| A basic assumption underlying traditional supervised learning algorithms is that labeled examples used to train a classifier are indicative of (drawn i.i.d. from the same distribution as) the test sample. However, a common problem in signal processing violates this assumption: given clean training examples, classify a signal that has propagated through a noisy linear time-invariant system. This traditional signal processing problem is recast as a dataset shift problem for machine learning, in which training and test distributions differ. Joint deconvolution and classification is proposed as a system-optimized framework for classifying a channel-corrupted signal from clean training features. In particular, classifiers are designed to account for the convolution relationship between test and training distributions. The joint MAP classifier jointly estimates a clean signal and a class label from a multipath-corrupted signal. The joint QDA classifier probabilistically accounts for the convolution relationship, and is extended for use with subband energy features. A set of kernels are proposed that measure similarity between a clean training signal and a corrupted test signal, and their use for channel-robust SVMs is proposed. With a focus on passive acoustic classification for multipath-corrupted signals, classifiers are tested in experiments to classify simulated narrowband acoustic signals, to identify Bowhead whales from their vocalizations in shallow water, and to acoustically identify trumpeters in a reverberant environment. |
