Abstract
Background: Perception can be influenced by the temporal context from both past and future stimuli. Recent work has shown that the perception of a target stimulus is affected by the contrast of not only earlier but also later stimuli, demonstrating bidirectional temporal suppression over hundreds of milliseconds. Temporal normalization is the idea that neural responses are divisively suppressed by their past activity, and has been used to explain adaptation. However, models incorporating temporal normalization have mostly focused on predicting neural time courses, rather than perception, and cannot capture the perceptual suppression of earlier stimuli by later ones. Goal: Here we tested whether temporal receptive fields could be used as a mechanism of temporal normalization, and whether this mechanism could capture bidirectional temporal suppression. Methods: We incorporated temporal receptive fields into the Normalization Model of Dynamic Attention (Denison, Carrasco, & Heeger, 2021), by allowing for excitatory and suppressive drives in a population of modeled sensory neurons to be affected not only by the current stimulus input, but also by recent stimulus history. We simulated the model’s response to a sequence of two target orientations that independently varied in contrast. Across simulations, we manipulated the durations of both the excitatory and suppressive temporal receptive fields via their exponential time constants. Results: A model without extended temporal receptive fields failed to capture contrast-dependent suppression. Adding suppressive temporal receptive fields to the model yielded forward suppression, such that a higher contrast first stimulus reduced responses to the second stimulus. Lastly, adding both excitatory and suppressive receptive fields yielded bidirectional suppression, with higher contrasts of each stimulus suppressing model responses to the other. Conclusion: Integrating temporal receptive fields into a dynamic normalization model captured contrast-dependent suppression both forwards and backwards in time, furthering the goal of developing real-time process models of dynamic perception.