For an unforgettable memory: special topic of Sun Jian

from ：AI Technology Review

“16 Years ago , Huangxutao goes to Hong Kong for Christmas , Inadvertently, I found that the leading figures of computer vision at the party were all different 10 year , Namely ：1936 Huang xutao, born in 、1946 Ma Songde, born in 、1956 Gao Wen, born in 、1966 Shenxiangyang, born in . We will discuss the next meeting , Whether we can find another young man 10 Year old computer vision leader , The discussion quickly points to the same person , That is shenxiangyang's student 、1976 Sun Jian, born in .”

2020 In the summer , Professor chenchangwen presented at the global conference on artificial intelligence and robotics （GAIR） “Thomas S. Huang Special commemorative session — Chinese computer vision 40 Years of inheritance ” link , Tells the story of “40 Years of inheritance ” The origin of .

As 40 The protagonist of the inheritance story in , Sun Jian also had three consecutive summers （2017-2019） Come to Leifeng GAIR The podium of the conference , With hundreds of slides , Without reservation, we will deliver sermons to thousands of future generations .

But just a week ago 6 month 14 Japan , Sun Jian died of illness , How can this not make people sigh .

1

Life and contribution

 

Dr. SunJian devoted his whole life to scientific research , He is a well-known scholar in the field of artificial intelligence, computer vision and deep learning . His research results have greatly promoted the development and application of artificial intelligence technology . Dr. Sun Jian's main research interests are computer vision and computational photography . since 2002 I have been in CVPR、ICCV、ECCV、SIGGRAPH、PAMI Published academic papers in five top academic conferences and journals 100 Yu Wen ,Google Scholar The number of references exceeds 28 Ten thousand times ,H-index 121, Twice obtained CVPR Best Paper Award(2009,2016).

2016 year 7 month , Dr. Sun Jian joined Kuangshi , Chief scientist 、 Dean of Kuang Shi Research Institute . Under the leadership of Dr. Sun Jian , Kuangshi Research Institute has developed from a small team of more than a dozen people to an industry-leading computer vision research institute , Developed efficient convolutional neural networks including mobile terminals ShuffleNet、 The open source deep learning framework looks at Tianyuan MegEngine、 AI Productivity platform Brain++ And many other innovative technologies , Leading edge AI applications . Dr. Sun Jian created the opening of the Research Institute 、 equality 、 The free culture and working atmosphere have established the mission of the Institute ： Continue to innovate and expand cognitive boundaries ; Extraordinary technology achieves product value .

 

 

2

Representative work ResNet

 

according to Google Scholar Statistics , The number of citations in Sun Jian's thesis is as high as 28 ten thousand 6 Thousands of times , Ranked first in terms of citations , It is “ Deep residual network （ResNet）”.ResNet It is the first deep neural network with more than 100 layers in the world , It is also one of the most important researches in the field of deep learning .

stay GAIR 2017 On , Sun Jian once introduced ResNet The background of the research ： Once people did not believe that deep learning networks could be trained , from 2012 year 8 Layer of AlexNet after , Everything has changed . But subject to various conditions , In this paradigm , The network 20 The performance of the left and right layers is the best ; until ResNet Appearance , Is to make the research of neural network take off , Entered a new stage .

All this is due to the fact that neural network research has always been “The Deeper,The Better”（ The deeper the better ） Belief ： The deepening of the network layers will improve the expression and abstraction ability of the network ; And on the other hand , The deepening of layers also greatly increases the number of parameters that can be adjusted , The greater the degree of freedom of adjustment . In other words , The increase in the number of layers of neural network can make up for the lack of network structure sophistication to a certain extent , It also makes “ Diao Shen Xia ” and “ Alchemist ” Our team is further expanded —— As Li Mu said , Suppose you're using a convolutional neural network , Half the chance is to use ResNet Or its variants .

But the increasing complexity of neural network also brings a series of problems , If the gradient disappears / Gradient explosion 、 Over fitting problem 、 Degradation problems, etc .ResNet It solves the problem of deep neural network training “ Degradation problem ”, It makes it possible to train deep neural networks .

If you make an analogy , Add a hidden layer to the neural network , Solved the famous “ Why can't a single-layer neural network fit XOR function ” Is a neural network from 0 To 1 Breakthrough ,AlexNet It's from 1 To 10 Breakthrough ,ResNet It is from 10 To 100 perhaps 1000, Even more breakthroughs .

Besides , SunJian's Faster RCNN、Spatial pyramid pooling、Single Image Haze Removal Using Dark Channel Prior、ShuffleNet And other representative studies have also made outstanding contributions in many fields of computer vision .

 

 

3

For the memory that has never been forgotten

 

We always believe in one word ： Anniversary , In order not to forget .

In the early hours of this morning CVPR 2022 At the opening ceremony , We are gratified to see that a short commemoration of Dr. Sun Jian was added to the conference . SunJian yes CVPR Community activists , And has twice obtained CVPR The best paper . Perhaps personal traces may be smoothed by time , But his wisdom will become a part of the world .

Confucius said ： A gentleman is not famous for his illness . It is famous in later generations , It is an inspiration for life ; In the vast river of history , Only by promoting historical progress 、 Those things that have left a mark on the world , Will really exist .

today , We launched Dr. Sun Jian's commemorative topic , Thinking about Dr. Sun Jian's life , Remembering Dr. SunJian's contribution , Thinking about life , Learn his academic thoughts , And carry it forward .

This commemorative topic is divided into four parts ：

• Life and contribution . We have made a brief summary in this article , If there is something wrong , Sorry for the Inconvenience. .

• Research interpretation . For countless AI researchers , Sun Jian is a good teacher and friend on their way to growth ; We selected Dr. Sun Jian's most representative research interpretation , I also hope that more people will be like what Sun Jian's student zhangxiangyu said in commemoration of Sun Jian , Take over the mantle , Self-motivated forward .

• Nostalgic article . We initially considered soliciting contributions from Sun Jian's relatives and friends , Memories and fragments of Dr. Sun Jian . But after comprehensive consideration , We decided to summarize the memorial articles written or published by our relatives and friends instead ;

• Speech arrangement . Including Sun Jian's research philosophy 、 Frontier research results and Thoughts on academic and industrialization .

As a writer , We believe in the power of words . Interpret Dr. Sun Jian's research 、 Memorial articles and speeches are summarized here , It is the best memorial we can think of for Dr. Sun Jian ; I also hope that the latecomers can compare themselves with this , Jian Xian , Siqi ; And follow the path pioneered by our predecessors , past , Open .