Echo State Networks-Based Reservoir Computing for MNIST Handwritten Digits Recognition

Reservoir Computing is an attractive paradigm of recurrent neural network architecture, due to the ease of training and existing neuromorphic implementations. Successively applied on speech recognition and time series forecasting, few works have so far studied the behavior of such networks on computer vision tasks. Therefore we decided to investigate the ability of Echo State Networks to classify the digits of the MNIST database. We show that even if ESNs are not able to outperform state-of-the-art convolutional networks, they allow low error thanks to a suitable preprocessing of images. The best performance is obtained with a large reservoir of 4,000~neurons, but committees of smaller reservoirs are also appealing and might be further investigated.