Modeling the neural mechanisms underlying reversal learning in spatial navigation

To adapt to a dynamic environment, animals must abandon outdated behaviors and acquire new ones, a phenomenon known as reversal learning (RL). Despite ample knowledge about neural representations of space, the mechanisms of RL in spatial navigation remain understudied. To fill this gap, we extended an existing closed-loop simulator of spatial navigation and learning (Ghazinouri et al. 2023). The model is based on a spiking neural network that captures features of the hippocampus, specifically place cells and boundary cells. These cells feed inputs into action selection neurons, whose activity drives movement in certain heading directions. Synaptic plasticity was implemented as either symmetric or asymmetric STDP coupled with an eligibility trace and a reward-dependent factor. Agents were tested in an RL task using an A-B-A design with two different goal locations. The agents learned the path to the reward, i.e., latencies decreased over trials. With symmetric STDP learning the first goal was fast, but RL failed, whereas with asymmetric STDP learning the first goal was slower and RLreverses to an equal level was comparable. This flexibility arises because the combination of synaptic depression and potentiation in asymmetric STDP allows the network to reconfigure the synaptic weights in the network to learn a new tasks. This flexibility comes at the cost of the agent partially forgetting the first goal when learning the second one. Our results are therefore another demonstration of the well-known plasticity-stability trade off.