Sun Jian

A computer-vision scientist

Sun Jian, who also published as Jian Sun, was a computer-vision scientist. He died aged 45 after a sudden illness: Megvii announced that he passed away at midnight on 14 June 2022^[3], and Caixin reported the death on 16 June 2022^[2]. Caixin described him as a prominent scientist in computer vision and computer graphics, and as managing director of Megvii Research^[2].

The two fields are close relatives. Computer vision is the branch of AI that tries to make machines interpret images and video — recognising what is in a picture, locating it, recovering structure from pixels. Computational photography uses computation to improve how a camera captures and processes a picture, for example by cleaning up noise in dim light or merging several shots into one. Both rest on the same machinery: large collections of images, and neural networks trained on them so that the network picks up the regularities of the data itself.

Thirteen years at Microsoft

Sun spent 13 years at Microsoft researching computer vision and computational photography^[1]. Recode China AI records that he was a Principal Research Manager there from 2015 to 2016^[3]. He held 35 patents, most of which were awarded while he was at Microsoft^[1].

A long run in an industrial research lab is a common shape for a computer-vision career. Such labs give researchers computing power, data and engineering support that a university group rarely has, and they measure output partly in published papers and partly in patents — legal records of an invention that the applicant believes is new and useful. Sun's record shows both kinds of output, and the patent count is the more institutional of the two^[1].

Co-creating ResNet

The arXiv listing for “Deep Residual Learning for Image Recognition”, submitted on 10 December 2015, names Kaiming He, Xiangyu Zhang, Shaoqing Ren and Jian Sun as its authors^[4]; Recode China AI describes Sun as a co-creator of ResNet^[3].

An image-recognition network is built as a long stack of layers, each transforming the representation handed to it by the layer below. In principle adding layers should never hurt: extra layers could simply pass their input through unchanged. In practice, very deep stacks proved hard to train. Residual learning changes what each block is asked to learn. Rather than learning the desired mapping directly, a block learns the difference — the residual — between its input and the output it should produce, and the input is added back to the block's output through a shortcut connection. The network can then start from something close to doing nothing and improve from there, which makes optimisation easier and lets much deeper networks be trained.

At Megvii

In 2016 Sun left Microsoft to join Megvii as chief scientist and managing director of research^[1][2]. There he oversaw the creation of ShuffleNet, a neural network optimized to run on mobile devices^[1], and managed the development of Brain++, Megvii's core AI productivity platform^[1].

Both projects point at the same practical problem. A phone has far less computing power, memory and battery than a server, so a model that is accurate but heavy is not usable on it; ShuffleNet was designed for that constraint^[1]. Brain++ sits at a different layer of the stack: a company that builds many models needs shared tooling for producing and running them, which is what an AI productivity platform provides^[1].

What he left behind

Sun's career ran through two kinds of institution^[1][3]. At Microsoft he belonged to the corporate-lab tradition in which fundamental vision research is done, and the ResNet paper carries his name as a co-author^[4][3]. At Megvii he moved to a company whose business is building and shipping AI systems, and his work there was organised around models light enough for a phone and platforms for making them in quantity^[1]. He was 45 when he died^[2].