Yixin Yuan¹, Mikio Aoi¹, John Serences¹; ¹UC San Diego, ²UC San Diego, ³UC San Diego

Recent studies with longitudinal neural recordings in mice suggest that neural activity associated with repeated stimuli gradually deviates from its initial pattern over time, even when task performance stays constant. This phenomenon, termed ‘representational drift’, challenges the common assumption that after fully learning a task, neural responses stabilize to support robust representations of task-relevant stimuli. Here we sought to determine whether representational drift occurs in human subjects while they are engaged in a visual working memory task, and to determine if drift is context specific. We recruited 4 participants (3F, 1M) to undergo a 5-session long fMRI study spread across a span of 3-5 weeks. During each session, the participants performed a two-alternative forced choice task that required recalling whether a probe object matched the target object shown prior to a 3-10 second delay period. Importantly, we manipulated the relevant stimulus components so that there were two task contexts: on half of the trials, the participants were instructed to remember the location and color of the object, while on the other half they were instructed to remember the location and shape. The participants were behaviorally trained and the task difficulty level was staircased so that accuracy stabilized at ~75% prior to the first scan. We then ran ridge-regularized circular regression on activation patterns in functionally determined regions-of-interest to predict the spatial location of the target object across trials. On average, within-session decodability of the target object’s spatial location is higher than that of cross-session decoding. These results suggest that representations of spatial locations encoded in working memory drift from session-to-session, and this effect can be readily detected in the human cortex via fMRI.