It occurred to me that functional connectivity is sort of like RSA (representational similarity analysis): calculating a bunch of distances (Pearson correlation, often), but over timepoints (e.g. volumes) instead of voxels.
I can make a similar sort of picture for linear SVM classification:
but it's quite a bit different. With classification we're analyzing (often) the test set accuracy which comes from testing the fitted classifier (represented by the purple dashed line) with new data. But with functional connectivity and RSA we're working with the correlation (or whatever distance metric) more directly, interpreting how related the BOLD is between two regions (over timepoints) or between two trials (over voxels).