Speaker
Description
While a great deal is known about how neurons of the early visual system respond to simple stimuli, our understanding of how they process natural stimuli is still limited. Machine learning models have been invaluable to predict how these neurons respond to natural stimuli. However, the increasing complexity of these models make them difficult to interpret, and their ability to generalize to new stimuli is limited.
I will describe recent works from my lab where we address these issues at the level of the retina. We developed a new perturbative approach to probe the selectivity of individual neurons during natural scenes, and to understand the features they extract. We use this method to characterize and model how short-term adaptation impact how the retina processes natural scenes. Our results show that adaptive mechanisms are not just here to normalize the response to the average luminance and contrast: during natural scene stimulation it also reshapes the selectivity of specific types of ganglion cells. Finally, we show that the ability of artificial neural networks to generalize can be dramatically enhanced in specific ganglion cell types by incorporating in the model a new geometry constraint derived from a natural hunting behavior.
Together our results suggest that scaling up model size may not be enough: including knowledge about the biological basis of visual processing in artificial neural network models might be necessary to understand sensory processing.
