Engineering photonics for efficient and scalable machine learning systems

On-device learning revolutionizes continuous learning and data protection, transcending traditional digital computing’s limitations in energy efficiency and scalability. Photonic technology, while increasing energy efficiency by a hundredfold, encounters scalability challenges. In this talk, we introduce photonic dimension engineering, designed to simultaneously amplify energy efficiency and scalability through two approaches: First, we manipulate RF modulation signals to craft compact high-dimensional photonic convolution processors, achieving a sixfold increase in efficiency compared to the Nvidia A100X GPU. Second, we optimize data flow to establish comprehensive machine learning systems for seamless on-device inference and training, boasting a hundred times greater energy efficiency than existing machine learning hardware, without compromising on latency. Our approach in photonic engineering not only heralds new pathways for the rational development and enhancement of artificial intelligence systems but also opens up possibilities for advancements in biomedical imaging and sensing.