Conference on Cognitive Computational Neuroscience

Place cells in the hippocampus have been proposed to encode a predictive representation of states, known as successor representation (SR) in the reinforcement learning literature. Here, we first examine how the SR can be learned in simple rate-based models by local learning rules. We find that we can learn the SR using the same learning rule at recurrent collaterals and feedforward synapses – both have been implied as locations for learning of the SR. We then investigate the successor representations that arise when one uses temporally (a-)symmetric versions of the plasticity rule. Indeed, we find that using a symmetric rule learns the SR under a symmetrized version of the policy. This might have relevant functional implications: while an asymmetric successor representation might be better suited for predictions under the current policy, a symmetric version could possibly be useful for generalization.