MLP should use a Flatten() layer so we don't have to concat features