A new kernelized associative memory and some of its applications

The classical Bidirectional Associative Memory (BAM) allows for the storage of pairs of vectors, such that when either member of the pair is presented to the BAM, the other member may be successfully recalled. This work presents a novel BAM, improved with respect to its capacity and noise performance through the use of the kernel trick, a common technique in machine learning for transforming linear methods into nonlinear methods. By kernelizing the BAM's energy function directly and defining new methods for recall, the kernel BAM shows improved performance compared to both the original BAM as well as a previously existing nonlinear BAM. This is demonstrated with thorough experimentation on synthetic datasets, and several practical applications are given on real data.Peer Reviewe