Contrast-dependent response modulation in convolutional neural networks captures behavioral and neural signatures of visual adaptation

Human perception remains robust under challenging viewing conditions. Robust perception is thought to be facilitated by nonlinear response properties, including temporal adaptation (reduced responses to re-peated stimuli) and contrast gain (shift in the contrast response function with pre-exposure to a stimulus). Temporal adaptation and contrast gain have both been shown to aid object recognition, however, their joint effect on perceptual and neural responses remains unclear. Here, we collected behavioural measurements and electrocorticography (EEG) data while human participants (both sexes) classified objects embedded within temporally repeated noise patterns, whereby object contrast was varied. Our findings reveal an in-teraction effect, with increased categorization performance as a result of temporal adaptation for higher but not lower contrast stimuli. This increase in behavioral performance after adaptation is associated with more pronounced contrast-dependent modulation of evoked neural responses, as well as better decoding of object information from EEG activity. To elucidate the neural computations underlying these effects, we endowed deep convolutional neural networks (DCNN) with various temporal adaptation mechanisms, including intrinsic suppression and temporal divisive normalisation. We demonstrate that incorporating a biologically-inspired contrast response function to modify temporal adaptation helps DCNNs to accurately capture human behaviour and neural activation profiles. Moreover, we find that networks with multiplicative temporal adaptation mechanisms, such as divisive normalization, show higher robustness against spatial shifts in the inputs compared to DCNNs employing additive mechanisms. Overall, we reveal how interaction effects between nonlinear response properties influence human perception in challenging viewing contexts and investigate potential computations that mediate these effects.