2022 Conference Objects

Biological learning in key-value memory networks

Tyulmankov, Danil; Fang, Ching; Dong, Ling Liang; Vadaparty, Annapurna; Yang, Guangyu Robert

In neuroscience, classical Hopfield networks are the standard 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 augmented networks achieve impressive feats of memory compared to traditional variants, yet it remains unclear whether they can be implemented by biological systems. In our work, we bridge this gap by proposing a set of of biologically plausible three-factor plasticity rules for a basic feedforward key-value memory network. Keys are stored in the input-to-hidden synaptic weights by a "non-Hebbian" rule, controlled only by pre-synaptic activity, and modulated by local third factors which represent dendtritic spikes. Values are stored in the hidden-to-output weights by a Hebbian rule, with the pre-synaptic neuron selected through softmax attention which represents recurrent inhibition. The same rules are recovered when network parameters are meta-learned. Our network performs on par with classical Hopfield networks on autoassociative memory tasks and can be naturally extended to correlated inputs, continual recall, heteroassociative memory, and sequence learning. Importantly, since memories are stored in slots indexed by hidden layer neurons, unlike the fully distributed representation in the classical Hopfield network, they can be individually selected for extended storage or rapid decay. Finally, our memory network can easily be incorporated into a larger neural system, either as a memory bank for an external controller, or as a fast learning system used in conjunction with a slow one. Overall, our results suggest a compelling alternative to the classical Hopfield network as a model of biological long-term memory.

Keywords: learning, memory, synaptic plasticity, Hebbian, key-value memory, neural network, three-factor plasticity