Figure 5. Subspace rotation increased with the performance at richly trained RNNs.

(A) Beta-space subspace rotation angles of RNNs’ representations before and after training. (B) An example of increased subspace rotation and task performance over the course of training. (C) Pearson correlation (r) between the subspace rotation angles and training performance (accuracy) in Remember 1^st (R1) and Remember 2^nd (R2) contexts. ***: p < 0.001, evaluated using a two-sided paired t-test for (A) and two-sided t-test for (C).