The collaborative research work of Prof. Xingjun Wang's team and Weiwen Zou's team at Shanghai Jiao Tong University, Optical coherent dot-product chip for sophisticated deep learning regression, has been accepted for publication in the journal Light :Science& Applications. Based on a silicon-based integration scheme, this work proposes an optical dot product kernel chip that can be time-multiplexed, which can be reconfigured to enable flexible computation of matrix multiplication and convolution. A high-quality MRI medical image reconstruction task was performed using the AUTOMAP model (the best available network model for image reconstruction), and the quality of the reconstruction was close to what is desirable for a 32-bit computer.This work has broad implications for the further development of the field of optical neural networks as well as for new application areas.

The Peking University team is responsible for the design and flow of this optoelectronic chip. The chip breaks through the key technology of coherent modulation of arrayed optical devices and successfully realizes real-domain computation. The numerical accuracy of optical computation has been dramatically improved with the help of a novel on-chip feedback control algorithm. Compared to previous work, the chip's breakthroughs in number field completeness and numerical accuracy give it the ability to perform complex intelligent tasks.

Link to the paper: https://www.nature.com/articles/s41377-021-00666-8