Kernel-based associative memory

We propose a new approach to pseudo-inverse associative memories using kernel machine methodology. Basing on Hopfield-type pseudoinverse associative memories we developed a series of kernel-based hetero- and auto-associative algorithms. There are convergence processes possible during examination procedures even for continuous data. Kernel approach enables to overcome capacity limitations inherent to Hopfield-type networks. Memory capacity virtually does not depend on data dimension. We provide theoretical investigation for proposed methods and prove its attraction properties. Also we have experimentally tested them for tasks of classification and associative retrieval.