According to the official news of Tsinghua University, the research group of Professor Fang Lu of the Department of Electronic Engineering of Tsinghua University and the research group of Academician Dai Qionghai of the Department of Automation have found a new way, pioneered the full forward intelligent optical computing training architecture, and developed the "Taiji-II" optical training chip, which realizes the efficient and accurate training of large-scale neural networks of optical computing systems.
The research results were published online in the journal Nature on the evening of August 7, Beijing time, titled "Full Forward Training of Optical Neural Networks".
The following is the original reprint of the official public account of Tsinghua University:
The Department of Electronics of Tsinghua University is the first unit of the paper, Professor Fang Lu and Professor Dai Qionghai are the corresponding authors of the paper, Xue Zhiwei, a doctoral student of the Department of Electronics of Tsinghua University, and Zhou Tianzhen, a postdoctoral fellow, are the joint authors, and Xu Zhihao, a doctoral student of the Department of Electronics, and Dr. Yu Shaoliang of Zhijiang Laboratory participated in this work. This project is supported by the Ministry of Science and Technology of China, the National Natural Science Foundation of China, the Beijing National Research Center for Information Science and Technology, and the Tsinghua University-Zhijiang Laboratory Joint Research Center.
In their review, the Nature reviewers noted that "the ideas presented in this paper are very novel, and the training process of such optical neural networks (ONNs) is unprecedented. The proposed method is not only effective, but also easy to implement. As a result, it is expected to be a widely adopted tool for training optical neural networks and other optical computing systems. ”
Clever use of symmetry helps optical computing get rid of GPU dependence
In recent years, intelligent optical computing with high computing power and low power consumption has gradually stepped onto the stage of computing power development. The advent of the general-purpose intelligent optical computing chip "Taiji" is a microcosm of this, which for the first time pushed optical computing from principle verification to large-scale experimental applications, and brought the dawn to the "reasoning" of large-scale complex tasks with a system-level energy efficiency of 160TOPS/W, but failed to release the "training energy" of intelligent optical computing.
Compared with model inference, model training requires more large-scale computing power. However, existing optical neural network training relies heavily on GPUs for offline modeling and requires precise alignment of the physical system. Because of this, the scale of optical training is greatly limited, and the advantages of optical high-performance computing seem to be imprisoned in invisible shackles.
At this time, Fang Lu and Dai Qionghai's research group found the key to "photon propagation symmetry", and the forward and backward propagation in neural network training are equivalent to the forward propagation of light.
According to Xue Zhiwei, the first author of the paper and a doctoral student in the Department of Electronics, under the Taiji-II architecture, the backpropagation in gradient descent is transformed into forward propagation of the optical system, and the training of the optical neural network can be achieved by using data-error two forward propagation. The natural alignment of the two forward propagations ensures the accurate calculation of the physical gradient. The high accuracy of the training achieved in this way can support large-scale network training.
Since there is no need for backpropagation, the Taiji-II architecture no longer relies on electrical computing for offline modeling and training, and the accurate and efficient optical training of large-scale neural networks has finally been realized.
Full forward intelligent optical computing training architecture
Efficient and accurate, intelligent light training can do everything
Using light as the computing medium and the controllable propagation of light to construct a computing model, optical computing naturally has the characteristics of high speed and low power consumption, and the full forward propagation of light to achieve training can greatly improve the speed and energy efficiency of optical network training.
The paper study shows that Tai Chi-II is able to train on a variety of different optical systems and exhibits excellent performance in a variety of tasks.
Large-scale learning: The contradiction between computing accuracy and efficiency is broken, the training speed of optical networks with millions of parameters is increased by 1 order of magnitude, and the accuracy of representative intelligent classification tasks is increased by 40%.
Intelligent imaging of complex scenes: In low-light environments (where the light intensity per pixel is only subphotons), all-optical processing with an energy efficiency of 5.40×10^6 TOPS/W is realized, and the system-level energy efficiency is improved by 6 orders of magnitude. Intelligent imaging with kilohertz frame rate is realized in non-line-of-sight scenarios, and the efficiency is increased by 2 orders of magnitude.
In the field of topological photonics, it can automatically search for non-Hermitian singularities without relying on any model prior, which provides a new idea for efficient and accurate analysis of complex topological systems.
Universal intelligent light training empowers complex systems
Join hands with Taiji to promote the voyage of AI optical computing power
The launch of Tai Chi-II, following the Tai Chi I chip, further reveals the huge potential of intelligent optical computing.
For example, the two instruments are separated, and Taiji I and II achieve energy-efficient AI inference and training, respectively.
Another example is the harmony of the two instruments, and the Taiji I and II together constitute the complete life cycle of large-scale intelligent computing.
Fang Lu said: "We describe the dialectical collaboration architecture of the Taiji series as 'the way of Tai Chi and the way of combining the positive and the negative', and we believe that they will work together to inject new impetus into the development of computing power into the future AI model and build a new foundation for optical computing power."
On the basis of principle samples, the research team is actively moving towards the industrialization of intelligent optical chips, and has carried out application deployment on a variety of end-side intelligent systems.
It is foreseeable that through the unremitting efforts in the field of optical computing, including the Taiji series, the intelligent optical computing platform is expected to open up a new path for high-speed and energy-efficient computing of large artificial intelligence models, general artificial intelligence, and complex intelligent systems with lower resource consumption and smaller marginal cost.