The suppression of learning at the hidden units of neural networks

Graduation date: 1990Under certain conditions, a neural network may be trained to perform a\ud specific task by altering the weights of only a portion of the synapses.\ud Specifically, it has been noted that certain three layer feed-forward networks may\ud be trained to certain tasks by adjusting only the synapses to the output unit.\ud This paper investigates the conditions under which this hobbling of the process\ud may be possible.\ud The investigation assumes that the existence of a set of weights which\ud perform a task with low error implies that the task is learnable. Thus an\ud algorithm is developed which attempts to find such a set, given the weights at the\ud hidden units as fixed. Success of the method is equated with the ability to\ud suppress learning at the hidden units.\ud The result is a classification of tasks for which suppression is possible. In\ud general, classification problems require learning at the hidden units, but\ud approximation of simple continuous functions does not