Holly Kular¹, Robert Kim², John Serences¹, Nuttida Rungratsameetaweemana³; ¹University of California San Diego, ²Cedars-Sinai Medical Center, ³Columbia University

Natural visual environments are highly structured and statistical regularities can be used to form expectations that guide behavior. However it is not well understood how top-down information such as prior expectation modulates information processing along the cortical hierarchy. Here, we test the hypothesis that expectations can shape processing at all levels of information processing, particularly when stimulus uncertainty is high. We isolated effects of expectation using a hierarchical continuous-time recurrent neural network (RNN) trained on a perceptual decision task while simultaneously manipulating stimulus uncertainty and set size. To establish top-down expectations in the network, the RNN was trained with the expected stimulus presented more frequently than others. We show that the RNN extracts sensory statistics and develops expectations that improve processing when stimuli are uncertain. Linear SVM classifiers were trained on RNN firing rates to decode stimulus identity, comparing conditions of low/high uncertainty and small/large stimulus sets. Decoding accuracy was highest for expected stimuli across all conditions. Additionally, the expected stimulus could be more readily decoded from activity in the final layer under the high uncertainty condition, suggesting a strong top-down influence from higher layers. This effect is more pronounced when the stimulus set size is large, indicating a greater role for top-down modulation from the final layer as the amount of incoming sensory information decreases. These findings show that the RNN developed expectations and relied on top-down signals under uncertainty and with larger stimulus spaces, supporting the hypothesis that expectations facilitate visual processing particularly when bottom-up sensory information is limited.