STATE-OF-THE-ART DNN MODELS OF PRIMARY VISUAL CORTEX ONLY PARTIALLY PREDICT RESULTS FROM CLASSICAL V1 EXPERIMENTS

Deep neural networks trained to predict neural responses are increasingly used as in silico models of visual cortex, primarily assessed by their predictive accuracy on natural-image stimuli. In parallel, classical neurophysiological experiments have established fundamental V1 response properties through controlled stimulus manipulations. An important unaddressed question is whether the data-driven models trained on natural images also express the elementary response characteristics revealed by these classical experiments. We use a standardized in silico framework to replicate classical visual neuroscience experiments on a state-of-the-art model of macaque V1 under matched stimulus conditions. We reimplemented a broad set of classical V1 experiments spanning center–surround, contrast, and orientation manipulations, as well as polarity, texture, and second-order structure. The classical stimuli were adapted to match the model’s visual field, resolution, and training image statistics. The model reproduced several qualitative trends reported in vivo, including surround suppression and its modulation by contrast and orientation, a bias toward black stimuli, and weak but consistent signatures of second-order sensitivity. However, quantitative agreement was limited: model response statistics deviated from in vivo measurements in a systematic way, with effect magnitudes and response variability often reduced by several fold in model response. These results demonstrate that high predictive accuracy on natural images does not ensure preservation of the fundamental response structure revealed by classical physiological experiments. Making these discrepancies explicit and comparable clarifies which aspects of cortical response structure are not constrained by natural-image predictivity alone, providing a practical basis for developing better-constrained models of visual cortex.