Transition from plasticity to dynamics-based value update for meta-learning

Value computation is fundamental to the survival of animals. Classical models suggest value is stored in synaptic weight through plasticity whereas more recent theories propose that recurrent network dynamics can encode and update value independently of synaptic change. Although these two mechanisms are not mutually exclusive, the condition under which each dominates remain unclear. Here we show that environmental stability determines the neural mechanism underlying value updating. In stable environments, mice rely on synaptic plasticity in the basolateral amygdala (BLA), supporting value memory that persists for over a week. In contrast, in dynamic environments, mice gradually transition to a recurrent dynamics-based value update, with value memory degrading after ~2 minutes. A recurrent neural network (RNN) endowed with online weight updating recapitulates experimental data, exhibiting a transition from plasticity to dynamics-based value computation. Synaptic plasticity blockade in the BLA impaired initial value updating in the stable environment but not expert value updating in the dynamic environment, consistent with the RNN model. Furthermore, electrophysiological recordings and optogenetic inactivation reveal that BLA encodes both stable and dynamic value and is required for performing in both environments. Lastly, the transition from plasticity to dynamics-based value computation allowed both RNNs and mice to use task knowledge to infer value, allowing value updating without direct experience. Overall, our work provides a new framework to understand value computation in the brain and suggests that meta-learning may emerge via a transition from plasticity to dynamics-based value computation.