Integrated Calendar

An increasing amount of evidence suggests that memory formation and retrieval are modulated by contextual signals, such as behavioral or emotional state. However, typical models of associative memory do not incorporate this dependency. Here we propose an extension to the Hopfield network which takes into account contextual modulation. The network is divided into a set of overlapping subnetworks, each representing a different context with a separate set of memory patterns. Only one subnetwork is active at any given time, thereby reducing interference from memories found in other contexts, which remain dormant through inhibitory control. Using theoretical and numerical methods, we show that these context-modular Hopfield networks have substantially increased memory capacity, as well as robustness to noise and to memory overloading. Their performance depends on two parameters—the number of subnetworks, and their relative size—and when chosen optimally, the capacity is up to ten times greater than the standard Hopfield model. We then show that adding context-dependent dendritic pruning further enhances the performance of the model. Improved performance comes at the cost of limited retrieval, because only memories stored in the active subnetwork can be recalled. To address this, we propose a system in which a controller network dynamically switches the memory network to a desired contextual state before storage or retrieval. Through simulations, we successfully show that this system is able to bias memory retrieval based on context. Overall, our work illustrates the benefits of context-dependent memory, and may have implications for our understanding of cortical memories and their interaction with contextual signals in the prefrontal cortex and hippocampus.