Figure 7. Forgetting as adaptive learning.

Our model assumes that animals create and update memory engrams to flexibly adjust their behavior to their environment. (a) We used a Rescorla-Wagner model that learns object-context associations in the object-based memory task. The key assumption is that stronger associations lead to higher memory performance, while weaker associations lead to more forgetting. (b) The Rescorla-Wagner model updated the strength of object-context associations (‘engram strength’) as a function of the prediction error (difference between experienced object and engram strength) and the learning rates governing the influence of the prediction error. (c) Based on learned representations, animals constantly predict what happens in the environment (e.g. the occurrence of objects), and if predictions are violated (prediction errors), engrams are updated to improve the accuracy of future predictions. Here, established engram cells are shown in green; non-engram cells in gray. (d) Positive prediction errors signaling the occurrence of an unexpected event (e.g. new object) induce a learning process that increases the probability of remembering. This might rely on the recruitment of new engram cells (shown in yellow). In contrast, negative prediction errors signaling the absence of an expected event (e.g. predicted object did not appear) induce forgetting. This might rely on ‘forgetting’ plasticity reducing access to engrams (light green cells). (e) Our model formalizes this perspective based on the notion of ‘engram relevancy’, i.e., the strength of the object-context association. Higher engram relevancy makes it more likely that an engram is behaviorally expressed, e.g., through exploration behavior. The presentation of a novel object (upper panel) leads to a high engram relevancy (middle panel) in response to a positive prediction error (lower panel). The absence of an expected object decreases engram relevancy through negative prediction errors. (f) Model simulations corroborate the behavioral effects of our data (Figure 3a). Gray lines and bars show the average exploration probability for the familiar and novel object according to the model; markers show simulated mice.

Figure 7—figure supplement 1. Dentate gyrus engram modulates object context associations.

(a) Behavioral paradigm for object-context associations. (b) Discrimination index. (c) Experimental timeline for optogenetic inhibition of object-context association. (d) Discrimination index. (e) Experimental timeline for optogenetic-induced object-context association. (f) Discrimination index. Bar graphs indicate average values in n = 6–11 per group (*p<0.05). Data graphed as means ± SEM. Panel a was created with BioRender.com. Panel c was created with BioRender.com. Panel e was created with BioRender.com.

Figure 7—figure supplement 2. In-sample model validation.

A comparison between the forgetting data and the forgetting curve predicted by the model suggests that the model captures the data accurately. (a) Control group (standard housing condition) of the environmental enrichment experiment. (b) Enrichment group of the environmental enrichment experiment. (c) Control group of the Rac1-inhibition experiment. (d) Rac1-inhibition group of the Rac1-inhibition experiment. Markers and bars show exploration probabilities of the mice. Lines show the average exploration probability for the familiar and novel object according to the model.

Figure 7—figure supplement 3. Out-of-sample model validation.

We examined the model fit using a simple cross-validation procedure. For the environmental enrichment data set, we estimated the free parameters based on half of the subjects. Then, we compared the predicted exploration behavior (dark lines) to the data of the other half of the subjects. (a) Control group. (b) Enrichment group (markers and bars show exploration probabilities of the mice). Furthermore, we compared the predicted exploration probabilities conditional on the parameter estimates from the environmental enrichment experiment to the observed exploration probabilities in the Rac1-inhibition experiment. (c) Control group. (d) Rac1-inhibition group.