Stimulus-specific breakdown of match errors suggests not mere recency.

(A) In a Match-First trial, a sequence of cues is presented and the agent must respond “yes” or “no” match by comparing the lead (cue A) and target (cue D) cues in the sequence. For a set size (i.e., possible cue types) of 5, there are 5 unique cues which we refer to as cue1 through cue5 (1–5, for short). We focus our analysis on the 25 neuron RNN which exhibited behavioral variation mimicking monkeys and humans over training on Match-First. (B) In the cue sequence labels, the number/letter represents the cue identity, A represents the cue of interest, and X represents any cue other than A. E.g. in the top row A = 1; dark blue bars indicate X-1-X-1 cue sequences (2–1-X-1, 3–1-X-1, 4–1-X-1, 5–1-X-1 from left to right) and light blue bars indicate X-X-1–1 cue sequences (2-X-1–1, 3-X-1–1, etc). A identities for each row is given on the right. For RNN trained on 20k trials and exhibiting recency behavior, both dark and light bars show uneven distribution of match errors across different lead cues. This refutes a simple recency-based strategy which would predict parity across all lead cues, i.e., 25% (horizontal dotted line in each graph) for all dark blue bars and 25% for all light blue bars. Error bars are SEM.