Variable projection support vector machines and some applications using adaptive Hermite expansions

Abstract

Summary: We introduce an extension of the classical support vector machine classification algorithm with adaptive orthogonal transformations. The proposed transformations are realized through so-called variable projection operators. This approach allows the classifier to learn an informative representation of the data during the training process. Furthermore, choosing the underlying adaptive transformations correctly allows for learning interpretable parameters. Since the gradients of the proposed transformations are known with respect to the learnable parameters, we focus on training the primal form the modified SVM objectives using a stochastic subgradient method. We consider the possibility of using Mercer kernels with the proposed algorithms. We construct a case study using the linear combinations of adaptive Hermite functions where the proposed classification scheme outperforms the classical support vector machine approach. The proposed variable projection support vector machines provide a lightweight alternative to deep learning methods which incorporate automatic feature extraction.