(A) fMRI decoding. Within-modulator decoding: classifier was trained and tested on data from each modulator separately, in a leave-one-run-out cross-validation approach. Across modulator decoding: classifier was trained on data from one modulator and tested on data from the other modulator (and vice-versa). Across-modulator decoding after shifting: data for one modulator was shifted by 90°, then the classifier was trained on that modulator and tested on the other (and vice-versa). (B) Multidimensional scaling. Top: population responses to all orientations for both modulators (solid lines, radial modulator; dotted lines, angular modulator) are plotted in 2D, such that distances between conditions reflect how different (i.e., dissimilarity = 1–r, where r is the correlation coefficient) the responses are. For the original data, orthogonal orientations from the two modulators have similar responses and are plotted near each other, while identical orientations taken from the two modulators evoked very different responses, and lay distant from each other. Bottom: after shifting the labels for the angular modulator by 90 deg, the two datasets align well, and similar orientations correspond to similar population responses regardless of the modulator.