Was there a particular change to the network or the way that it was trained that introduced the 'backtrack and error correct' mechanism?