Learning to combine sensory evidence and contextual priors under ambiguity

The neocortex is composed of a hierarchy of regions and is thought to perform sensory inference by combining two complementary information streams: 1) a feedforward stream propagating bottom-up in the hierarchy, representing sensory information, and 2) a feedback stream propagating top-down, representing expectations or priors derived from contextual information integrated in higher-order associative regions. Within a cortical region, pyramidal neurons in layers 2 & 3 are a key cellular component of the feedforward pathway which integrate these two streams at their basal and apical dendrites, respectively. How they combine these two streams to update feedforward representations based on contextual priors is unknown. Here we propose a functional model of integration of these two streams at basal and apical compartments of pyramidal cells based on known physiological principles. We developed an ambiguous MNIST dataset that implements parameterized ambiguity between digits by conditional generation, and trained the feedback projection onto apical dendrites in our model by gradient descent to complement the sensory representation at the basal dendrites and resolve ambiguity. Specifically, when input stimuli are ambiguous, contextual priors arriving at the apical dendrites are integrated in the sensory representation to rescue classification performance. Importantly, when stimuli are unambiguous, contextual priors which oppose sensory evidence are appropriately ignored. Our proposed model allows analysis of candidate local learning rules that could support learning of such models, and provides insight into how pyramidal neurons and neocortical circuitry could integrate sensory and contextual information to learn predictive models of the world.