Answers are grounded in this paper and cite their sources.
Despite the rapid progress of on-chip photonic computing, optical neural networks (ONNs) remain constrained by limited kernel expressivity, weak nonlinearities, and substantial hardware overhead. To address these limitations, we propose a compact microring resonator (MRR)-based photonic convolutional neural network (PCNN) for modulation format classification from constellation diagrams. This architecture is based on cascaded dual-microring units and add–drop MRRs; the former achieves dual-port mixing in the optical linear layer through a phase compensation mechanism while the latter utilizes differential pass-through and up/down-link readout to implement signed convolution kernel coefficients. On constellation diagrams derived from the RML2016.10a dataset, the proposed PCNN achieves accuracies of 82.58%, 87.85%, and 91.94% at input resolutions of 16×16, 32×32, and 48×48, respectively, with only 0.25 million trainable parameters and 0.001 GFLOPs. We further provide an explicit physics-to-network parameter mapping and a first-order implementation-budget analysis, indicating a compact and programmable photonic route to modulation format recognition under tight computational and hardware budgets.
No takes yet. Share an insight, caveat, or question.
Li et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: