Figure 3.

Relationship among GMF features, DNN activations, and observed behavior

(A) Mutual information (MI) between human behavior and test-set predictions from GMF features.

(B) y axis: MI between human behavior and test-set DNN predictions; x axis: redundant information about human behavior that is shared between DNN predictions and GMF shape feature predictions. These plots show that DNN prediction performance of human behavior increases on the y axis when the DNN embedding layers represent the same shape features as humans. Each data point in (A) and (B) represents the combination of one test set, one participant, and one familiar identity. Overlaid lines reflect the 95% (bold) and 50% (light) highest posterior density intervals (HPDIs) of the corresponding main effects of predictor spaces from Bayesian linear models fitted to the MI and redundancy values.

(C) Fractions of MI and redundancy data points exceeding noise threshold (95^th percentile of MI and redundancy distributions obtained from trial-shuffled data).

(D) Comparisons of the posterior distributions of the main effects for all predictor spaces from Bayesian linear modeling of the raw performances. For each pair in the matrices, the grayscale color map shows the fraction of samples of the predictor space color coded on the y axis that is larger than the predictor space color coded on the x axis (testing a hypothesis).

Colors in (C) and (D) correspond to those in (A) and (B). See also Figures S2–S8 and S21.