A flexible winner takes all neural network with the conscience mechanism realized on microcontrollers

A flexible, fully programmable Winner Takes All (WTA) Neural Network (NN) has been realized on two microcontrollers (μC) with the AVR and the ARM cores placed on a single board. The proposed device is equipped with an external analog-to-digital (ADC) and four digital-to-analog (DAC) converters, which enable operation with analog signals in parallel in the real time. One of the objectives of the proposed device is to be used as a hardware model of NNs realized as VLSI ASIC. In this paper we focus on the influence of the, so called, conscience mechanism on the learning quality of the NN. A properly tuned conscience mechanism enables activation of all neurons in the network, thus significantly decreasing the quantization error. In the paper we propose an alternative realization of such mechanism, which is especially suitable for transistor level implementations of both analog and digital NNs.