New learning rules for the ASSOM network

Abstract

The ASSOM is a self-organising neural network with the capability of adapting to linear subspaces. Here we propose two new methods to train the ASSOM network. A nonlinear system of equations is derived for network training. This system can be solved by a gradient-based approach or by the Levenberg–Marquardt method. Each of these two approaches gives a different learning rule. A comparison is carried out among the original Kohonen’s method and the proposed learning rules. Experimental results are reported, including a convergence speed experiment and a speech processing application, which show that the new learning rules have better performance than the original one.