The effect of training signal errors on node learning
John R. Pemberton, Vincent Vidal · 1989
Summary form only given. The response of three node learning rules to errors in the training signal was examined. The discrete (perceptron) learning rule is shown to be very susceptible to target signal errors, and on the average each target signal error causes an output error. In contrast, the linear (Widrow-Hoff) and nonlinear (generalized delta rule) learning rules are able to tolerate a large amount of noise in the target signal without affecting the node output function. The ability to tolerate training signal noise is shown to depend on the learning rate and the shape of the nonlinear threshold function. For example, a nonlinear threshold function, with scale factor equal to 1 and learning rate constant equal to 0.01, produces less than 5% output errors for a rate of 40% errors in the training signal. The training signal error tolerance of linear and nonlinear learning rules is explained in terms of the effect of the errors on the weight vector.>