Randy O’Reilly, University of Colorado-Boulder

Early developmental learning in babies appears largely passive, and yet forms the deep foundation of all that follows. Standard biologically-supported forms of self-organizing learning, e.g., Hebbian learning, which capture this largely passive and yet “magical” self-organizing aspect, are not computationally powerful enough to achieve many aspects of core cognitive function, including invariant object recognition. Instead, we propose a biologically-based form of error-driven predictive learning, which learns every 100 msec (10 Hz, i.e., the alpha frequency) from the difference between a prediction about what will be seen next, and what is actually seen. The deep layers of the neocortex drive predictions on the pulvinar nucleus of the thalamus, which is broadly interconnected with higher-order visual areas throughout the posterior neocortex, and serves as a kind of neural “projection screen” or “blackboard” (Mumford, 1991). A peculiar, strong, largely one-to-one projection from intrinsic bursting layer 5 neurons (5IB) in V1 and other lower-level areas provides the “ground truth” signal (bursting every 100msec). Where this signal differs from the preceding prediction, which is driven by much more numerous, weaker, and plastic layer 6 corticothalamic (CT) projections, there is a temporal difference error signal that drives learning throughout neocortex, via local synaptic mechanisms attuned to such temporal differences. This model is consistent with a wide range of detailed biological data, and we show that it can self-organize invariant, categorial object representations in its simulated inferotemporal (IT) cortex, based strictly on “passive” viewing of movies of objects moving through space, along with saccadic eye movements. Current work is focused on applying this framework to forward and inverse models in motor learning, and to learning a “predictive model of the self” that could support metacognitive awareness and enable full volitional control to emerge.