Autoencoder path ii : connectivity flavor i to connectivity flavor i . This path begins by stacking the upper triangular portion of each subject’s connectivity matrix into input {X}_{i} in R^{{n}_{edges}x {n}_{subj}} . A precomputed, fixed PCA transformation normalizes the data and reduces dimensionality to {X}_{i}^{'} in R^{256x {n}_{subj}} , equalizing the size of disparate input flavors. A single fully connected layer Encoder i , followed by L 2 normalization, transforms {X}_{i}^{'} into a latent hypersphere surface {z}_{i} in R^{128x {n}_{subj}} . Batch-wise encoding loss L z ( z i ) controls inter-subject separation in latent space. A single fully connected layer Decoder i transforms z i to {X}_{i}^{'} , and batch-wise reconstruction loss {L}_{r}({X}_{i}^{'},{X}_{i}^{'}) and L z ( z i ) are backpropagated to optimize Encoder i and Decoder i .
