A Simple Conceptor Model for Hand-written-digit Recognition

Traditional recognitions of the MNIST hand-written-digits need vast amounts of datasets to assure high accuracy based on artificial neural networks (ANNs). In this paper, we present a simple preprocessing method for image classification. Firstly, the image pixels are converted into spike streams by using the Poisson distribution method. Similar as the integration of synaptic current in brain, spike or binary streams are integrated into continuous signals which are used to feed into the input layer of the conceptor network. The conceptor network is a recurrent neural network used to generate high-dimensional dynamic information. We use the MNIST database to investigate the computational performance of this model. Our results show that this method can achieve high recognition accuracy with much smaller training samples (6000 in this model V.S. 60000 in traditional other methods). Note that in this model, information for each image is decoded into a continuous sequence and fully analyzed through the conceptor network. Therefore, the number of training samples can be remarkably reduced.