Overcoming sensory-memory interference in working memory circuits

Memories of recent stimuli are crucial for guiding behavior, but the sensory pathways responsible for encoding these memories are continuously bombarded by new sensory experiences. How the brain overcomes interference between sensory input and working memory representations remains largely unknown. To formalize the solution space, we examined recurrent neural networks that were either hand-designed or trained using gradient descent methods, and compared these models with neural data from two different macaque experiments. Here we report mechanisms by which neural networks overcome sensory-memory interference using both static and dynamic coding strategies: gating of the sensory inputs, modulating synapse strengths to achieve a strong attractor solution, and dynamic strategies – including the extreme solution in which cells invert their feature preference during working memory. Neural data from the medial superior temporal (MST) area of macaques, where sensory and working memory signals first interact along the dorsal pathway, best aligned with a solution that combined input gating and tuning inversion. Behavioral predictions from this model also matched error patterns observed in monkeys performing a working memory task with distractors. Taken together, our results help elucidate how working memory circuits preserve information as we continue to interact with the world, and suggest intermediate cortical visual areas like MST may play a critical role in this computation.