Fig. 3: Identifying shared patterns from neural population activity.

a Schematic of the decoding shared patterns across tasks: neural population activity from task A was reduced to a lower dimension (n = 16 dimensions) using a nonlinear method, and CNNs were trained separately to classify different visual stimuli and motor decisions. The reddish shadow area highlights neural activity used from the onset of visual stimulus presentation to the appearance of the choice buttons. b Example from monkey XW: classification accuracy of CNN decoders trained to decode visual stimuli and motor decisions from Task A/B/C on each session (n = 3 sessions). Trained decoders were then tested on Revisit-A task; Bars show mean accuracy, and error bars represent standard deviation across three recording sessions (A&B&C vs. chance level: Stimulus: P = 0.02, g = 2.27; Decision: P = 0.003, g = 6.36; Stimulus: Revisit-A vs. chance level: P = 0.30, g = 0.46; ABC vs. Revisit-A: P = 0.005, g = 4.48; Decision: Revisit-A vs. chance level: P = 0.008, g = 3.69, ABC vs. Revisit-A: P = 0.33, g = 0.91). The dark grey dashed line indicates chance level for stimulus classification (0.17), and the light grey dashed line indicates chance level for decision classification (0.5). Colored scatter dots represent individual sessions. c Generalization accuracy of decision classifiers trained on Task A when applied to Task B, Task C, and Task Revisit-A across three monkeys. Bars show mean accuracy across all sessions (9 sessions total: 3 from monkey AB,3 from monkey ZZ, 3 from monkey XW); Box plots show the median (centre line), 25th and 75th percentiles (box bounds), and whiskers extend to minimum and maximum values. Each dot indicates performance from one session, with different colors representing individual monkeys. The grey line denotes the decision classification chance level (0.5). Data analyzed by one-way ANOVA (P = 0.07, Cohen’s f = 0.56). d Decompose neural population activity from an example session of monkey AB into the decision- and the stimulus-related subspace, showing principal components explaining the highest proportion of variance in each subspace. Color intensity reflects different tasks (A-B-Revisit A) within a session, ranging from darker to lighter shades. Red and blue represent the trial average activity for choices toward the upper and lower buttons, respectively. e Neural activity during subsequent learning tasks from one session of monkey XW was projected into the primary principal component of the decision subspace using the projection matrix from task A. The top panel shows task B, and the bottom panel shows Revisit-A. St-A1: Stimulus 1 in Task A; St-A2: Stimulus 2 in Task A; St-Revisit-A1: Stimulus 1 in Task Revisit-A; St-Revisit-A2: Stimulus 2 in Task Revisit-A; *, P < 0.05, **, P < 0.01, ***, P < 0.001; two-tailed t-tests.