A groundbreaking study from the lab of院士 Harish Bhaskaran at the University of Oxford has been published in the prestigious journal Nature, showcasing the potential of photonic computing in analyzing Parkinson’s patients’ gait with an impressive 92.2% accuracy. This development signals a significant acceleration in the dawn of the photonic computing era, a promising alternative to traditional electronic computing.
The AIxiv column, a dedicated platform for academic and technical content by Machine Heart, has been instrumental in reporting on over 2,000 articles from top global universities and corporations, fostering academic exchange and dissemination. Researchers and institutions with groundbreaking work are encouraged to submit their contributions for consideration. For inquiries or submissions, interested parties can contact liyazhou@jiqizhixin.com or zhaoyunfeng@jiqizhixin.com.
Dr. Dong Bowei, the first author of the study, is a member of Bhaskaran’s research team. In 2022, several researchers from the院士’s group jointly established Photonic Bit Technology, a Chinese-based company specializing in photonic computing chips. The company made headlines at the World Artificial Intelligence Conference in July, announcing that their 128×128 matrix-scale photonic computing chip had achieved commercial standards for computational density and precision. Dr. Dong has since partnered with the company to further the commercialization of photonic computing in artificial intelligence, focusing on advancements in light sources, phase-changing materials, and silicon photonic interconnect architectures.
The increasing demand for computational power, driven by the growth of technologies like artificial intelligence, has exposed a mismatch between the capabilities of traditional electronic computing and AI’s computational requirements. This has fueled the search for new computational breakthroughs. Photonic computing, with its high degree of parallelism, energy efficiency, and speed, holds great potential in addressing this imbalance, particularly in the construction of large-scale matrix-matrix parallel computing systems.
In recent years, the field of photonic computing has seen a flurry of research and progress. The Oxford study, titled Partial coherence enhances parallelized photonic computing, demonstrates that reducing optical coherence can enhance photonic convolution processing. The researchers present a photonic convolution processing system that utilizes reduced temporal coherence (partially coherent systems) to boost parallel processing without significantly compromising accuracy. This approach eliminates the need for precise control of numerous phase shifters or micro-ring resonators and alleviates strict feedback control and thermal management requirements by employing partially coherent light sources.
The study’s findings were showcased in two separate photonic platforms for computational applications. In the first, a 3×3 photonic tensor core using phase change material-based photonic memory was employed for parallel convolution processing. The system classified gait patterns of 10 Parkinson’s patients with a remarkable 92.2% accuracy. In the second demonstration, a 9×3 silicon photonic tensor core with embedded electroabsorption modulators was used as a high-speed 0.108 TOPS convolution processor for vector encoding and weight setting. It was combined with on-chip photodetectors to classify the MNIST handwritten digit dataset with an accuracy of 92.4%.
Photonic computing, primarily carried out on chips, is poised to revolutionize the artificial intelligence landscape. Photonic chips, fabricated using mature CMOS electrical chip process nodes, have found applications in communication, sensing, and computing. The integration of photonic switches onto silicon photonics chips has led to significant reductions in size and power consumption in the telecommunications sector. Similarly, solid-state lidar devices are being replaced by silicon photonic chips in the sensing field for miniaturization and cost-effectiveness.
In the realm of computing, photonic chips, leveraging advancements from communication and sensing, offer a pathway to large-scale, energy-efficient computing clusters that align with the demands of the AI market. Designed for linear operations, which form the foundation of mainstream AI algorithms, photonic chips are a natural fit for tasks such as training and inference in large models, autonomous driving, and embodied intelligence. They promise lower hardware costs, both in terms of initial investment and operational expenses, while delivering exceptional energy efficiency compared to electronic chips.
As the demand for computational power continues to soar, photonic computing’s potential to provide massive computational capacity with minimal energy consumption makes it an attractive solution. The 92.2% accuracy achieved in gait analysis for Parkinson’s patients is a testament to the transformative potential of this technology, paving the way for more advanced applications in healthcare, AI, and beyond. With ongoing advancements in photonic computing, the future of high-performance, energy-efficient computing is on the horizon.
【source】https://www.jiqizhixin.com/articles/2024-08-28
Views: 3