Zvi N. Roth¹, Elisha P. Merriam²; ¹Gonda Brain Research Center, Bar-Ilan University, ²Laboratory of Brain and Cognition, National Institute of Mental Health

Several recent studies have reported that stimulus representations are not temporally stable, but rather exhibit cumulative, directional changes over hours, days, and months, a phenomenon termed 'representational drift'. Many of these studies used a stimulation protocol in which stimuli were repeated multiple times within each recording session. Repetition suppression is a well-established phenomenon whereby repeated presentations of a stimulus evoke decreasing neural responses. Here, we ask whether repetition suppression affects measures of representational drift, or alternatively, whether the two are entirely distinct phenomena. Representational drift has recently been demonstrated in human V1 using the Natural Scenes Dataset (NSD), which consists of fMRI responses to a large number of unique images repeated up to 3 times over the course of a year. Computer simulations revealed that repetition suppression could, in principle, produce changes similar to those we have attributed to representational drift. We next tested whether repetition suppression does indeed account for representational drift. First, we reanalyzed the NSD, limiting ourselves to first-repetition trials only, while subsampling each session to equalize the number of trials across sessions. While this subsampling procedure drastically reduced the amount of data used in the analysis, we nonetheless found robust evidence for representational drift. Second, we analyzed the THINGS dataset, which contains a large set of images that were not repeated (we didn't analyze the subset of 100 stimuli that were repeated). Remarkably, two out of three subjects exhibited significant drift, even though the THINGS dataset is considerably smaller than NSD. These findings indicate that although repetition suppression poses a major potential confound in representational drift studies, the two phenomena are distinct. Our results suggest that neural representations are subject to multiple time-dependent changes at different time scales. The role of each of these processes in cognition, perception, and learning remains an important and unresolved question.