Learning representations of back-propagation errors