A Multiple Kernel Machine with Incremental Learning using Sparse Representation

In this paper, a new multiple kernel machine with in-situ learning capability is introduced for pattern classification using sparse representation framework. A vector-valued score function for a given pattern is generated as a linear combination of multiple kernel functions, each of which measures the similarity between the pattern to be classified and the representative samples from every class. The representative samples for different classes are selected through the sparse approximations, by forcing the representation of the vector-valued score function to be sparse in a dictionary of kernel functions with respect to all data samples from different classes. After the sparse weight matrix is found, the vector-valued score function for a new pattern with unknown class label can be computed. An incremental learning process is developed which allows for incremental updating of the system without impacting the previously learnt patterns. The proposed multiple kernel machine was then implemented and benchmarked against other multi-kernel methods for classifying underwater objects using low frequency sonar data.