Tracking concept drift in a single neuron

We consider the performance of a variety of learning algorithms for single linear threshold neurons where the weights of the neuron change as training examples are presented. We restrict the weight changes to small changes referred to as the concept drift problem. The performance of the different learning algorithms (tracking algorithms) is defined by the average generalization error which is dependent on the concept drift, the nature of the tracking algorithm, the information given to the tracking algorithm, and the number of inputs, n. We analytically determine the average generalization error for a wide variety of tracking algorithms and different types of concept drift. We model this problem as a system identification problem with a single target neuron and a single tracking neuron