Fig. 10: Decoding analysis pipeline. a. Computing dynamic correlations from timeseries data.

Given a timeseries of observations as a T × K matrix (or a set of S such matrices), we use Equation (4) to compute each participant’s DISFC (relative to other participants in the training or test sub-group, as appropriate). We repeat this process twice-- once using the analysis kernel (shown here as a Gaussian in the upper row of the panel), and once using a δ function kernel (lower row of the panel). b. Projecting dynamic correlations into a lower-dimensional space. We project the training and test data into K-dimensional spaces to create compact representations of dynamic correlations at the given order (estimated using the analysis kernel). c. Kernel trick. We project the dynamic correlations computed using a δ function kernel into a common K-dimensional space. These low-dimensional embeddings are fed back through the analysis pipeline in order to compute features at the next-highest order. d. Decoding analysis. We split the training data into two equal groups, and optimize the feature weights (i.e., dynamic correlations at each order) to maximize decoding accuracy. We then apply the trained classifier to the (held-out) test data.