Reservior Computing Based on GIST Feature Extraction and Nonlinear Synaptic Integration

Traditional image classification relies heavily on vast datasets to achieve high accuracy when using artificial neural networks (ANNs). In this paper, we introduce a novel reservior computing method for image classification that can attain high recognition accuracy with significantly fewer training samples. Specifically, we first convert grayscale images into spike streams based on their pixel values, following a Poisson distribution. These spike streams are then integrated into continuous signals, mimicking the integration of synaptic currents in the brain. Finally, we compute the GIST descriptors of these continuous signals, which are fed into the input layer of a spiking neural network for training and classification. We demonstrate the effectiveness of our method by applying it to the MNIST database using an Echo State Network (ESN). Our results show that our approach achieves high recognition accuracy with far fewer training samples (2000 in our model compared to 60,000 in traditional methods). Thus, our method significantly reduces the number of training samples required.