Problem of Efficient Initialization of Large Self-Organizing Maps Implemented in the CMOS Technology

Initialization of neuron weights is one of key problems in artificial neural networks (ANNs). This problem is particularly important in ANNs implemented as Application Specific Integrated Circuits (ASICs), where the number of the weights becomes large. When ANNs are implemented in software, the weights can be easily programmed. In contrast, in parallel systems of this type realized as ASICs it is necessary to provide programming and addressing lines to each weight that causes a large increase in the complexity of such designs. In this paper we present investigations that demonstrate that Self-Organizing Maps (SOMs) in many situations may be trained without the initialization (with zeroed weights). We present example results of several thousands simulations for different topologies of the SOM, for different neighborhood functions and two distance measures between the learning patterns and particular neurons in the input data space. Simulations were performed for zero initial values, for small values (up to 1 % of full scale range) and for neurons randomly distributed over the overall input data space. The results are comparable that allows to reduce the complexity of the SOM implemented in the CMOS technology.