Undersampling depth and frequency masks occurrences of recency-like behavior by RNNs.

(A) Example of undersampling depth. Serial position curve is constructed based on occurrence of error trials with particular cue arrangements. Task variations with large cue set size (100 vs 5 cues; 20x difference) require more sampling. Increased test sampling depth (5000 → 30000 → 100000 test samples) of same network reveals recency-like behavior when none was apparent at lower test sampling depth (5000 test samples). (B) Recency-like behavior could also be masked by sampling frequency for networks/tasks that rapidly progressed from random to target-selective strategy. Increased sampling frequency (training points ∈{0, 11, 20, 40, 50} → {0, 5, 11, 15, 20, 30, 40, 50} → {0, 5, 7, 9, 11, 13, 15, 20, 30, 40, 50}x103) reveals recency-like behavior when none was apparent at lower frequency sampling. (EPS)