On using backpropagation neural networks to separate single echoes from multiple echoes

Applications of neural networks to pattern classification problems in underwater acoustics have been an active area of research. Often, due to lack of a sufficient amount of data, the training data may not accurately represent the probability distributions of the classes to be classified. The authors give a simple and illustrative simulation example of a neural network performing unsatisfactorily under such circumstances. During training, a back propagation neural network classifier learns to recognize two classes of waveforms. Waveforms in Class 1 have two major peaks and low SNR. Waveforms in Class 2 have one major peak and high SNR. In testing it was found that the neural network classifier tuned in to their difference in SNR rather than the number of peaks.<>