Figure 4.

Predicting Subjects’ Responses Within and Across Task with Different Models

(A and B) Individual subjects. Performance of cognitive tomography is shown for within-task (red) and across-task predictions that is using subjective distributions inferred from one task to predict behavior in the other task (pink). The dashed line shows chance performance. Subjects are ordered by their average consistency on the two tasks (as in Figure 2).

(C and D) Group averages (mean ± SE) comparing cognitive tomography (red and pink bars) to alternative predictors. Replacement of subjective distributions with moment-matched Gaussians, thus ignoring the fine structural details of the subjective distributions, decreases performance (dark blue, within task; light blue, across task). A Gaussian process (GP) classifier that is directly optimized to fit subjects’ stimulus-to-response mappings without assuming the existence of subjective distributions also performs worse and is unable to generalize across tasks (green bars). ^(∗)p < 0.10, ^∗p < 0.05, ^∗∗p < 0.01.

(E and F) Within-task predictive performance of cognitive tomography for each subject (symbols color coded as in Figure 2) against their consistency levels. Boundary of gray shaded area shows expected upper bound on the performance of any predictor as a function of consistency. Error bars show 95% confidence intervals.

See also Figure S4 for a more detailed analysis of predictive performance.