Experimental and Theoretical Analysis of Deep Residual Time-Delay Reservoir Computing Based on Clipping Algorithm

Photonic reservoir computing (RC), as a crucial variant of recurrent neural networks, has boasted substantial potential owing to its straightforward training processes and facile hardware integration. In this work, we introduce and validate both experimentally and numerically a novel post-processing technique, referred to as the clipping algorithm (CA), utilized on the high-performance deep residual time-delay RC (DR-TDRC). The proof-of-concept model, i.e., a four-layer DR-TDRC, comprises 960 interrelated neurons (240 neurons per layer) based on four injection-locked distributed feedback lasers. Our results demonstrate that the implementation of the CA can effectively mitigate issues associated with the redundant layers and improve the performance of deep architectures.