Model-free optimization and parallel architecture towards monolithic-hybrid-photonic-electronic reservoir computing

Physical reservoir computing (PRC) is a recently developed variant of neuromorphic computing, where the output from a nonlinear physical system is utilized to perform various machine learning tasks. In this work, we theoretically analyze the performance of a photonic waveguide mesh (WGM) with electro-optic phase shifters for monolithic-hybrid-photonic-electronic reservoir computing (MHPE RC), where the phase-to-intensity relations in the photonic circuit provide nonlinearity and high dimensionality, while the electronic circuit provides the input and feedback with tunable parameters. First, we numerically demonstrate the efficiency and performance superiority of a parallel architecture comprising fabricated WGM. Next, we present the Lyapunov filtered-minimal redundancy maximal relevance (Lf-mRMR) algorithm, which optimizes the electronic parameters of parallel WGMs by analyzing the Lyapunov exponent and the mutual information between the output of the corresponding WGMs and the required task. The Lf-mRMR algorithm is computationally less complex, substantially improves the performance of MHPE RC, and can tolerate fabrication errors. We present the selective parallel architecture for reservoir computing (SPARC), which, assisted by the Lf-mRMR algorithm, can achieve performance close to convolutional neural networks. Finally, we experimentally employ on-chip silicon photonics with thermo-optical phase shifters and external off-chip digital memory and control unit to validate the advantageous performance of Lf-mRMR-assisted RC.