Reservoir Computing is a sub-field of Machine Learning. The idea is to develop analog computers that can perform specific classification or prediction tasks. These computers fulfill the following conditions, which are necessary and sufficient from an information theory viewpoint: the Separation Property (SP), which allows the system to distinguish two neighboring but distinct states; the Approximation property (AP), which permits robust target identification robustness regardless of small perturbations; and finally, Fading-memory Property (or FP), that guarantees that the system does not indefinitely keep past states in memory. In general, only nonlinear systems with a large phase-space dimensionality can eventually fulfill these three conditions. The conceptual interest of these analog computers is that they operate in a radically different paradigm with regards to Turing Von-Neumann Machines. We know that these latter, despite their indisputable superiority as far as logical operations are concerned, are relatively inefficient for certain tasks such as pattern recognition (voice, faces, etc.).

The main interest in using photonic systems for Reservoir Computing is that it offers a large bandwidth and therefore, ultra-fast computation capability. Optoelectronic oscillators with delayed feedback are ideal candidates for this purpose, as they feature strong nonlinearity and large delay-induced dimensionality. On-going research focuses on a twofold objective: developping new architectures in order to increase the bandwidth of these computers, and investigating the potential of these computers with new benchmark tests featuring increasing difficulty, such as prediction tasks (requiring higher versatility), visual tests (requiring larger memory), or real-time pattern recognition (requiring wider bandwidth).

Further reading:
– Martinenghi et al. (2012)
– Larger et al. (2017)