28

2024

-

06

Light | Lifelong Learning Architecture for Intelligent Light Computing

Author:


Reading Guidance

Artificial intelligence tasks are increasingly complex, diverse and dynamic, traditional machine learning methods are often based on fixed data sets training tasks, can not adapt to environmental changes, how to make AI systems have the ability to multi-task continuous learning has become an urgent problem to be solved. In recent years, with the stagnation of Moore's law, the optical architecture represented by diffractive neural network as a new computational mode shows great potential in terms of function and efficiency. However, the problem of "catastrophic forgetting" (catastrophic forgetting) is still prevalent in current optical networks. Learning a new task will lose the memory of the previous task.

Inspired by the neurosynaptic mechanism of the human brain to realize sustainable learning, based on the sparseness and parallelism of light natural physics, Fang Lu's research group of the Department of Electronic Engineering of Tsinghua University pioneered the intelligent light computing lifelong learning architecture, breaking through the single function limitation of the optical network and supporting brain-like parallel multi-task learning. Through key modules such as sparse optical connectivity and multispectral parallel computing, optical computing systems are endowed with functionality, throughput and energy efficiency beyond existing technologies, providing a light-speed solution for multitasking deployment of intelligent systems on edge hardware.

The research results were published online in Light: Science & Applications under the title "Photonic neuromorphic architecture for tens-of-task lifelong learning. Dr. Cheng Yuan, Department of Electronics, Tsinghua University, is the first author of the paper, and Associate Professor Fang Lu is the corresponding author of the paper. This project is supported by the National Natural Science Foundation of China, the Basic Science Center Project, the "New Generation Artificial Intelligence" major project 2030 by the Ministry of Science and Technology, and the Tsinghua University-Zhijiang Laboratory Joint Research Center.

Industry background and technical challenges

Most of the current artificial intelligence models can only handle a given task. When the environment changes, there is nothing they can do, which is far from the general intelligence that people need. The sustainable learning ability of intelligent systems is an important bottleneck in the development of AI field, which is attracting widespread attention. On the other hand, in the process of continuous growth, the human brain can learn and master various skills for life. How to learn from the lifelong learning mechanism of the human brain and explore the computational paradigm of large-scale sustainable evolution has become a very challenging research direction. However, as the computing power and power consumption of traditional silicon-based electronic computing chips are severely limited, it is difficult to meet the needs of large-scale high-throughput data processing urgently needed in reality, so exploring new computing technologies and modalities is an inevitable trend for the development of future intelligent systems.

In recent years, the construction of photoelectric neural network based on optical computing and photoelectric fusion has become a hot research in the international forefront. However, although various optical networks represented by diffractive neural networks continue to emerge, the basic optical computing units in the existing optoelectronic systems are generally limited by the solidified structure and low scalability, which greatly limits the physical potential and computing power of the light itself. As a result, current optical networks can only achieve simple single tasks, and the problem of catastrophic forgetting cannot be overcome when learning new tasks.

L2ONN: Sparse Reconfigurable Optical Computing Supporting Lifelong Learning Intelligent Systems

In response to the above problems, the Fang Lu research group of the Department of Electronic Engineering of Tsinghua University proposed the lifelong learning optical computing architecture L2ONN(Lifelong Learning Optical Neural Network)(as shown in Figure 1), which innovatively realizes a multi-task intelligent system for sustainable learning. Benefiting from the sparsity and parallelism inherent in large-scale optical connectivity, optical computing naturally mimics the lifelong learning mechanisms of neurons and synapses in the human brain. Each task is learned by adaptively activating sparse optical connections in the coherent optical field, while gradually activating optical calculations to continuously obtain empirical information for various tasks. Multi-task optical features are processed in parallel by multi-spectral representations assigned with different wavelengths. The proposed optical lifelong learning architecture theoretically ensures the scalability and versatility brought by the physical characteristics of light itself, constructs a new neuromorphic computing system, and endows the intelligent system with the ability to calculate at the speed of light.

 

In order to evaluate the sustainable learning ability of L2ONN, the project team conducted simulation verification on several challenging data sets (as shown in Figure 2). A large number of experiments have confirmed that L2ONN avoids the catastrophic forgetting problem of ordinary optical neural networks and completes multi-task lifelong learning on a variety of data (visual classification, speech recognition, medical diagnosis, etc.), and its learning ability is more than 14 times higher than that of ordinary optical neural networks.

 

Further, the team actually manufactured an optical lifelong learning chip based on the diffraction calculation model (as shown in Figure 3), and used it to achieve continuous learning and parallel processing, completing the actual deployment of optical intelligent systems in edge hardware. The final results show that the chip uses limited computational resources and the computational energy efficiency is more than an order of magnitude higher than typical electrical neural networks.

 

prospects and outlook

The optical lifelong learning architecture proposed in this paper shows the extraordinary learning ability of optical computing. With the development of silicon optical chip integration technology, the optical chip based on the intelligent optical computing lifelong learning architecture effectively supports multi-task parallel processing, and provides a sustainable learning light speed solution for the actual deployment of intelligent systems in edge hardware. With its advantages in function, throughput and energy efficiency, the optical computing architecture is expected to significantly improve the scalability and versatility of artificial intelligence systems, enabling machine intelligence to adapt to complex tasks in the real world like humans.

Paper Information

  Cheng, Y., Zhang, J., Zhou, T. et al. Photonic neuromorphic architecture for tens-of-task lifelong learning. Light Sci Appl 13, 56 (2024).