The world's first all-optical multilayer neural network

[ Instrument Network Instrument Development ] Recently, researchers from the Hong Kong University of Science and Technology have developed the world's first all-optical neural network that can be used for deep machine learning. It not only allows artificial intelligence to deal with the relationship between identifying things or risk assessment. On the issue, closer to humans, it is also possible to calculate at the speed of light with a significant reduction in energy consumption. The research results were recently published in "Optics."

In order to master deep learning skills, artificial intelligence requires a multi-layer neural network with a nonlinear start function. However, in existing opto-electric hybrid neural networks, the nonlinear start-up function that mimics the response pattern of the human brain is mainly achieved by electricity, which limits the instruction cycle and capability of the optical network.

Now, the team led by Du Shengwang, a professor of physics at the Hong Kong University of Science and Technology, and Liu Junwei, an assistant professor, has developed the first all-optical multilayer neural network, which is a step toward building a large-scale optical neural network.

In order to break through the limitations, the research team realized the nonlinear activation function by using the electromagnetic wave that can operate with only a very low laser power in a cold atomic medium, and created a two-layer all-optical neural network. In order to test the results, the researchers used the network to classify the ordered and disordered phases in the condensed matter physics symplectic model and found that it is as accurate as the results of high-performance computer neural network operations.

Du Shengwang said: "Although the result is only a proof-of-concept test, it shows that a new generation of optical artificial intelligence, that is, fast calculations under low-energy conditions is possible."

In the future, researchers hope to expand the scale of this technology and build a larger, more complex all-optical neural network for practical applications such as image recognition.