Summary

The hippocampus as a predictive map

10 Sep 2021 - Tianqi Xu

In traditional, place cells are thought to represent its memorized place per se. However, there are emerging evidence show that place cells not only code the position they fire maximally. But, they contain some information about future states. Stachenfeld et al. approach the puzzle from a reinforcement framework, in which the state(place) and the reward of a state are decomposed in the Value function. They argue that a well-trained place cell can represent its successors, thus they build a Successor Representation (SR) model. The SR model implements a series of equations adopted from reinforcement learning framework. The key point is the policy of state transition function transform from random work to directional biased. From this aspect, the place field of SR skewed toward the unbiased direction. They then compare simulations of SR model with real experimental data. The results show pretty well similarities between them. Furthermore, since grid cells show periodic firing patterns, and the downstream place cells do not, they hypothesis the eigendecomposition of SR matrix can mimic grid cells. The simulation results of eigenvectors fo SR matrix acting for grid cells also reproduce similarities to experimental data. Those results indicate that place cells in the hippocampus formulate a predictive map.

Original author： Tianqi Xu
Link： https://CNeuroUSTC.github.io/2021/09/10/TianqiXu.html
Copyright Notice： All articles on this WenLab site are subject to a BY-NC-SA license agreement, except for special notices. Reprint please indicate the source!