Digit image classification using magnonic reservoir computing

Abstract

Reservoir computing represents a special type of recurrent neural networks that allow effective solving hard tasks, such as speech recognition, object classification, and time series prediction, without requiring large time and energy costs for training. It is also desirable to use cheap components as well as cheap assembling technology for mass production of physical reservoirs. Magnonics is one promising technology to achieve this goal. However, conventional material for magnonics—single-crystal yttrium iron garnet (YIG) films—are quite expensive and require sophisticated fabrication by liquid-phase epitaxy. Here, we propose and validate a magnonic reservoir computing with an YIG ceramic slab (fabricated with a simple, very well established, and cheap ceramic technology) and a feedback loop for the classification of digits in the form of images. The findings demonstrate that this reservoir architecture is capable of effectively classifying images of printed and handwritten digits ranging from zero to nine. The spin waves propagating within the YIG slab provide sufficient memory capacity and nonlinearity, achieving classification accuracies exceeding 76% for 5 × 4 pixel printed digits with 20% pixel errors, 98% for 10 × 10 pixel printed digits with the same level of pixel errors, and 71% for handwritten digits from the MNIST (Modified National Institute of Standards and Technology) data set. Such ferrite ceramic slabs represent promising platforms for the development of magnonic computational devices.