INFORMATION TECHNOLOGY

Scientists have developed an ultra-integrated optical convolution processor


Recently, the team of researcher Li Ming-Academician Zhu Ninghua of the Microwave Optoelectronics Research Group of the State Key Laboratory of Integrated Optoelectronics, Institute of Semiconductors, Chinese Academy of Sciences has developed an ultra-highly integrated optical convolution processor. The research results are published in Nature Communications under the title Compact optical convolution processing unit based on multimode interference.

Convolutional neural networks are artificial neural networks developed inspired by the biological visual nervous system. It consists of a multilayer convolutional layer, a pooling layer, and a fully connected layer. As the core component of the convolutional neural network, the convolutional layer extracts features of different levels and abstraction through local perception and weight sharing of input data. In a complete convolutional neural network, the amount of convolution operations usually accounts for more than 80% of the entire network operations. While convolutional neural networks have been successful in areas such as image recognition, they also face challenges. Traditional convolutional neural networks are mainly based on the electrical hardware implementation of von Neumann architecture, and the storage unit and processing unit are separate, resulting in an inherent contradiction between data exchange speed and energy consumption. With the increase of data volume and network complexity, it is increasingly difficult for electronic computing solutions to meet the demand for high-speed, low-energy computing hardware for real-time processing of massive data.

Optical computing is a technology that uses light waves as a carrier for information processing, which has the advantages of large bandwidth, low latency, and low power consumption, and provides a computing architecture of “transmission is computation, structure is function”, which is expected to avoid the data tidal transmission problem in the von Neumann computing paradigm. Optical computing has attracted much attention in recent years, but in most of the reported optical computing schemes, the number of optical components has shown a secondary growth trend with the scale of the computing matrix, which makes the scale expansion of optical computing chips face challenges.

The optical convolution processing unit proposed by Li Ming-Zhu Ninghua’s team constructs three 2×2 related real-valued convolution kernels through two 4×4 multimode interference couplers and four phase shifters (Figure 1). The team innovatively combines wavelength division multiplexing technology with multi-mode interference of light to characterize Kernel elements by wavelength, input-to-output mapping to realize the multiplication operation in convolution, wavelength division multiplexing and photoelectric conversion to realize the addition operation in convolution, and the correlation convolution kernel reconstruction by adjusting the four thermal modulated phase shifters (Figure 2).

The optical convolution processing unit experiment proposed by the team verifies the feature extraction and classification ability of handwritten digital images. The results show that the image feature extraction accuracy reaches 5 bit. Ten classes of handwritten digits from the MNIST Handwritten Digits Database were performed with 92.17% accuracy. Compared with other optical computing schemes, this scheme has the following advantages: (1) High computing density: combining optical wavelength division multiplexing technology with optical multi-mode interference technology, four control units are used to realize three 2×2 real-value kernel parallel operations, and the computing power density reaches 12.74-T MACs/s/mm2. (2) Linear scalability: The number of regulatory units increases linearly with the scale of the matrix, which has a strong potential for large-scale integration.

The research work was supported by the National Natural Science Foundation of China and the Youth Innovation Promotion Association of the Chinese Academy of Sciences. (Source: Institute of Semiconductors, Chinese Academy of Sciences)

Figure 1.Optical convolution processing unit

Figure 2.Results of image feature extraction using optical convolution processing units. (a) Pictures of five handwritten digits entered; (b) the results of feature extraction using a computer; (c) Results of characterization using the proposed optical convolution processing unit.

Related paper information:https://doi.org/10.1038/s41467-023-38786-x

Special statement: This article is reproduced only for the need to disseminate information, and does not mean to represent the views of this website or confirm the authenticity of its content; If other media, websites or individuals reprint and use from this website, they must retain the “source” indicated on this website and bear their own legal responsibilities such as copyright; If the author does not wish to be reprinted or contact the reprint fee, please contact us.



Source link

Related Articles

Leave a Reply

Your email address will not be published. Required fields are marked *

Back to top button