Convex combination initialization method for kohonen neural network implemented in the CMOS technology
The paper presents a new CMOS implementation of the initialization mechanism for Kohonen self-organizing neural networks. A proper selection of initial values of the weights of the neurons exhibits a significant impact on the quality of the learning process. A straightforward realization of the initialization block in software is simple, but in hardware it requires providing the programming signal to all weights of each neuron. This makes the layout of the chip very complex, especially in case of large networks. This paper presents a new approach, in which to program particular neuron weights we use the same lines that are used by the adaptation block. This proposal is the first known transistor level implementation of the Convex Combination Method (CCM) that so far was implemented only in software.