News Release

Optical neural networks: A powerful boost for optical computing

Peer-Reviewed Publication

Light Publishing Center, Changchun Institute of Optics, Fine Mechanics And Physics, CAS

Fig. 2. Timeline of optical neural networks (ONNs) and related optical implementations.

image: 

Selected partial key milestones and publications are displayed to retrospect the developments of ONNs.

view more 

Credit: by Tingzhao Fu et al.

The development and challenges of optical neural networks

Research on ONNs began as early as the 1960s. To clearly illustrate the development history of ONNs, this review presents the evolution of related research work chronologically at the beginning of the article, spanning from the 1960s to the present, as shown in Figure 2.

 

This review categorizes ONNs into two major types: free-space non-integrated ONNs and on-chip integrated ONNs. Further, it subdivides ONNs into seven types based on different optical components in free-space and on-chip integration, as shown in Figure 3. The research work on these different types of ONNs is interconnected. For example, ONNs designed based on the 4f system typically have their computing units designed on the mask at the focal plane, which is limited. In contrast, diffractive optical neural networks (DONNs) in free space allow computing units to be freely designed at each hidden layer, resulting in a large scale of computing units and high computing capacity, thus effectively addressing the difficulty of further enhancing the computing capacity of ONNs designed based on the 4f system. Additionally, the introduction of tunable optical components such as spatial light modulators (SLMs), digital micromirror devices (DMDs), and information metasurfaces provides free-space DONNs with reconfigurable functionality, enabling different architectures of ONNs with various functions. However, the alignment process between the input layer, hidden layers, and output layer in free-space DONNs introduces additional system errors, with error accumulation leading to significant performance degradation. Moreover, due to the discrete diffractive components in free-space DONNs, their integration, stability, and portability need improvement. Therefore, research on on-chip integrated DONNs has been proposed to better address the issues of low integration, stability, and portability in free-space DONNs. However, since the computing units of on-chip integrated DONNs consist of sub-wavelength structures, programming these units is very challenging, making it difficult to achieve reconfigurable functionality. Hence, the reconfigurability problem of on-chip integrated DONNs remains to be solved.

 

Additionally, ONNs designed with other free-space optical components also face issues such as low integration, stability, and portability. To address these problems, researchers have used on-chip integration methods. For instance, ONNs designed with on-chip integrated optical components like Mach–Zehnder interferometers (MZIs) and micro-ring resonators (MRRs) have improved integration, stability, and portability. However, these on-chip integrated optical components require continuous energy consumption during the operation of ONNs, and when computing units are scaled up, thermal crosstalk between adjacent computing units becomes an inevitable issue, compromising the performance of on-chip integrated ONNs. To address the energy consumption and thermal crosstalk issues, researchers have introduced phase change materials (PCM). Due to PCM's non-volatile and programmable nature, using PCM to design on-chip computing units can support all-optical inference processes and provide reconfigurability in on-chip integrated ONNs. In addition, PCM enables ONNs to achieve nonlinear functions in the optical domain, further enhancing their inference capabilities. Although PCM offers many benefits to on-chip integrated ONNs, the current implementation scale for programmable and nonlinear functions is very limited. In the future, as the computing units in on-chip integrated ONNs are scaled up, the insertion loss of PCM must be considered, as it directly affects the feasibility of the ONNs system. This review analyzes and discusses the strengths and weaknesses of various types of ONNs at relevant points in the text and provides an outlook on their future development directions.

 

Although the development of ONNs still faces many challenges, significant technical success has already been achieved. Academic research is also moving towards practical applications. For example, a research team at Princeton University has applied on-chip integrated ONNs for nonlinear compensation in submarine optical fiber links, and another research team at the University of Cambridge has developed an edge-computing architecture based on photonic deep learning, achieving computational efficiency surpassing that of electronic computing hardware. In addition to academic efforts, the industry, with companies like Lightmatter and Lightelligence, is also exploring practical applications of optical computing.

 

Future development trends of optical neural networks

Research on using light for signal processing or computation can be traced back to the 1960s. However, the rapid development of ONNs has only prevailed in recent years. This is due to the increasing application of artificial intelligence and the growing demand for computing power across all trades and professions. Modern electronic computing hardware will be unable to meet the computing needs of societal development. There is an urgent need to find a new computing paradigm to provide stronger computational support. ONNs represent such a new paradigm, offering many advantages over modern computing hardware. However, the current operation of ONNs still relies on the assistance of electronic circuits and achieving nonlinear functions in the optical domain is very challenging. Therefore, in the short term, a hybrid optoelectronic ONNs system may be a feasible solution for practical applications. Figure 4 shows the architecture of a hybrid optoelectronic ONNs system, which includes the principle layer, the optical computing layer, the electronic circuit assistance layer, and the application layer. In this system, ONNs handle the majority of computational tasks, while electronic computing hardware is used for less computationally intensive tasks, such as routing, storage, and nonlinear functions. It is important to note that in a hybrid optoelectronic ONNs system, the energy consumption and speed of the optoelectronic conversion are critical to the overall system performance. Future research should aim to minimize the energy consumption of each optical-to-electrical (or electrical-to-optical) conversion and maximize the conversion speed.

 

Applications and prospects of optical neural networks

In fact, the application of ONNs in real-world scenarios is not yet mature, with most research still at the laboratory stage. According to the summary and analysis of ONNs-related research since the 1960s presented in this review, it is clear that ONNs currently face technical bottlenecks in areas such as storage, nonlinearity, and large-scale reconfigurability. Therefore, the application of ONNs will be concentrated in certain specialized fields in the short term, as shown in Figure 5. Even so, it may take a considerable amount of time to continuously optimize ONNs system architecture or hybrid optoelectronic frameworks to achieve better performance, enabling ONNs to outperform electronic computing hardware in certain specialized fields. During this period, both academia and industry should comprehensively consider the construction of the ONNs ecosystem, including aspects such as software, hardware, protocols, optical algorithms, industry standards, manufacturing techniques, and application scenarios.


Disclaimer: AAAS and EurekAlert! are not responsible for the accuracy of news releases posted to EurekAlert! by contributing institutions or for the use of any information through the EurekAlert system.