Abstract
Our visual system continually adapts to its sensory environment. As a result, both sensory encoding and perceptual behavior change according to the recent history of sensory input. Earlier psychophysical experiments have shown that adaptation to a single stimulus orientation substantially increases encoding accuracy at the adaptor orientation (Mao/Stocker, VSS 2020). Here we tested the hypothesis that adaptation optimally prepares the perceptual system for future stimuli by establishing efficient sensory representations for the next expected sensory input. First, we asked whether natural input statistics predict the observed increase in accuracy at the adaptor orientation. Specifically, we analyzed the spatiotemporal orientation distributions in the retinal inputs of freely behaving human subjects in natural environments. We found that after relative stable visual input, the distribution of local orientations in the next frame peaked at the mean orientation over that short stable stimulus history. Thus, an increase in encoding accuracy at the adaptor is consistent with an efficient representation of the next stimulus under natural stimulus conditions. We further tested the hypothesis using PredNet, a recurrent neural network trained to predict the next frame in naturalistic videos. We computed the representational accuracy of image orientation in the network after presenting it with short image sequences containing just a single orientation (adaptor). PredNet exhibited the same increase in encoding accuracy at the adaptor orientation as observed in human subjects. Because PredNet contains no local adaptation mechanisms, the increase in encoding accuracy is solely a function of the stimulus history and the task of the network to best possible predict the next video frame. Together, our results suggest that adaptation induced changes in encoding accuracy and perceptual behavior reflect the visual system’s goal to be best possibly prepared for future sensory input.