Clustering-Based Nonlinear Training Algorithm for Precision Constrained Photonic Micro-Ring Convolution Chip

The photonic convolutional neural network (CNN) represents a groundbreaking approach that promises to deliver immense computational power for artificial intelligence (AI) applications, including feature recognition, climate analysis, and disease diagnosis. However, limited by the resolution loss in the analog opto-electronic devices, photonic CNN currently suffers serious accuracy barrier, which is far lower than that of digital accelerators with up to 64 bit precision floating point format. In this paper, we propose a nonlinear distributed training method by embedding a nonlinear quantizer in the straight-through-estimator (STE) training algorithm applied on integrated photonic 2D-convolution chip based on time-frequency interleaved modulation. We demonstrate by experiment a higher accuracy of 88$\%$ over 62.8$\%$ using the Fashion MNIST recognition task with only 2 bit precision in the whole opto-electronic interfaces.