If this works, is there any theory why training models with low rank layers (y = (A.B).x + b) directly doesnt work? (or do they?)