Adaptive kernel classifiers for short-duration oceanic signals

Two kernel networks are presented for the classification of short-duration acoustic signals characterized by wavelet coefficients and signal duration. These networks combine the positive features of exemplar-based classifiers such as the learned vector quantization method and kernel classifiers using radial basis functions. Results on the DARPA Data Set 1 show that these networks compare favorably with other classification techniques, with almost 100% accuracy achievable in identifying test signals that are similar to the training signals. A method of combining the outputs of several classifiers to yield a more accurate labeling is proposed based on the interpretation of network outputs as approximating posterior class probabilities. The authors also provide a technique for recognizing deviant signals and false alarms.<>