A Versatile ONN Computing Platform through Optically-Collaborative Algorithms

Optical neural networks (ONNs) are emerging as a revolutionary platform for machine learning to address the challenges posed by rapidly increasing data volumes and computational demands. While extensive research has led to numerous optical hardware innovations, it is crucial to recognize that neural network (NN) algorithms remain integral to system performance. Although conventional electronic NNs (ENNs) have been advancing quickly, they are typically defined with precise, strict coding standards. In contrast, directly translating these algorithms into ONNs can lead to significant deviations owing to the different data carrier – light – resulting in subpar performance or complex implementations. Aiming at overcoming these challenges, we specifically explored how optical-and-algorithm fine collaborations can significantly enhance ONN performances, and proposed a versatile and highly efficient ONN framework, coined as BONN, utilizing the binarized parameter quantization technology. Rather than relying on direct conversions, we achieve an algorithm-accurate optical network dataflow, highlighting subtle collaborations between optical components and NN algorithms. Compared with similarly structured ENNs, enhanced performances are achieved across multiple datasets. Additionally, by minimizing electronic–optical interactions and system complexity, we design a Generative Pre-trained Transformer (GPT)-oriented multi-task scheduler driven by BONN as a versatile black-box computing platform. The in-system BONN is able to deliver an ultra-fast maximum speed of 428 TFLOPs, a significant 96.88% memory usage reduction, together with a low-power consumption of just 2.5 W. Furthermore, leveraging the combined advantages of ONNs, BNNs, and 2D algorithms, the framework also demonstrates higher robustness against environmental vibrations and input noise over long-term operations.