Fig. 3: Decoding stimulus representations from MEG.

A Localizer Task. The Localizer Task was completed prior to the risky-decision task and to learning choice-outcome probabilities. On each trial participants were shown an outcome or choice stimulus, and, on the next screen, selected a word corresponding to the stimulus they had just observed. B Activation Probability measure. We trained lasso-regularized logistic regression classifiers to discriminate MEG data from when a given outcome stimulus was presented compared to data from presentation of all other images and inter-trial intervals. Each classifier output an estimated probability that its stimulus was being presented (Activation Probability). Separate classifiers were trained at successive 10 ms bins of MEG data around stimulus presentation. In the example, lines display the group-mean (+/− s.e.m.) activation probability measure for the classifier corresponding to O2, for each training timepoint, applied to held out data from the same corresponding test timepoint. Color designates the true outcome stimulus presented. C Decoding accuracy. Cross-validation accuracy is the proportion of trials for which the classifier corresponding to the presented outcome (for held-out data) had the highest activation probability. Lines denote mean accuracy (+/− s.e.m.) for each set of 10 ms time-binned outcome classifiers, applied to the same time-bin on held out examples. Dashed line designates permutation threshold corresponding to the 95 percentile peak threshold for accuracy lines generated with shuffled labels. D Temporal specificity. Classifiers trained on each 10 ms time bin were also tested on every time bin from −350 to 800 ms following presentation of stimuli from held out data. The resulting accuracy image demonstrates temporal selectivity. Classifiers identify with good accuracy representations of stimuli specific to the timepoint on which they were trained. B–D Values reflect group means across 19 participants. A Scissor image was created by scarlab and published on svgrepo.com under an MIT license.