Fig. 3: Coupling of hippocampus spike entropy (PE) to late-layer image features specifically predicts memory performance.

All p values have been computed using non-parametric permutation tests (see Online Methods). a: Zero-order eta estimates for early-layer, late-layer, and all-layer coupling vs. performance (left, bar graphs including bootstrapped 95% CI). Recognition performance was captured by a principal component score that combined accuracy, dprime and confidence. Note that criterion was not related to spike PE coupling (Fig. S3), represents bias rather than performance, is very weakly correlated with all performance measures (see Methods), and was hence omitted in this case. Zero-order relationship between hippocampal spike coupling estimates (individual latent correlations) and recognition performance (principal component capturing accuracy, dprime, and confidence) for early-layer models (middle) and late-layer models (right). All correlations are non-parametric (Spearman; n = 34). b: Unique links of hippocampal spike coupling to late-layer features after controlling for coupling to early-layer features (left), coupling of spike rate to late-layer features (middle), and a set of other potential between-subject confounds (number of neurons and trials, task variant, encoding duration, age; right). c: Effects are specific to the hippocampus. Coupling of amygdala spike entropy to either early-layer (left) nor late-layer features (middle) did not predict recognition performance (n = 30). Finally, controlling for amygdala spike PE coupling to late-layer features did not reduce the link of hippocampus late-layer coupling to performance (right).