Noise robustness of EBNN learning

A variety of methods have recently been proposed for constraining neural networks to fit various constraints while being trained. One such approach is to constrain the function approximated by the network to fit desired slopes, or derivatives. Such slopes may be provided by the designer, as in Simard´s character recognizer network which was constrained so that the slope of the output with respect to translations, rotations, etc. of the input should be zero. Alternatively, target slopes may be generated automatically by program as in explanation based neural network (EBNN) learning. While slope information is known to improve generalization, sometimes slope information as well as value information is corrupted by noise. This paper explores the effects of noise in value and slope information on EBNN learning, compared with standard backpropagation. Experimental results show several characteristics of noise robustness of EBNN learning.