Figure 8. Learning model captures the dynamics of forgetting.

Environmental, optogenetic, and pharmacological manipulations might modulate the speed of forgetting by altering key parameters of our model. Simulations with different learning-rate parameters explain the forgetting dynamics of the different experimental conditions. (a) The enrichment and Rac1-inhibition conditions were successfully captured using a low learning rate (0.01, similar to the empirical estimates). (c) In contrast, assuming a larger learning rate (0.5), we could capture faster forgetting as observed in the Rac1 activator and context-only conditions. (e) Moreover, improved memory performance after reminder cues can be explained by assuming that these interventions induce a positive prediction error boosting object relevancy. Here, we assumed a learning rate of 0.07 (based on the empirical estimate). (b) Development of engram relevancy and (d) prediction errors across conditions. (f) Probability of exploring the novel object plotted separately for each condition.

Figure 8—figure supplement 1. Model comparison.

We tested if our computational model (Rescorla-Wagner [‘RW’]) describes the data better than a control model (‘Baseline’) that explores each object with a probability of 0.5. The model comparison was based on a comparison of the cumulated Bayesian information criterion (BIC). Here, higher values indicate a better model fit. In both the environmental enrichment experiment (a, experimental and control condition) and the Rac1-inhibition experiment (b, experimental and control condition), the model comparison favored the RW-learning model. These results show that our model explained the data above chance level.

Figure 8—figure supplement 2. Parameter recovery.

We performed a parameter-recovery study to examine whether the free parameters of our learning model could be estimated accurately. Similar to the enrichment experiment, we assumed n=12 mice per inter-test interval (24 hr, 1 week, 2 weeks, 3 weeks). The recovery study indicates that the (a) learning-rate, (b) slope, and (c) exploration-variability parameters can sufficiently be estimated given the limited amount of data for model fitting. We validated that the variability of the parameters, especially of the learning rate, is lower when more subjects are included (e.g. n=50; not shown).