当前位置：网站首页>Has aaig really awakened its AI personality after reading the global June issue (Part 1)? Which segment of NLP has the most social value? Get new ideas and inspiration ~

Has aaig really awakened its AI personality after reading the global June issue (Part 1)? Which segment of NLP has the most social value? Get new ideas and inspiration ~

2022-06-23 13:17:00 【Zhiyuan community】

AAIG Look at the world -6 monthly ( On ) Release ！

AAIG Our experts use a multidisciplinary and professional perspective , Collect and screen global trends on algorithm governance and sustainable development from the world every month . Focus on new technologies in the computing age 、 New risks 、 New practice 、 And new explanations for them .

6 Monthly magazine ( On ) common 10 An article 3430 word , branch 【 Insight this month 、 Industry trends 、 Cutting edge technology 、 Industrial practice 】 Four plates spread out .

Artificial intelligence technology is widely and deeply applied to production and life , Especially for automated decision making , It has profoundly affected thousands of users . While users enjoy the convenience and bonus brought by artificial intelligence , It may also be due to the defect of automatic decision-making of artificial intelligence 、 Prejudice, etc . What are the user's feelings and comments on AI applications ？ Whether they know artificial intelligence , Trust artificial intelligence ？ What doubts and worries do users have about artificial intelligence ？ The provider of artificial intelligence services , It is necessary to explain to the user , Bridging the cognitive gap , Enhance consensus , Promote artificial intelligence to promote social development better and faster .

Click the title to view the full text

1. 《 User perception survey and analysis report of algorithm application （2021）》 Blockbuster released

In order to better understand the public's real feelings and comments on the current algorithm application , Promote large-scale Internet platforms to listen to public voices , Implement algorithmic ethical governance ; Research Center for digital economy and legal innovation, University of international business and economics 、 The digital economy research center of Renmin University of China and the ant group research institute were established in 2021 At the end of the year, we jointly carried out “ The algorithm applies user awareness ” Large scale empirical investigation , Received a valid questionnaire 6941 Share , On this basis, it has formed 《 User perception survey and analysis report of algorithm application （2021）》.

Reports are transparent around algorithms 、 Algorithm security 、 Algorithm fairness, algorithm goodness and other major algorithm problems , Carried out detailed research and analysis , The report found ：1、 The public is calling for increased algorithm transparency 2、 Balance algorithm security and use 3、 Pay more attention to the procedural fairness of the algorithm 4、 Let the algorithm better serve life 5、 Actively improve users' algorithm literacy .

2. Tsinghua huangminlie ： Google AI Is the personality really awakened ？

Professor Huang minlie analyzed from a professional point of view AI Is it possible to have personality , For humans , Is it “ threat ” still “ comfort ”？ He thinks that LaMDA Learned a lot of human dialogue data thus I've learned that “ Average ” Your personality . So-called “LaMDA Have personality ” It's just human language speaking style , Not at all LaMDA Spontaneously formed . Huangminlie thinks ,AI The focus of the next phase of the development of the dialogue system is “ More ethical 、 More moral 、 More secure ”.AI You must know what reply is safe , No risk , This requires AI Have ethics and correct values . He from AI The history and future development of dialogue system are explained in detail , He Said to AI for Social good Really looking forward to , Pay special attention to AI In social connection 、 psychology

1. NLP Which subdivision direction has the most social value ？

This article was published this year ACL Findings The collected papers put forward a set of proposals for NLP Evaluation index of social benefit of the task , From the perspective of maximizing social benefits, this paper gives priority to NLP Topics , The author first passes the classical theories and viewpoints of ethics , A qualitative method for evaluating good social benefits is given , secondly , Through the causal structure model, the existing NLP Technical classification , Thus, different kinds of NLP Different impacts of technology on social benefits . after , With the help of the analytical framework of global priority research areas , Propose a series of useful indicators to measure technology priority , Last , The author through to ACL 2020 570 The analysis and research of this thesis , This paper presents a set of thinking methods and methods for the research significance of the evaluation subject based on social benefits NLP Topics that should be given priority in the field .

2. AI Face changing 「Deepfakes」8 Annual evolutionary history

GAN A brief history of development ：1、2014 year Deepfake The first year of its birth .2、2015 year ：CNN Able to process large amounts of data in parallel , And the operation efficiency on the graphics card is particularly high .3、2016 year ：Deepfake Glasses and face processing .4、2017 year ： NVIDIA promotes quality leap , The first paragraph deepfake Video comes out .5、2018 year ：GAN Control is strengthened ,deepfake land YouTube channel .6、2019 year ：Deepfake Officially become the Lord flow .7、2020 year ：Deepfake Supervision and Disney megapixel deepfakes.8、2021 year ：Deepfake Tour 、 Live broadcast and face rental .9、2022 year ：3D GAN、DALL-E 2 With ZELINSKY deepfake.

3. Basic computer problems , A breakthrough has been made in the maximum flow problem ： New algorithm 「 It's ridiculously fast 」

A research team of computer scientists , This paper presents a very fast algorithm for the classical maximum flow problem in the computer field . The maximum flow problem is a combinatorial optimization problem , Discuss how to make full use of the device's capabilities , Maximize the flow of transportation to achieve the best results . at present , The new method is mainly theoretical improvement , Because the improvement of algorithm speed is only applicable to networks that are much larger than those encountered in the real world , For these networks , The maximum flow problem can be answered quickly （ At least without considering cost minimization ）.

4. Gato after , Google also launched 「 generalist 」 agent Multi-Game Decision Transformers

The initial steps for Google research to take a scalable approach , A versatile agent with superior performance . The researchers answered the following questions through experiments ：1、 How do different online and offline methods behave in multi game systems ？2、 How different methods compare to model dimensions ？3、 How effective are the different ways to switch to a new game ？4、 Multi game decision making Transformer Whether the training data has been improved ？5、 Whether best behavior reasoning improves behavior cloning ？6、 Whether the mixed training of expert and non expert data is more beneficial than only expert training ？7、 send use Transformer What are the benefits of Architecture ？

5. Zero sample learning surpasses CLIP！ Google proposes the first multimodal sparseness model LIMoE, It can also reduce computing costs

Google brings the latest results LIMoE, For the first time, the thinning method is applied to the image text hybrid model . The model based on this time is MoE, stay Transformer On the basis of Architecture , Added “ Experts ”. It is a parallel FNN, It replaces the original feedforward network . The specific term , Is to make LIMoE Contrast learning . Using a lot of images - Text to training , The image model inside the network extracts the image representation , Text model extracts text representation . For the same image - The text is right , The model will shorten the distance between the image and the text representation . conversely , For different images - The text is right , Will make the corresponding representations away from each other . The immediate benefits of this , Is to achieve zero sample learning .

1. On a par MIT、 Stanford , Alibaba ranks among the top ten computer network research institutions in the world

Authority AMiner published 2022 year AI 2000 Artificial intelligence influence list , Alibaba is the only Chinese organization on the list , It shows that its basic research field related to computer network has reached the level of world-class universities . In other words , In the basic field of computer network , Alibaba is an enterprise 、 Institutions of higher learning 、 All Chinese institutions in the scientific research institute , Have a place . It also shows that , Today's enterprise is not only the application field of Technology , It can also be the source of original theories and basic scientific research , And represents the technical and theoretical level of China . The continuous investment of enterprises in basic scientific research , It will help accelerate the improvement of China's overall scientific and technological strength .

2. Baidu heavyweight released the industry's first trusted system integrating trusted analysis and enhancement AI Toolset -TrustAI

Baidu released the first credible... Integrating trusted analysis and enhancement AI Toolset -TrustAI, help NLP Domain developers understand the prediction mechanism of neural network models , Build more trusted 、 More effective models , Help neural network model more secure 、 Reliable landing in practical application .TrustAI The function of ：1. Confidence analysis method ： Feature level evidence analysis , Case level evidence analysis .2. Trusted measurement data ： Natural language understanding task trust evaluation data , Pre training model evaluation data , Credible assessment indicators .3. Trusted enhancements .

3. 170 Million parameters ,28 Open test set SOTA, The largest visual multi task unified model in the industry is coming

WAVE SUMMIT 2022 In depth learning developer Summit , Baidu has released the largest visual multi task model in the industry VIMER-UFO 2.0, The model parameters reach 170 Billion , Single model 28 Item public dataset SOTA, Based on flying oars Task MoE framework , Automatically select and activate the optimal area according to different tasks , So as to achieve 100 Multiple parameter compression , At the same time, it supports the rapid expansion of downstream tasks .