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

Studies of human decision-making demonstrate that environmental regularities, such as the nonuniform image statistics of natural scenes, can be exploited to improve efficiency (termed `efficient-coding'). Consistent with efficient coding, previous work has shown that the distribution of tuning preferences in early sensory areas changes during development to track the distribution of features in the environment. In addition to these hard-coded changes, top-down feedback alters selectivity in early sensory areas to bias processing in favor of more likely stimuli when expectations change rapidly with task context. However, top-down feedback might also evolve over development to complement hard-coded changes in sensory areas to help balance efficient coding with the need to sometimes process infrequently encountered stimuli. To better understand the role of top-down feedback in exploiting naturally occurring statistical regularities, we trained a 3-layer hierarchical continuous-time recurrent neural network (ctRNN) to reproduce one of six possible inputs under biased conditions (stimulus 1 more probable than stimuli 2-6). Across all hidden layers, more information was encoded about high-probability stimuli, consistent with the efficient-coding framework. Importantly, reducing feedback from the final hidden layer of trained models selectively magnified representations of high-probability stimuli, at the expense of low-probability stimuli, across all layers. Thus, in addition to enhancing the selectivity of responses to expected stimuli, these results suggest that feedback pathways evolve to protect the processing of low-probability stimuli by regulating the impact of biased input statistics.

Acknowledgements: This work was funded by the National Eye Institute award (RO1-EY025872).