Fig. 6.

Denoising of brain-based RDMs improves correspondence with behavior-based RDMs.

In Experiment 4, participants were asked to make behavioral judgments about the stimuli used in the fMRI experiment, and an RDM was constructed from these behavioral judgments. We averaged the Test and Re-test RDMs derived from the fMRI data and then correlated the lower triangle of this brain-based RDM to the lower triangle of the behavior-based RDM. (A) Results using GLMdenoise. This panel shows the correlation observed for each participant, before and after denoising. Black dots indicate the median across participants. Nearly all (19/20) participants show an increase in correlation and this increase in correlation is statistically significant (two-tailed sign test). (B) Comparison to other approaches. Each ‘x’ indicates for one participant the increase in brain-behavior correlation that is observed compared to Baseline (no denoising). The black dot indicates the median increase across participants (error bars indicate 68% confidence intervals obtained by bootstrapping participants). Although differences in performance across the three approaches are not statistically significant, notice that GLMdenoise consistently improves performance, whereas the Motion regressors approaches yield more mixed results (some datasets improve, some datasets worsen).