Review of AI hotspots this week: the Gan compression method consumes less than 1/9 of the computing power, and the open source generator turns your photos into hand drawn photos

                       

​01​ This open source animation generator makes your photos become hand-painted in seconds

​ Although recently 2019 The Turing prize in was awarded to computer graphics 、 To Pixar 3D Animation , But many people may think that two-dimensional animation is more interesting . Like Hayao Miyazaki 、 New Haicheng these masters hand-painted animation , It's the soul , Every piece can be a wallpaper , And the whole day diffuse also takes two dimensions as the core .​

​ If there is a model, it can transform the real picture into a hand-painted picture in the style of sun man , That must be cool . Recently, the heart of machines found that there are indeed these models , from CartoonGAN To AnimeGAN Can generate very interesting images .​

​ Pictures originate from ： The effect picture generated by the heart of the machine according to the real shop photos ​

​ Although many of the best examples given by the original project are street views , But I found that all kinds of scenes were ok , The following is a trial of the original image and the generated effect . Take a look at the generation effect of the first Cherry Blossom path , Suddenly there was a kind of 《 Spirited away 》 The feeling of .​

​ Project address ：​ ​​​​​​ Source of information ：​ ​ Almost Human ​​​

​02​15 Industrial level algorithms were launched 、35 A high-precision pre training model goes online ！

2020 year ,“ New infrastructure ” It is bringing new and significant opportunities to China's scientific and technological development , The AI infrastructure is facing a comprehensive upgrade . The in-depth learning framework is an important infrastructure to promote the advancement of industrial intelligence . In recent days, , Baidu PaddlePaddle has achieved a major upgrade in the field of intelligent vision .​ this ,PaddleCV The latest panorama is exposed for the first time . among ,PaddleDetection、PaddleSeg、PaddleSlim and Paddle Lite Heavyweight upgrade ; New release 3D Visual and PLSC Very large scale classification 2 Ability . meanwhile ,PaddleCV Added 15 Algorithms widely used in industrial practice , The overall number of high-quality algorithms has reached 73 individual ;35 A high-precision pre training model , Total number reached 203 individual .PaddleCV Panorama

​PaddleCV Relying on the core technology at the bottom of the propeller and the integration of software and hardware of Baidu brain AI Advantages of mass production platform , Through the core technology 、 Ecological Applications , And then to the whole system of commercial solutions , Support Baidu vision to become the largest in the industry 、 The technology stack is the most comprehensive 、 The most perfect visual technology platform in the ecosystem , Form a closed loop of self sustainable iterative optimization .​ As shown in the panorama ,PaddleCV Mainly from three aspects to update the core technical capabilities ：​​PaddleDetection Overall improvement of module type and performance ,YOLOv3 Greatly enhanced , Accuracy improvement 4.3%, Train faster 40%, Speed up reasoning 21%; Face detection model BlazeFace newly added NAS edition , Volume compression 3 times , Reasoning speed up 122%; newly added IoU Loss function type , Improve the accuracy 1%, Do not increase forecast time . In terms of models , newly added 3 A type of , be based on COCO Data sets are the most accurate open source models CBNet, the height is 53.3%;Libra-RCNN Model accuracy is improved 2%;Open Images V5 Become the best single model of target detection competition .

When the target detection model is actually deployed , Because of time and memory consumption , There are still big challenges . Based on this ,PaddleSlim Provides a variety of efficient model compression methods , boost PaddleDetection Performance reaches new heights . The validation accuracy can be improved by using the distillation model compression scheme 2%; The clipping model compression scheme greatly reduces FLOPs; Distillation + Clipping model compression scheme , be based on COCO Data set for testing , Can speed up 2.3 times . Besides ,PaddleDetection It also provides developers with an end-to-end process from training to deployment , And provide a cross platform image detection model C++ Forecast the deployment plan .

First ,PaddleCV be based on Paddle Lite Advantages of multi hardware support capability , Deep joint optimization with Kunlun chip , Realize the complete leadership and independent control of the end-to-end software and hardware integration capability . Take manufacturing as an example , Baidu and Weiyi intelligent manufacturing have jointly created intelligent automatic monitoring equipment “ Visual inspection equipment for surface defects ”, It is different from the traditional manual visual inspection of electronic parts , It can ensure the quality and efficiency of quality inspection , It has also further alleviated the problem of lack of manpower caused by the epidemic .

​ Source of information ：​ ​ Flying propeller PaddlePaddle​​

​03​ Han song 、 Zhujunyan et al GAN Compression method ： It takes less than 1/9, Now open source

​ Generate models GAN It is one of the most important development directions in the field of machine learning . But this kind of algorithm needs a huge amount of computing power , It's hard for most researchers to come up with new results . In recent years , This direction is quite monopolized by large institutions .​

​ But in recent days, , From MIT （MIT）、Adobe、 Researchers at Shanghai Jiaotong University have proposed a compression condition GAN The general method of . This new technology keeps visual fidelity at the same time , take pix2pix,CycleGAN and GauGAN And so on GAN The calculation of the model is reduced to 1/9~1/21. This method is suitable for many generator architectures 、 Learning goals , Paired or unmatched settings .​

​ At present, the research paper has been CVPR 2020 Included in the conference .​

​ In what the researchers showed Demo in , Use CycleGAN Adding zebra stripes to the horses in the video doesn't take much computing power 1/16, Three times more frames , And the effect has been improved . The hardware platform used by the Institute is the edge of NVIDIA AI Computing chip Jetson Xavier GPU. According to the official data ,Jetson Xavier Of INT8 The calculation force is 22+10TOPS, Xiao dragon 865 It is 15TOPS. Compressed GAN Now it seems that it can run in the robot 、 Drones and other small devices , The future is just around the corner .​

​ Project links ：​ ​​​​​

​ Source of information ：​ ​ Almost Human ​​

​04​ Physics breaks through the bottleneck of deep learning theory ？Google- Stanford explains the mechanism of deep learning

​3 month 16 Japan , Researchers from Intel Research Institute and Cornell University in the United States are in 《 natural - Machine intelligence 》（Nature Machine Intelligence） A paper was jointly published in the magazine , Shows intel Loihi Neural mimicry studies the ability of the chip to smell and recognize danger . For details, please see ：​ ​1 Billion neurons , Intel releases neuromorphic computing system ​​​
Although deep learning has revolutionized many applications , But the theoretical mechanism behind it has not been uniformly explained . Recently, scholars from Google brain and Stanford jointly Annual Review of Condensed Matter Physics Published a review paper on in-depth study of statistical mechanics 《Statistical Mechanics of Deep Learning》, common 30 page pdf, This paper expounds the relationship between deep learning and various physics and mathematics topics from the perspective of physics .​

​ Although deep neural network has achieved amazing success in the field of machine learning , But this raises deep questions about the theoretical principles behind their success . The author reviews recent work , Among them, physical analysis methods rooted in statistical mechanics have begun to provide conceptual insights into these problems . These insights have created links between deep learning and various physical and mathematical topics , Including random landscape 、 Rotating glass 、 interfere 、 Dynamic phase transition 、 chaos 、 Riemann geometry 、 Random matrix theory 、 Free Probability and nonequilibrium statistical mechanics . in fact , Statistical mechanics and machine learning have long enjoyed a rich history of strong coupling interaction , The latest progress in the cross field of statistical mechanics and deep learning shows that , These interactions will only deepen .​

​ Address of thesis ：​

 ​ ​​ ​https://www.annualreviews.org/doi/abs/10.1146/annurev-conmatphys-031119-050745​​​​

​ Source of information ：​ ​ Machine learning research group ​​​

​05​ Paper recommendation this week

【CVPR2020】 A new feature learning method for long tail data ： Construct feature clouds for tail samples

​​ author ：​Jialun Liu, Yifan Sun, Chuchu Han, Zhaopeng Dou, Wenhui Li​

​​ Paper introduction ：​​

​ Feature learning under long tail data is particularly difficult . Solve the problem that there are few tail category samples 、 The fundamental approach to the lack of intra class diversity is to increase the number of samples . So how to increase ？ In this work, the author proposes to construct clouds for tail categories 、 use “ Characteristic cloud ” To enrich the methods of the tail class . Different from previous work , Others are using generative confrontation networks （GAN）, Generate additional for the tail category “ Fictitious ” sample , Made better progress . However , Do you really need to add in the original image space ？GAN The training and calling of are very resource consuming , Is there any more practical 、 An efficient way ？ The author's work gives a more ingenious 、 Efficient solutions .​

​ The author's method directly focuses on the feature space , In the learned feature space , For the tail ID Add some virtual samples . These new virtual samples are like electron clouds winding around atomic nuclei , Wrap around the actual sample , Therefore, it is vividly called “feature cloud”. whatever cloud The essence of is a probability model , How to create this “feature cloud” The probability model of , Here it is raised to the head ID Study , Make the tail ID keep “ Short to admit , Stand firm after being hit ” Low attitude , Since the number of samples is small , This leads to insufficient intra class diversity , Then with a modest attitude to the head ID Study , Put the head ID Within class diversity of transfer To the tail ID. therefore , The author's method can be intuitively described as —— Fill the thin feature space of tail category with feature cloud , Just like an electron cloud filled with empty atoms .​

​​ Address of thesis ​：​

​ ​

 ​​

 ​​

​END​