Meta-learning Hebbian plasticity for continual familiarity detection

Memories are stored and recalled throughout the lifetime of an animal, but many models of memory, including previous models of familiarity detection, do not operate in a continuous manner. We consider a family of models that recognize previously experienced stimuli and, importantly, operate and learn continuously. Specifically, we investigate a learning paradigm in which stimuli are presented in a streaming fashion with repetitions at various intervals, and the subject/model must report whether the current stimulus has previously appeared in the stream. We propose a feedforward network architecture with ongoing plasticity in the synaptic weight matrix. Parameters governing plasticity and static network parameters are meta-learned using gradient descent to optimize the continual familiarity detection process. This architecture, unlike recurrent networks without ongoing plasticity, generalizes easily over a range of repeat intervals even if trained with a single interval. We show that an anti-Hebbian plasticity rule (co-activated neurons cause synaptic depression) enables repeat detection over much longer intervals than a Hebbian one, and this is the solution most readily found by meta-learning. This rule leads to experimentally observed features such as repeat suppression in the hidden layer neurons. In contrast to previous theoretical work, the capacity of these networks remains constant across their lifetimes, meaning that pairs of stimuli with a given temporal separation are stored and recognized as familiar independent of the network's input history. We also consider learning rules that use an external gating circuit to control plasticity. Collectively, these models demonstrate a range of different psychometric curves that we compare to human performance. Keywords: learning, memory, recognition, familiarity, novelty detection, meta-learning, Hebbian, synaptic plasticit