Clements-Enhanced Complex-Valued Coherent Mesh With Balanced Detection Units for Photonic Neural Networks

Photonic neural networks (PNNs), as emerging analog computing paradigms, have the potential to break both the power consumption wall and processing speed boundary of digital computing architectures. However, most reported PNNs were mainly focused on real-valued regions, where phases of optical fields were only utilized as indirect information to affect optical amplitudes. Moreover, the quantization of phase information at each node of multi-layer PNNs is disregarded. In this work, we propose and demonstrate a Clements-enhanced complex-valued mesh (CCM) with balanced detection units (BDUs), which can capture both the amplitude and phase information at one layer and pass them on to the next layer. Besides, by introducing <tex-math notation="LaTeX">$N/2$</tex-math> additional phase shifters to the Clements architecture, the proposed CCM possesses a complete characterization ability for arbitrary N-order complex unitary matrices. In this way, both the input vectors and the weighted matrices can be mapped into complex-valued regions and further implemented on the proposed CCM, in which the corresponding computing density is increased by four times with the comparison of real-valued meshes. To validate the proposed CCM with BDUs, we also fabricated a silicon-based on-chip 4×4 mesh with a footprint of 7.86 mm <tex-math notation="LaTeX">²</tex-math> and experimentally verified its performances on a two-class classification task of King-Rook vs. King (KRK) in chess and a ten-class classification of handwritten digits from MNIST database, yielding tested accuracies of 94.06 <tex-math notation="LaTeX">$%$</tex-math> and 91.75 <tex-math notation="LaTeX">$%$</tex-math> , respectively. The proposed passive on-chip CCM together with BDUs might provide another solution for next-generation artificial intelligence hardware with enhanced performances.