Quantifying shared convergence between different plasticity rules.

A. Example learning trajectories in weight space. For an example plasticity rule (square) operating on an example representational space (ResNet50/avgpool), weight vectors near the beginning (open dots) and end (closed dots) of different learning simulations for a subtask are shown. One example trajectory (gray) for a single simulation is shown. In red is a trajectory from a different rule. B. Weight convergence. For the example encoding stage in panel A, the expected (over subtasks, simulations) squared pairwise ℓ2 distance in weight space between independent simulations from all n = 7 plasticity rules (red) remains close to the lower limit (gray), across trials. Shaded regions are the ± standard deviation over rules. C. Functional convergence. The probability that two distinct rules (red) generate the same behavioral prediction for a random, held-out test image is shown, and compared to the probability a rule agrees with itself across two independent simulations (dashed gray). D. Summary of weight convergence across all encoding stages. Shown is a normalized rule-to-other rule distance metric (normalized between random walk; rule-to-self) at trial 100 for all encoding stages (median: 0.02, max: 0.07). D. Summary of functional convergence. A normalized functional convergence metric (norm. between random guessing, rule-to-self) for all encoding stages is shown (median: 0.95, min: 0.92). See Section 2.3 in S1 Appendix for details. (TIF)