Comparison of Feature Extraction Techniques for Handwritten Digit Recognition with a Photonic Reservoir Computer

Reservoir computing is a bio-inspired computing paradigm for processing time-dependent signals. Its photonic implementations have received much interest recently, and have been successfully applied to speech recognition and time-series forecasting. However, few works have been devoted to the more challenging computer vision tasks. In this work, we use a large-scale photonic reservoir computer for classification of handwritten digits from the MNIST database. We investigate and compare different feature extraction techniques (such as zoning, Gabor filters, and HOG) and report classification errors of 1% experimentally and 0.8% in numerical simulations.