Adaptive least error rate algorithm for neural network classifiers

We consider sample-by-sample adaptive training of two-class neural network classifiers. Specific applications that we have in mind are communication channel equalization and code-division multiple-access (CDMA) multiuser detection. Typically, training of such neural network classifiers is done using some stochastic gradient algorithm that tries to minimize the mean square error (MSE). Since the goal should really be minimizing the error probability, the MSE is a "wrong" criterion to use and may lead to a poor performance. We propose a stochastic gradient adaptive minimum error rate (MER) algorithm called the least error rate (LER) for training neural network classifiers.