Default Feature Representations of the Cognitive Map

Cognitive maps allow animals and artificial agents to predict where they are likely to go next, and computational models show how such predictive maps guide planning. Environments, however, rarely stay fixed. When the goal changes, a door closes, or a barrier appears, the predictive map needs to be updated. However, given that predictive maps look far ahead, a local change in the environment potentially influences what is predicted from many other states, including distant ones, even though the agent experiences it only locally. Here, we propose Default Feature Representations (DFR) to address this problem. Our solution separates a predictive map into a fixed set of features shared across environments and an operator that carries the current environment. We show that the operator has a model-based closed form when the structural change is known (DFR-MB), and we present a temporal-difference rule for learning it from sampled transitions (DFR-TD) and prove that it converges to the same solution. Using samples alone, DFR-TD efficiently learns the perturbed map, plans about as well as the model-based solution, and outperforms successor-representation baselines on replanning. The model also reproduces the local remapping of grid cells under environmental change, mirroring the stability of entorhinal grid fields across environments.