An Optimized Learning Algorithm Based on Linear Filters Suitable for Hardware implemented Self-Organizing Maps

Abstract. In this study, we present a fast and energy efficient learning algorithm suitable for Self-Organizing Maps (SOMs) realized in hardware. The proposed algorithm is an extension of the classical algorithm used in Kohonen SOM. It is based on the observation that the quantization error that is a typical quality measure of the learning process, does not decrease linearly along the learning process. One can observe the phases of the increased ‘activity’, during which the quantization error rapidly decreases, followed by ‘stagnation ’ phases, during which its values are almost the same. The activity phases occur just after decreasing the neighborhood radius, R. A set of finite impulse response (FIR) filters is used to detect both phases. This enables an automatic switching the radius R to a smaller value that shorts a given stagnation phase and starts a new activity phase. Comprehensive investigations carried out by means of the software model of the SOM show that the learning process can be shorten even by 80-95% that allows for reduction of energy consumption even by 70-90%.