Fig. 7: Effects of residual encoding on behavior during learning of orthogonal rules.

A Correlation between behavioral accuracy and population decoder accuracy according to the non-match rule (blue) and cue-identity rule (yellow) in Problem (A) (left) and (B) (right) during the non-match training phase. Classifiers were trained separately for each rat and session. Decoder accuracy was significantly correlated with behavioral accuracy according to the non-match rule in both Problem (A) (simple linear regression: R² = 0.57, β = 0.97, p = 0.0001) and (B) (R² = 0.50, β = 0.75, p = 0.0008). No significant correlation was observed for the cue-identity rule (Problem (A): R² = 0.12, β = –0.13, p = 0.086; Problem (B): R² ≈ 0, β = 0.02, p = 0.90). Full regression and interaction models are shown in Supplementary Table 10. B Same as (A), but during the cue-identity training phase. Decoder–behavior correlations remained strongly positive for the cue-identity rule in both Problem (A) (R² = 0.88, β = 0.92, p = 2.2 × 10⁻¹²) and Problem (B) (R² = 0.56, β = 0.61, p = 0.0003). Correlations with the irrelevant non-match rule reversed in later sessions of Problem (A) (R² = 0.23, p = 0.016). Session-level slope analysis confirmed this reversal: early sessions showed positive slopes (0.89, 0.37; r = 0.86, 0.95), whereas later sessions flattened or inverted (0.03, –0.46, –0.34; r = 0.06, –0.58, –0.87). Early slopes were positive on average (0.63) and late slopes negative (−0.26), with a significant Fisher Z difference (z = 4.20, p = 2.7 × 10⁻⁵; Supplementary Table 11). C Same as (A–B), for the controls. No significant decoder–behavior relationship was observed under the non-match rule (Problem (A): R² = 0.002, p = 0.84; Problem (B): R² = 0.10, p = 0.20). For the cue-identity rule, no correlation appeared in Problem (A) (R² = 0.13, p = 0.12), but a positive correlation was found in Problem (B) (R² = 0.26, β = 0.62, p = 0.033). D Correlation between irrelevant-rule decoder accuracy and the proportion of trials labeled “go” by this decoder but ignored by the animal (R² = 0.27, β = 1.14, p = 0.002; Supplementary Table 12). E Correlation between irrelevant-rule decoder accuracy and behavioral accuracy for the relevant cue-identity rule (R² = 0.34, β = 0.40, p < 0.001; Supplementary Table 12). F Correlation between cue-identity behavioral accuracy during its acquisition and non-match behavioral accuracy during its initial learning. No significant relationships were observed (Pearson p > 0.1; Supplementary Table 13), indicating that cue-identity learning speed was not predicted by earlier non-match performance.