Representational Drift in Human Visual Cortex

Eli Merriam, Ph.D.

August 11(Thu) - August 11(Thu), 2022

4PM

N Centre 86314 & ZOOM (ID: 728-142-6028)

CNIR Seminar

Title: Representational Drift in Human Visual Cortex

Abstract: Neural representations in primary sensory cortical regions are often thought to encode stimulus information in a stable manner across time. But recent rodent studies suggest this is not the case. Instead, visual representations in mouse V1 gradually change, or drift, over time, at timescales ranging from minutes to weeks (Deitch et al., 2021; Marks and Goard, 2021). The origin and function for this drift is not known. Mice do not depend heavily on visual cues, so perhaps representational drift in mouse V1 simply reflects the allocation of resources to other more ecologically relevant sensory modalities. Alternatively, perhaps representational drift is a fundamental property of neural representation, present across brain systems and species. To address these questions, we asked whether representational drift occurs in human visual cortex. We analyzed fMRI BOLD activity in the Natural Scenes Dataset, collected while subjects viewing a massive database of naturalistic scenes in many scanning sessions over many months (Allen et al., 2021). We used single fMRI sessions to fit image-computable models to individual voxel responses. We then measured model generalizability across sessions, training the model on each session and testing it on all other sessions. We found that generalizability decreased with the amount of time between training and testing, implying that the representation, as characterized by the model, had changed. Critically, the relative similarity between representations of multiple stimuli were unchanged across sessions, suggesting that downstream cortical areas can read-out a stable representation, even as the representation within V1 drifts. The drift that we observed was not due to trivial changes in the fMRI signal, such as changes in SNR or head motion. Instead, the drift arose from gradual and systematic changes in the mean activity of individual voxels, with some voxels exhibiting increases and others exhibiting decreases over sessions. After mean activity changes were removed from the data, the drift disappeared. Our results suggest that drift is a general feature of neural representations, present in both mouse and human visual cortex. While the function of representational drift remains elusive, our results demonstrate that cortical networks are dynamic, even when their putative function involves simple feedforward computation.