Hundred-layer photonic deep learning

Abstract

In the artificial intelligence era propelled by complex computational models, photonic computing represents a promising approach for energy-efficient machine learning; however, error accumulation inherent to the analog nature limits their depth to around ten layers, restricting advanced computing capabilities towards large language models (LLMs). In this study, we identify that such error accumulation arises from propagation redundancies. By introducing perturbations on-chip to decouple computational correlations, we eliminate the redundancy and develop deep photonic learning with a single-layer photonic computing (SLiM) chip that exhibits error tolerance. The SLiM chip overcomes the depth limitations of optical neural networks, allowing for error rates to be constrained across more than 200 layers, and extends spatial depth from millimeter to hundred-meter scale, enabling a three-dimensional chip cluster. We experimentally constructed a neural network with 100 layers for image classification, along with a 0.345-billion-parameter LLM with 384 layers for text generation, and a 0.192-billion-parameter LLM with 640 layers for image generation, all achieving performances comparable to ideal simulations at 10-GHz data rate. This error-tolerant single-layer chip initiates the advancement of state-of-the-art deep learning models on efficient analog computing hardware.