Figure 6.

Novelty-gated Hebbian learning. (A) The target maze (left) and test maze (right). Each one of the 16 rooms in the maze is represented by a random MNIST digit. Transitions are only possible between neighbor rooms. The test maze was generated from the target one by randomly replacing a few rooms (indicated with yellow borders). A sample trajectory is shown with the green arrows. (B) Evolution of the slow moving average of the free energy as a function of the amount of observed data for the target maze (blue) and the mean free energy of the same model when “teleported” to the control maze (red) every 500 s. (C) Fine-grained temporal details of the fast free energy traces e_N(t) while traversing the sample trajectory for the target maze (blue) and for the test maze (red). Visited rooms are indicated by the annotated coordinates at each change-point. During the intervals between two consecutive change-points the agent remains in the same room.