Cue-dependent sorting directions of activity state space underlie recency-like phenotype.

(A) Activity states of an RNN trained on 50k episodes after first three cues. Coloring by first cue (leftmost of colored PC plots), second cue (middle), and third cue (right) reveal distinct sorting directions that organize points within clusters. Sorting angles approach orthogonality in PC3 space. (B) Same as (A) for the network after 20k episodes. Activity space clusters are less distinct, and sorting angles less close to orthogonality. In particular, θ12 is more acute at 20k than 50k. (C) Independently trained networks consistently showed an increase of the angle between the first and second sorting directions, θ12, from acute toward orthogonal with increased training from 20k to 50k episodes (after 20k training, θ12 = 38.5 ± 4.9°; after 50k training, θ12 = 61.5 ± 5.9°, average over 10 independently trained RNNs; *: p = 0.001, one-sided paired samples t-test. (D) Aggregated slope vs performance curve shows recency-like phenotype at 20k training episodes. (E) Same plots as in (D) but stratified by target cue identity. E.g. cue 1 plot measures serial position curve only when cue 1 is the target cue, i.e., 1-X-X-1, X-1-X-1, X-X-1-1. Curves stratified by cue 1 and 5 show strong recency-like phenotypes, whereas those for cues 2, 3, 4 do not. **: p p t-test. Curves for (D),(E) are averages over 10 independently trained RNNs. For (A) and (B) PCA was performed on all correct trials (97% and 83% of all trials, respectively) and top 3 PCs explain 94% and 88% of activity state variance, respectively.