Abstract

Reservoir Computers (RC) are recurrent neural networks where most connections are fixed at random, and only output weights are trained [1]. Exhibiting memory about past inputs, RCs are perfect candidates for tasks involving signal processing, such as distortion-compensation in telecom channels. Due to their low level of complexity, RCs are easy to implement on unconventional computing platforms. A photonic RC can deliver low-cost non-electronic computation, without conversion to the electronic domain, with applications in scenarios like edge-computing [2].

© 2023 IEEE