We could support the automatic mirrored factors by adding an "M" factor of the same shape as W, but flipped along the "L" axis. In other words:
- double the number of rows (K) in H to form H_expanded
- factorize X into the convolution of cat(W, M) and H_expanded, where cat(W, M) stacks W and W next to each other along the "K" axis
This would allow us to search for "backwards replay" events that mimic forward replay events, but that are exactly reversed.
We could support the automatic mirrored factors by adding an "M" factor of the same shape as W, but flipped along the "L" axis. In other words:
This would allow us to search for "backwards replay" events that mimic forward replay events, but that are exactly reversed.