Neural Representation of Temporal Prediction in Mouse Primary Visual Cortex

Abstract

Perception requires the nervous system to infer latent causes from inherently ambiguous sensory measurements that do not uniquely specify the states of the world that produced them. Predictive coding frameworks propose that the brain addresses this ambiguity using an internal generative model that captures learned statistical regularities linking sensory input to latent causes and latent causes to one another. Within this model, each latent variable is represented by combining current sensory input with predictions generated by applying these regularities to estimates of other latent variables. Such predictions may arise from earlier estimates at the same representational level or concurrent estimates at higher levels operating over broader temporal scales. This dissertation investigates how the current stimulus, recent history, and broader sequence structure shape representations in mouse primary visual cortex and how their contributions combine.The first study examined how earlier activity at the same representational level shaped responses to subsequent input. Patterned optogenetic stimulation imposed controlled sequences of population activity in mouse V1 while holding the current input pattern fixed. Responses depended on the preceding patterns and were enhanced for sequences that matched temporal transitions characteristic of natural vision relative to sequences that did not. This sequence dependence extended beyond directly stimulated neurons and included excitation and suppression within the surrounding population, supporting a contribution from recurrent circuit dynamics rather than an effect confined to neurons driven by the preceding input. The network therefore filtered input sequences according to the temporal statistics of natural vision, consistent with a prior over how represented variables evolve over time. The second study examined how predictions derived from broader temporal context shaped V1 representations. In hierarchical predictive coding accounts, higher regions may generate these predictions by integrating information over longer temporal spans and convey them to V1. Mice were exposed to structured visual sequences in which the next element depended on more than the immediately preceding stimulus, while a Neuropixels probe spanning all V1 layers recorded population activity. This allowed identical sensory events to be compared across contexts implying different predicted states. Population representations differed according to the prediction implied by the preceding context, even when the current stimulus and immediate history were held constant. Information about contextual prediction was weak and heterogeneous across individual neurons and distributed across cortical layers rather than confined to a specialized population. Analyses of population geometry indicated that the current stimulus, immediate history, temporal order, and contextual prediction contributed separable components to population structure and combined approximately additively. Together, these studies show that temporal statistics at two timescales reshape representations in V1, the earliest cortical stage of visual processing. At the shorter timescale, predictions from recent activity were expressed through recurrent dynamics within V1, whereas at the longer timescale, predictions were derived from sequence structure extending beyond the interval over which V1 retained information specific to individual stimuli. The approximately additive contributions observed in the second study are consistent with linear-Gaussian formulations of predictive coding, although the findings do not establish a specific anatomical source or circuit implementation for broader contextual predictions.