Combining self-organizing maps

Abstract

The author proposed a learning rule for a single-layer network of modules representing adaptive tables of the type formed by T. Kohonen's vector quantization algorithm (Rep. TKK-F-A601, Helsinki Univ. of Technol., 1986). The learning rule allows combination of several modules to learn more complicated functions on higher dimensional spaces. During learning each module learns a function, which is adjusted such as to minimize the average square error between the correct function and the function represented by the network. Although this is a single-layer system, the capability of each module to learn an arbitrary nonlinearity gives the system far more flexibility than a perceptron. At the same time, for output nonlinearities that are a product or a sum of monotonous functions of their arguments there is a unique minimum to which the system is guaranteed to converge.<>