当前位置:网站首页>Zhiyuan community AI weekly 78: CVPR 2021 publishes best paper

Zhiyuan community AI weekly 78: CVPR 2021 publishes best paper

2021-06-23 23:58:22 Author: Lou

Reading guide

In order to help Chinese AI research 、 Practitioners can better understand the latest information in the field of global artificial intelligence , This week, the editorial team of Zhiyuan Research Institute compiled the second 78 period 《 Zhiyuan community AI weekly 》, From the academic ( Papers and new ideas 、 Academic conferences, etc ), Industry and policy ( Technology industry policy 、 Project fund application 、 Technology investment and financing, etc ), figure ( Personnel changes and awards of scholars )、 data ( Data sets ), Tools ( New tools and applications recommended ) And so on , A bird's-eye view of what's happening in AI in the past week .

In the past week (2021/06/14~2021/06/20), The following are noteworthy contents 3 aspect :

One 、 In recent days, ,CVPR 2021 Published the best paper 、 Best student thesis, etc . Researchers from Marple Institute and tibingen University in Germany won the best paper award , Researchers at Caltech and Northwestern University won the best student thesis Award . Besides ,FAIR Two Chinese scholars, including he Kaiming, were nominated for the best paper , And another Chinese scholar 、 Lin Shanchuan, a graduate student in the Department of computer science at the University of Washington, was nominated for the best student thesis .( See this week for details “ meeting ” The column )

Two 、 In recent days, ,ACM SIG The results of the new election , Chen Yiran 、 Liu Xue 、 Wang Wei 、 Wang xiaofeng 、May Dongmei Wang、Lili Qiu And so on , The term of office is 2021 year 7 month 1 solstice 2023 year 6 month 30 Japan .( See this week for details “ figure ” The column )

3、 ... and 、 The study was conducted by DeepMind Chief research scientist 、 Professor at University College London David Silver Leading , The research inspiration comes from their research on the evolution of natural intelligence and the latest achievements of artificial intelligence , At the time of writing the paper, it is still in the pre proof stage . The researchers believe that , Reward maximization and trial and error experience are enough to develop the ability to show behavior related to intelligence . thus , They came to the conclusion that , Reinforcement learning is a branch of artificial intelligence based on reward maximization , It can promote the development of general artificial intelligence .( See this week for details “ Point of view ” The column )

Here are the details of each point .

Paper recommendation

tsinghua 、 National People's Congress 、 Fudan, etc | Pre training model : In the past 、 Present and future

Pre-Trained Models: Past, Present and Future

We went deep into the history of pre training , In particular, it is related to transfer learning and self supervised learning , reveal Large scale pre training model in AI The key position in the development spectrum . Besides , We have a comprehensive review of the latest breakthroughs in large-scale pre training models . These breakthroughs are due to the surge of computing power and the increase of data availability , In four important directions : Design an effective Architecture 、 Take advantage of the rich context 、 Improve the efficiency of calculation and carry out interpretation and theoretical analysis . Last , We discuss a series of open problems and research directions of large-scale pre training model , We hope that our views can inspire and promote the future research of large-scale pre training model .

Check the details and download the paper

Transformer | THUNDR: Tag based Transformer Of 3D Human body reconstruction

THUNDR: Transformer-based 3D HUmaN Reconstruction with Markers

THUNDR By using 3d The mark indicates , The predictive ability of the model free output architecture is compared with that of the statistical human surface model ( Such as GHUM) Regularization of 、 A combination of anthropometric characteristics , We deduce a statistical model based on the whole body 3d Mannequins , Achieve end-to-end training . be based on Transformer The prediction channel can focus on the image area related to the task , Support self-monitoring mechanisms , And make sure the results are consistent with anthropometry .THUNDR The structural diagram of is as shown in the figure 1 Shown , Network runtime , Marker pose prediction channel based on constrained markers , The prediction channel automatically encodes the initially generated body grid consistent with anthropometry into a set of markers through the linear layer of features , Then markers are used to predict GHUM Parameters , Generate human posture grid .

Check the details and download the paper

Robust learning | Against visual robustness based on causal intervention

Adversarial Visual Robustness by Causal Intervention

at present , Confrontational training is recognized as the most promising means to defend against confrontational samples . However , Its passive characteristic inevitably makes its unknown attackers immune . In order to achieve active defense , In addition to the popular bounded threat model , We also need to have a more basic understanding of the opponent's example . In this paper , The author puts forward a causal perspective of confrontational security risks : Because there are many confounding factors in the learning process , The attacker can just take advantage of the hybrid effect . therefore , A fundamental solution to robustness is causal intervention . In general, confounding factors cannot be observed , The author suggests using instrumental variables to intervene , There was no need to observe confounding factors . This robust training method is called causal intervention based on instrumental variables (CiiV). It has a differentiable sampling layer and consistency loss , Stable and unaffected by gradient ambiguity .

Check the details and download the paper

GNN | Rotation invariant graph neural networks using spin convolution

Rotation Invariant Graph Neural Networks using Spin Convolutions

By effectively simulating the atomic system , Can significantly accelerate the energy breakthroughs needed to address climate change . Simulation technology based on first principles , For example, density functional theory (DFT), Because of the high cost of Computing , It is limited in practical application . Machine learning methods may be similar to DFT Efficient way of Computing , Thus, the impact of computational simulation on real-world problems can be significantly increased . The approximate DFT There are several challenges . These include accurate simulations of subtle changes in relative positions and angles between atoms , And enforcement constraints , For example, rotation invariance or conservation of energy . In this paper, a new method is introduced to model the angle information between adjacent atomic groups in graph neural network . By using the local coordinates of each side and the new spin convolution on the remaining degrees of freedom , The side message of network realizes rotation invariance . Two variants of the model are proposed for the application of structural relaxation and molecular dynamics . meanwhile , This paper is based on a large scale Open Catalyst 2020 State of the art results are shown on the data set . Also on the MD17 and QM9 Data sets are compared .

Check the details and download the paper

Medical image segmentation | Unsupervised domain adaptive optimal latent vector alignment in medical image segmentation

Optimal Latent Vector Alignment for Unsupervised Domain Adaptation in Medical Image Segmentation

In this paper , This paper discusses the problem of domain drift in medical image segmentation . In this paper, a self encoder based on variation is proposed (VAE) And optimal transmission (OT) The lightweight unsupervised domain adaptation method of the theory OLVA. Due to the use of VAE, The model proposed in this paper learns a shared cross domain potential space which obeys normal distribution , This reduces domain drift . To ensure effective segmentation , The author designed a shared potential space to model shape changes , It's not a change in intensity . The author of this paper further relies on OT Loss to match and align the remaining differences between two domains in space . The author in MM-WHS It is proved on the data set OLVA The effectiveness of segmentation for multiple heart structures , Where the source domain is annotated by 3D MR Image composition , The target domain consists of unlabeled 3D CT Image composition . The results show that , Compared with the current generative training method , The method proposed in this paper is simple Dice Improved scores 12.5%.

Check the details and download the paper

Point of view

DeepMind Chief research scientist 、 Professor at University College London David Silver: Reinforcement learning can promote general AI

The study was conducted by DeepMind Chief research scientist 、 Professor at University College London David Silver Leading , The research inspiration comes from their research on the evolution of natural intelligence and the latest achievements of artificial intelligence , At the time of writing the paper, it is still in the pre proof stage . The researchers believe that , Reward maximization and trial and error experience are enough to develop the ability to show behavior related to intelligence . thus , They came to the conclusion that , Reinforcement learning is a branch of artificial intelligence based on reward maximization , It can promote the development of general artificial intelligence .

Look at the original  |  Participate in discussion

Industry and Policy

Google's autopilot Division Waymo Recapture 25 Us $100 million financing

According to Reuters 6 month 16 It's reported that , Google's autopilot Division ——Waymo Get... In a new round of financing 25 Billion dollars investment . It is reported that , Google parent Alphabet Participated in this round of financing , Other investors include Andreessen Horowitz、Silver Lake and Tiger Global etc. . According to the investor website PitchBook The data of ,Waymo The latest valuation of is over 300 Billion dollars .Waymo Founded on 2009 year , It has become a leader in the field of completely driverless .Waymo Is the world's first commercial taxi Hailing service for passengers Waymo One, Already in Phoenix 、 San Francisco and the bay area are in operation .Waymo Express , This round of financing includes 10 Multiple investors , The financing will be used to advance the company's autonomous driving technology Waymo Driver, And used to expand Waymo The team .

Look at the original  |  Participate in discussion

figure

ACM SIG The list of new presidents is announced ! Chen Yiran 、 Liu Xue and six other Chinese scholars were elected

In recent days, ,ACM SIG The results of the new election , Chen Yiran 、 Liu Xue 、 Wang Wei 、 Wang xiaofeng 、May Dongmei Wang、Lili Qiu And so on , The term of office is 2021 year 7 month 1 solstice 2023 year 6 month 30 Japan .

Look at the original  |  Participate in discussion

Code

Texas A & M University 、 University of Texas at Austin | Self destructive comparative learning

Recent breakthrough learning through comparison accelerates the deployment of unsupervised training on real world data applications . However , In reality, unlabeled data is usually unbalanced and long tailed , And it's not clear what role the latest contrastive learning methods play in practice scenarios . This paper proposes a clear solution to this challenge , Through a comparative study called self destruction (SDCLR), Automatic balancing means learning without knowing the category . Our main inspiration comes from the phenomenon that recent models have hard to remember samples , And those may be exposed through network pruning . It is more natural to assume that the long tail sample is also due to the lack of examples , It makes it difficult for models to learn well . therefore ,SDCLR The key innovation is to create a dynamic self competition model in contrast to the goal model , This is a trimmed version of the latter . In training , Comparing the two models will lead to adaptive online mining of the most easily forgotten samples of the current target model , And implicitly emphasize that they are more in contrast to losses . A large number of experiments across multiple data sets and imbalance settings show that ,SDCLR Significantly improved overall accuracy and balance .

Look at the original  |  Participate in discussion

The university of Hong Kong | HR-NAS: Use lightweight Transformer Search for efficient high resolution neural architecture

High resolution representation (HR) For intensive forecasting tasks ( For example, segmentation 、 Detection and attitude estimation ) crucial . In the past, neural architecture search focused on image classification (NAS) In the method , Study HR The expression is usually ignored . This work presents a new NAS Method , be called HR-NAS, It can effectively encode multi-scale context information while maintaining high-resolution representation , Find effective and accurate networks for different tasks . stay HR-NAS in , We updated it NAS Search space and search strategy . In order to be in HR-NAS Better coding of multi-scale image context in search space , We first designed a lightweight Converter , Its computational complexity can be changed dynamically according to different objective functions and computational budgets . In order to maintain the high-resolution representation of the learning network ,HR-NAS Adopt multi branch architecture , suffer HRNet Inspired by the , Convolutional coding that provides multiple feature resolutions . Last , We propose an effective fine-grained search strategy to train HR-NAS, It effectively explores the search space , And find the best architecture given a variety of tasks and computing resources . HR-NAS It can be used in three intensive prediction tasks and one image classification task FLOP To achieve the most advanced trade-off between , As long as the budget is small . for example ,HR-NAS Beyond the design for semantic segmentation SqueezeNAS, At the same time, it improves 45.9% The efficiency of .

Look at the original  |  Participate in discussion

Well quickly |MlTr: Using converters for multi label classification

The task of multi label image classification is to recognize all the object labels in the form of images . Although it has been advancing for many years , But small objects 、 Similar objects and objects with high conditional probability are still the main bottlenecks . Previously based on Convolutional Neural Networks (CNN) Model of , Restricted by the representation ability of convolution kernel . Recent visual converter networks use self attention mechanism to extract pixel granularity features , Express more abundant local semantic information , Not enough to mine global spatial dependencies . In this paper , We point out three key issues based on CNN Methods encounter and explore specific possibilities for converter modules to solve them . We propose a multi label converter architecture (MlTr) Partition by window 、 Pixel attention in the window 、 Cross window attention building , Especially, the performance of multi label image classification task is improved . Proposed MlTr Shows state-of-the-art results on a variety of popular multi label datasets , for example MS-COCO、Pascal-VOC、NUSWIDE, Respectively 88.5%、95.8%、65.5%.

Look at the original  |  Participate in discussion

New tools

Netease is open source EMLL: High performance end side machine learning computing library

With the development of artificial intelligence technology , We have higher and higher requirements for computing performance . Most of the traditional computing processing is based on the cloud side , Put all the images 、 Audio and other data are transmitted to the cloud center through the network for processing, and then the results are fed back . But as the data grows exponentially , Relying on cloud computing has shown many shortcomings , For example, real-time data processing 、 Network constraints 、 Data security, etc , So end-to-end reasoning becomes more and more important . In this context , NetEase has a way AI The team independently designed and developed a high-performance end side machine learning computing library ——EMLL(Edge ML Library), And it has been open source recently .

Look at the original  |  Participate in discussion

Chinese University of Hong Kong | DistillFlow: A self supervised learning framework for optical flow estimation

We proposed DistillFlow, This is a knowledge distillation method for learning optical flow .DistillFlow Training multiple teacher models and student models , Among them, the challenging transformation is applied to the input of student model to generate illusion occlusion and less confident prediction . then , A self supervised learning framework is constructed : Confidence predictions from teacher models are used as annotations , To guide the student model for those less confident prediction learning optical flow . Self supervised learning framework enables us to learn optical flow effectively from unlabeled data , Not only for non occluded pixels , It also applies to occluded pixels . DistillFlow stay KITTI and Sintel State of the art unsupervised learning performance on datasets . Our self supervised pre training model also provides excellent initialization for supervised fine tuning , This shows that compared with the current supervised learning method which highly relies on synthetic data pre training , It's an alternative training paradigm . At the time of writing , Our fine tuning model is KITTI 2015 Ranked first of all the monocular methods for benchmarking , And in Sintel Final Performance in benchmarking is better than all published methods . what's more , We've shown... In three ways DistillFlow Generalization ability : Frame generalization 、 Corresponding generalization and cross dataset generalization .

Look at the original  |  Participate in discussion

University of Chinese Academy of Sciences | Uformer: General purpose for image restoration U shape Transformer

In this paper , We proposed Uformer, This is an effective and efficient method based on Transformer The architecture of , We use Transformer Block building layered encoders - Decoder network for image restoration . Uformer There are two core designs , Make it fit for the task . The first key element is the locally enhanced window Transformer block , We use non overlapping window based self attention to reduce computational requirements , The depth convolution is used in the feedforward network to further improve its capture potential . The second key element is that we've explored three ways to skip connections , To effectively transfer information from the encoder to the decoder . With the support of these two designs ,Uformer High ability to capture useful image recovery dependencies . A large number of experiments on multiple image restoration tasks have proved that Uformer The advantages of , Including image denoising 、 Go to 、 To blur and to dream . We hope that our work will encourage further research , In order to explore low-level visual tasks based on Transformer The architecture of .

Look at the original  |  Participate in discussion

application

Massachusetts institute of technology, | Learning protein language : evolution , Structure and function

In recent years, language model as a powerful machine learning method , It has been used to extract information from large-scale protein sequence databases . Only from the ready-made sequence data , These models can discover the evolution of the whole protein space 、 Structural and functional areas . Using language model , We can encode amino acid sequences as vector representations , To capture its structural and functional characteristics , And evaluate the evolutionary adaptability of the sequence mutants . This paper discusses the latest development of protein language models and their applications in the prediction of downstream protein properties . The author considered how to enrich these models with previous biological knowledge , And a method of coding protein structure knowledge into the characterization is introduced . The knowledge extracted from these models enables us to improve the downstream function prediction through transfer learning . The deep protein language model is revolutionizing protein biology , They offer new ways to design proteins and therapies . However , Coding powerful biological prior knowledge into protein language models , Further development is needed to enrich its application .

Look at the original  |  Participate in discussion

Indian Institute of technology | 2D Comparative semi supervised learning of medical image segmentation

Comparative learning (CL) It's a recent way of expressing learning , It achieves gratifying results by encouraging inter class separability and intra class compactness in learning image representation . Because medical images usually contain multiple categories of interest , So the standard image level of these images CL Do not apply . In this work , We propose a novel semi supervised 2D Medical image segmentation solution , The solution will CL Apply to image blocks , Not the whole picture . these patches It is constructed meaningfully using semantic information of different classes obtained through pseudo tags . We also propose a novel uniform regularization scheme , It works with contrastive learning . It solves the problem of confirmation deviation often observed in semi supervised setup , And encourage better clustering in feature space . We evaluated our approach on four public medical segmented datasets and the new histopathological datasets we introduced . Our method consistently improves the most advanced semi supervised segmentation methods for all datasets .

Look at the original  |  Participate in discussion

New developments in machine learning are published again Nature cover : Solve the problem of medical data privacy

It is reported that , Recently, new progress in machine learning has been published in international academic journals once again 《 natural 》(Nature) cover . Researchers from the University of Bonn in Germany, together with Hewlett Packard and from Greece 、 Germany 、 A number of research institutions in the Netherlands have jointly developed a combination of edge computing 、 Distributed machine learning method for peer-to-peer network coordination based on blockchain —— Group learning (Swarm Learning, hereinafter referred to as SL), For data integration between different medical institutions . The researchers based on 1.64 Ten thousand copies of blood transcriptome and 9.5 Ten thousand breasts X X-ray image data , Use SL For leukemia 、 Tuberculosis and lung disease 、COVID-19 Developing disease detection classifiers , Find out SL It is superior to the classifier developed by a single medical institution while meeting the confidentiality standard . The accuracy of the algorithm to identify the sick individuals , In the blood transcriptome data set, the average is 90%, stay X The X-ray image data set shows that 76%-86%.

Look at the original  |  Participate in discussion

meeting

CVPR 2021 Awards come out : The best papers go to Mapu , He Kaiming was nominated , The first Huang xutao Memorial Award was issued

In recent days, ,CVPR 2021 Published the best paper 、 Best student thesis, etc . Researchers from Marple Institute and tibingen University in Germany won the best paper award , Researchers at Caltech and Northwestern University won the best student thesis Award . Besides ,FAIR Two Chinese scholars, including he Kaiming, were nominated for the best paper , And another Chinese scholar 、 Lin Shanchuan, a graduate student in the Department of computer science at the University of Washington (Shanchuan Lin) Got the best student thesis nomination .

Look at the original  |  Participate in discussion

Above is 《 Zhiyuan community AI weekly 》 The first 78 The content of the issue , The editorial team of Zhiyuan Research Institute will be based on “ Provide real experts AI information ” The goal of , Constantly optimize and improve our content services , If you have any criticism , Or good advice , Please point out in the comments section below . Thank you. .

版权声明
本文为[Author: Lou]所创,转载请带上原文链接,感谢
https://chowdera.com/2021/06/20210623235709142K.html

随机推荐