Filter bank learning for signal classification

This paper addresses the problem of feature extraction for signal classification. It proposes to build features by designing a data-driven filter bank and by pooling the time–frequency representation to provide time-invariant features. For this purpose, our work tackles the problem of jointly learning the filters of a filter bank with a support vector machine. It is shown that, in a restrictive case (but consistent to prevent overfitting), the problem boils down to a multiple kernel learning instance with infinitely many kernels. To solve such a problem, we build upon existing methods and propose an active constraint algorithm able to handle a non-convex combination of an infinite number of kernels. Numerical experiments on both a brain–computer interface dataset and a scene classification problem prove empirically the appeal of our method. [Display omitted] •We propose a method of feature extraction, using a large-margin framework.•We extend generalized multiple kernel learning to infinitely many kernels.•We take a fresh look at learning convolutional features for signal classification.