Learning orthogonal working memory representations protects from interference in a dual task

Working memory (WM) is a cognitive function that allows for the short-term maintenance and manipulation of information when no longer accessible to the senses. It relies on temporarily storing stimulus features in the neuronal activity. However, the mechanisms protecting WM from task-irrelevant influences are unknown. Recent studies involving WM tasks have suggested that neural activity before and after distractors can be decomposed into two orthogonal subspaces, one invariant across distraction, thus protecting WM from interferences. However, whether orthogonalization is a general mechanism for WM preservation remains an open question, and the network mechanisms supporting it are unclear. Here, we investigated WM representations in calcium imaging data from the prelimbic cortex (PrL) in mice learning to perform a recently developed olfactory dual task. The task consists of an outer delayed paired-association task (DPA) combined with an inner Go-NoGo task. We studied how PrL reflected the process of learning to protect the memory representation of DPA sample odors against Go/NoGo distractor odors. We evaluated the activity overlap with the low-dimensional encoding manifold of the sample/distractor odors as the mouse learned. In the first training days, we found that PrL early delay activity overlapped with both sample and distractor axes, but the overlap with the distractor axis vanished in the late training phase. We give a mechanistic account of these PrL activity overlaps in a spiking network model of strongly recurrent neurons with low-rank connectivity and short-term facilitation. Our model associates learning with (1) orthogonalization of sample and distractor representations and (2) orthogonalization of WM representations with corresponding irrelevant sensory representations. We confirmed both predictions with the help of photoinhibition of the Anterior Cingulate Cortex (ACC) to PrL inputs. Altogether, our results suggest that rotations of WM representations in PrL play a fundamental role in preserving WM from interfering tasks.