Optoelectronic reservoir computer at optimal computational power based on regulated light scattering

Abstract

The development and analyses of artificial neural networks have played a vital role in the rapid growth of artificial intelligence. Among various designs of artificial neural networks, recurrent neural networks has facilitated the processing of continuously fed input like a real brain, significantly contributing in the analyses of sequential data such as spoken audio. Yet the difficulty in training the network parameters has remained a major limitation of recurrent neural networks. Reservoir computing has emerged to circumvent this difficulty, initially by utilizing a randomly configured recurrent neural network with the least effort of training. More recently reservoir computing has been attempted using dynamic physical substrates instead of artificial neural networks, in the pursuit of higher efficiency in time and energy. These physical reservoir computing has been experimentally implemented with various substrates including electronic circuits, mechanical bodies and biological neurons. Optical reservoir computing has also emerged along other approaches, offering exceptionally high potential in the computational speed. In this thesis, we realize a reservoir computing platform based on the scattering of light, as well as searching and suggesting the optimal design that secures a high ratio of the cognitive capability to the reservoir dimension. The working principle of the scattering-based reservoir system is to harness the spreading of optical information in scattering, in analogy with how information propagates into a larger dimension in biological and artificial neural networks. Reservoir computing based on light scattering offers a unique benefit in that the reservoir can be realized with a very large dimension. To further maximize this advantage, it is necessary to be able to obtain high performance at a given dimension, and we search for the suitable design parameters based on our knowledge in the control and measurement of light scattering. The parameters of interest include the intensity of measurement, characteristics of the scattering medium, ratio of the input nodes, and the weight of input signals. We identify a sharp criticality of performance in the intensity of measurement, as well as locating a consistent range of the two input-related parameters that display high computational performance. Based on these findings, we achieve a reservoir computing system that is effective in complicated problems including the prediction of a chaotic time-series and audio recognition.