QUANTIZATION EFFECTS IN DIGITALLY BEHAVING CIRCUIT IMPLEMENTATIONS OFKOHONEN NETWORKS

Implementing a neural network on a digital or mixed analog and digital chip yields the quantization of the synaptic weights dynamics. This p aper addresses this topic in the case of Kohonen's self-organizing map s. We first study qualitatively how the quantization affects the conve rgence and the properties, and deduce from this analysis the way to ch oose the parameters of the network (adaptation gain and neighborhood). We will see that a spatially decreasing neighborhood function is far more preferable than the usually rectangular neighborhood function, be cause of the weight quantization. Based on these results, an analog no nlinear network, integrated in a standard CMOS technology, and impleme nting this spatially decreasing neighborhood function is then presente d. It can be used in a mixed analog and digital circuit implementation , that will therefore be consistent with the conclusions obtained in t he first part of the paper.