Capturing rapid learning in an extended successor representation theory of the cognitive map

Abstract

Animals sometimes adapt their behavior after a single exposure to novel information. Here, we integrate abstract theory with recent mechanistic insights to understand this and other examples of rapid adaptation. We ground our investigation in studies of animal learning in spatial contexts, where it is widely held that learning depends on the formation of a cognitive map reflected in learned representations in hippocampal areas CA1 and CA3. We extend the successor representation (SR) theory of the cognitive map, incorporating the weighting of this representation by perceived salience (PS) signals, and using it to represent future availability of environmental features that allow behavior to be guided flexibly by an animals goals and needs at later times. To capture rapid adaptation, we rely on behavioral time-scale synaptic plasticity (BTSP), a form of plasticity that captures one-shot formation of new place fields. Through simulations, we show that our hippocampal model network can learn the PS-weighted SR rapidly and set up the ability to replay experiences in the environment after a single exposure through BTSP. Spontaneous replay activity during subsequent offline periods strengthens the PS-weighted SR. Predictions of environmental features also develop rapidly, and behavioral simulations show how these learned representations give rise to need-dependent choice, rapid behavioral adaptation to changes in the availability of reward, and one-trial shock avoidance learning.