Abstract

A long-standing hypothesis in computational neuroscience is that the neocortex learns a hierarchical model of the world by predicting incoming sensory data, and learning from unexpected events. This hypothesis predicts that: (1) There should be distinct responses to expected and unexpected stimuli. (2) As a circuit learns about stimuli, the responses to both expected and unexpected stimuli should change. (3) There should be differences between the manner in which top-down and bottom-up driven responses change during learning. (4) The response changes should be long-lasting and predictable from the differences in the responses to expected and unexpected stimuli. Here, we use chronic two-photon imaging of supra- and sub-granular pyramidal neurons in mouse visual cortex to test these hypotheses. We habituated the mice to sequences of random Gabor frames with embedded patterns to help shape expectations. We then conducted imaging over three days during which time we exposed the mice to sequences that violated the previous patterns. We find that all four of the above predictions are borne out in the data: pyramidal neurons respond differently to unexpected stimuli, the responses to both expected and unexpected stimuli evolve over time, the evolution of these responses is different in the distal apical dendrites and the somata, and the differences in responses predicted this evolution over days. Altogether, this data supports the hierarchical predictive model hypothesis, and provides greater information about how this learning is implemented in the neocortical microcircuit.