04

2024

-

08

Tsinghua University has made progress in the direction of pre-sensing optical calculations.

Author:


In the era of the Internet of Things, visual image sensors, as the key equipment of the intelligent society, are embedded in various devices such as mobile communication terminals, smart wearable devices, automobiles and industrial machines. With the continuous expansion of the application, the system power consumption, response speed, security and other aspects of the sensor also put forward higher requirements. In the traditional "sensing and computing" link, the access speed and communication bandwidth of memory have gradually become the main bottleneck limiting the power consumption and speed of the system. The calculation unit is close to the sensing unit, so that the near sensing end of the system has a certain data processing capacity, which has gradually become a powerful way to solve this problem. Compared with other proximity computing methods, pre-sensory computing has the advantages of high speed, high bandwidth and low power consumption. However, at present, most optical neural networks need coherent laser as the light source, the hardware system is bulky and complex, and can only complete linear operations, lack of interlayer nonlinear activation, limiting the application of pre-sensory optical computing in edge scenes.

Figure 1. Computing paradigm of near sensor in machine vision link

Professor Chen Hongwei's team from the Department of Electronic Engineering of Tsinghua University proposed a compact passive multilayer optical neural network (MONN) architecture, which consists of a passive mask and a quantum dot film to complete multilayer optical calculations with interlayer nonlinear activation under incoherent light illumination. The optical length of the architecture is as short as 5mm, which is 2 orders of magnitude smaller than existing lens-based optical neural networks. Experiments show that the multi-layer computing architecture is superior to linear single-layer computing in the performance of various visual tasks, and can transfer up to 95% of the calculations from the electrical domain to the optical domain. The architecture has the advantages of small size, low power consumption and high practicability, and is expected to be deployed in mobile visual scenes such as automatic driving, intelligent manufacturing and virtual reality in the future.

Figure 2. Multilayer presensory optical neural network architecture and interlayer nonlinear activation function measurement

At the same time, the absorption and emission spectra of CdSe quantum dots overlap in the wavelength range. Through the design, the absorption and excitation spectra of the front and rear quantum dots can be aligned, and the cascade and the expansion of the existing three-layer architecture to more layers can be realized. The pre-sensing MONN architecture can also be combined with other proximity computing paradigms to complete complex computing functions.

Recently, the relevant research results were published in Science Progress (Science Advances) under the title of "Presensory Computing Based on Compact Multilayer Optical Neural Networks" (Pre-sensor Computing with Compact Multi-layer Optical Neural Network). The Department of Electronic Engineering of Tsinghua University is the first unit of the paper, Chen Hongwei is the corresponding author of the paper, and Huang Zheng, a 2020 doctoral student in the Department of Electronics, is the first author of the paper. The research was supported by the National Natural Science Foundation of China and the Beijing Municipal Science and Technology Commission.

Chongqing Chongfan Technology Co., Ltd. Optical platform film thickness meter OCT optical coherence tomography LabVIEW control