Answers are grounded in this paper and cite their sources.
Photonic reservoir computing (RC) systems are attractive for neuromorphic information processing due to their minimal training overhead and reduced hardware complexity compared with conventional neural network architectures. However, prevailing delay-based RC systems face a fundamental speed-accuracy trade-off and limited task complexity due to their serial processing nature. In this work, we introduce an incoherent photonic RC system with parallel optical state update and parallel nonlinear activation nodes, operating without time-division multiplexing, and overcoming the intrinsic computing-frequency limitations of time-multiplexed implementations. The proposed architecture employs a programmable 4×4 Mach–Zehnder interferometer (MZI) mesh to realize the recurrent linear transformation, combined with four semiconductor optical amplifiers (SOAs) acting as parallel nonlinear neurons. By operating with an incoherent broadband light source, the system performs intensity-based computation and avoids coherent interference, while recurrence is intrinsically embedded in the spatial photonic network without relying on a conventional external delay-based virtual-node architecture. As a result, the architecture eliminates the fundamental speed bottleneck of delay-based RC and enables a simplified hardware design supporting processing rates up to 15 GS/s in experiments. The system is experimentally validated on temporal and classification benchmarks, including Mackey–Glass (MG) chaotic time-series prediction and MNIST handwritten digit recognition. At the maximum processing rate of 15 GS/s, the system maintains an NRMSE of 0.088, For pattern recognition, it attains an MNIST classification accuracy of 85.6% via only four physical nonlinear neurons. This work establishes a scalable and high-speed hardware paradigm for photonic neuromorphic computing, demonstrating a practical path toward real-time brain-like signal processing.
No takes yet. Share an insight, caveat, or question.
huili et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: