Representational Drift Across Short Timescales in the Mouse Visual Cortex

Recent studies on representational drift in the visual cortex have produced an intriguing puzzle, demonstrating that neuronal responses to the same external stimuli change over time while population-level activity remains stable (Deitch et al., 2021; Marks & Gourd, 2021). The purpose of drift, and how the brain either compensates for or takes advantage of this phenomenon to maintain robust perception, remains largely unknown. A recent study found that the geometric properties of drift in the primary visual cortex resemble that of continual learning in the presence of a large amount of dropout regularization in artificial neural networks (Aitken et al., 2021), adding to a growing body of work theorizing that drift may be computationally advantageous (Kappel et al., 2018; Qin et al., 2021). However, these studies focused on representational changes over long timescales (days to weeks). In this study, we focus on short timescales to ask: Are cumulative changes observable over the course of hours, and are they consistent with continual learning? Openly available datasets presented limitations for analyzing short timescales due to the large battery of stimuli used, which limited the number of repeats for any individual stimulus. Therefore, we recorded neural activity with Neuropixels probes in six areas of the mouse visual cortex during 200 repetitions of a natural movie. In the cortical visual areas, we found short timescale drift demonstrated unique geometric properties from drift across longer timescales, with differences in how the magnitude of drift and its alignment with the neural state space varied over time. We conducted a follow-up analysis of features in artificial neural networks to probe these properties. This study sheds light on the potential role of drift in enabling beneficial malleability of the neural code needed for continual learning from external stimuli.