrix are dense and nontopographic, they represent a random “background noise” input, which can increase variance in selected matrix responses to cortical inputs, making the striatally-selected motor response to a cortical input somewhat nondeterministic. The resulting behavior should appear “exploratory,” involving a range of different responses to a given stimulus. Synaptic increases in patch, in addition to causing accurate “predictions” of reward size, as described, also increasingly inhibit TANs, diminishing the breadth of exploratory variability. Thus as rewards occur, not only will reward-associated responses be increasingly selected by matrix, but the variability among those responses will decrease. Analyses suggest detailed comparisons of basal ganglia and standard reinforcement learning systems (Schultz et al., 1997; Dayan et al., 2000; see Table 1). Table 1. Sample simplified basal ganglia algorithm 1) Choose action A. Set reward_estimate ← 0 Set max_randomness ← R > 0 2) randomness ← max_randomness – reward_estimate 3) reward ← Eval( A + randomness ) 4) If reward > reward_estimate then A ← A + randomness reward_estimate ← reward