Biological key-value memory networks

In neuroscience, Hopfield networks are the classical biologically plausible model of long-term memory, relying on Hebbian plasticity for storage and attractor dynamics for recall. In contrast, memory-augmented neural networks in machine learning commonly use a key-value mechanism to store and read out memories in a single step. Such networks can achieve impressive feats compared to traditional variants, yet their biological relevance is unclear. Here, we propose a biological implementation of basic key-value memory that stores inputs using a combination of Hebbian and non-Hebbian plasticity rules. Similar plasticity rules are recovered when network parameters are meta-learned. Our network performs similarly to Hopfield networks on autoassociative memory tasks and can be naturally extended to continual recall, heteroassociative memory, and sequence learning. Our results suggest a compelling alternative mechanism to Hopfield networks for biological long-term memory.