Experience shifts value updating from synaptic plasticity to neural dynamics

Abstract

Adaptive behavior requires durable value memories and rapid updating. Experience in dynamic environments accelerates value updating, but the underlying neuronal mechanisms remain unclear. Mice exposed to repeated reversals of odor-outcome contingencies acquired faster value updating at the cost of faster memory decay. Recurrent neural networks (RNNs) trained with continual plasticity revealed two routes to faster learning, one through more efficient synaptic plasticity and one through a transition to value computation by neural dynamics. Only the latter reproduced the trade-off between updating speed and memory persistence observed in mice. Consistent with this transition, inhibiting dopamine neurons at reward delivery or pharmacologically inhibiting CaMKII in the amygdala impaired initial learning but largely spared expert reversal learning. In vivo recordings in the amygdala and hippocampus revealed representations of cue value and latent context necessary for rapid value updating. Context and value signals decayed over prolonged inter-trial intervals, paralleling behavioral forgetting. Conditional deletion of CaMKIIa in Emx1-cre-expressing neurons spared initial associative learning but impaired expert reversal performance, as did freezing the recurrent weights of RNNs, suggesting distinct loci for initial value learning and meta-learning. Together, these findings reveal how experience shifts value computation from synaptic plasticity toward neural dynamics, providing a mechanistic framework for the trade-off between memory stability and behavioral flexibility.