S-128 | Context-based remapping in artificial neural networks shows shared representational features previously found in the rodent hippocampus

Theoretical and Computational Neuroscience
Author: Mauro Navarra | Email: mauro93866@gmail.com

Mauro Navarra^1°, Camilo J. Mininni^1°

^1° Instituto de Biología y Medicina Experimental

Changes in firing patterns of navigational brain regions (‘remapping’) occur (although not exclusively) after changes in contextual cues. In particular, hippocampal place cells remap by either changing their place field location or place field firing rate. Previous experimental studies in rodents have shown that the maps at the neural activation level that emerge in two distinct contexts are not completely independent, but share a representational geometry even across different individuals. Can this shared representational geometry in the remapping phenomenon be understood as a consequence of the computational problem being solved by the system? To answer this question, we trained recurrent neural network models to solve a spatial navigation task with two contexts that are a geometric transformation from one another (in particular, rotations). Using principal component analysis on neural activations and Procrustes transformations (rotation, translation and scaling) over said components, we observed that the representations in both contexts aligned. This result shows that the experimental observations in previous studies can be replicated in artificial neural networks, suggesting that the mechanism underlying the differences in firing patterns at the neural activation level arises as a solution to the optimization problem of path integration and context discrimination.