Artificial neural networks (ANNs) have limited adaptive capabilities, hindering their ability to reliably predict neural output under dynamic input conditions. This research develops a new deep learning model of the retina incorporating the biophysics of photoreceptor adaptation at the front-end of conventional convolutional neural networks (CNNs). CNNs including this photoreceptor layer outperform conventional CNN models at predicting primate and rat retinal ganglion cell (RGC) responses to naturalistic stimuli with dynamic local intensity changes and large ambient illumination changes. This underscores the potential of embedding neural adaptation models in ANNs to better encode dynamic, wide-range natural inputs.

Saad Idrees, Michael B. Manookin, Fred Rieke, Greg D. Field, Joel Zylberberg