Speech: Xie Yutao
Editor: Chen Ping
Are there any useful research tools in the FIELD of AI? Xie Yutao's team created AI tools to significantly improve the efficiency of AI research.
On March 23, at the Annual Conference on AI Technology of the Heart of Machines, Xie Yutao, Engineering Director of IDEA Research Institute and Head of AI Platform Technology Research Center, delivered a keynote speech entitled "If You Want to Do Good, You Must First Use It - Tools for AI Innovation".
Scan the code to look back
Video address: https://www.bilibili.com/video/BV1ki4y1k7xe?spm_id_from=333.999.0.0
The following is the content of Xie Yutao's speech at the Machine Heart AI Technology Annual Conference, and the Machine Heart has been edited and sorted out without changing the original meaning:
Hello everyone, I am Xie Yutao, from the Guangdong-Hong Kong-Macao Greater Bay Area Digital Economy Academy IDEA, thank you very much for the invitation of Machine Heart to participate in the Machine Heart AI Technology Annual Conference, I was a user of Machine Heart before, and it is a great honor to share with you as a guest today. The previous teachers shared wonderful content on the research and industrialization of AI, and I took this opportunity to share my thoughts on scientific research tools, and the title I shared was "If you want to do something good, you must first use it - a tool for AI innovation".
I mainly introduce three aspects: the first part is that we were born in the era of AI, which is a very exciting era; the second part is the desire to do good, in this part I will share some of our exploration of scientific research tools, I hope to be able to help you a little; the third part introduces some of our thinking on innovation in the field of AI research.
The AGE OF AI
The concept of artificial intelligence (AI) dates back to 1956, and during these 60 years, AI has experienced many ups and downs. We have completely lived in the era of artificial intelligence today, such as face recognition, fingerprint recognition, etc. in mobile phones are AI; during the epidemic, some people are isolated at home and it is inconvenient to go out, takeaway APP recommends favorite menus for you, which is also AI; how to allocate riders after placing an order, how to optimize the path, how to predict road conditions, etc., behind which there are shadows of artificial intelligence, the purpose is to let users quickly eat the food they want to eat. AI provides great convenience for our lives, and it can be said that it is everywhere. The previous teachers also shared that the current artificial intelligence technology has entered a full-scale commercialization stage, and has had different degrees of impact on various traditional industries and various participants, changing the ecology of various industries.
The four words of China's digital economy have appeared in the "Government Work Report" for the fifth time, which is a very important topic. In my opinion, digitalization is divided into two parts: one informatization and one intelligent. Shallow digitization is informatization, and deep digitization is intelligence.
Informatization in China has been quite mature, such as convenient and fast mobile phone payment, and there are no relevant convenient tools in other overseas countries, which China has been at the forefront of the world.
Intelligence can be said to be the right time, AI in the entire national economic development including industry, financial industry, etc. has played a great role, if there is no AI technology, the digital economy is only an information technology, it only brute force has no IQ, and this development is very limited. If there is no intelligent demand in the digital economy, then the industrial potential can not be tapped, for AI technology, researchers, experts and scholars can only stay in the ivory tower forever, technology can not land.
Research-led AI innovation
So the ERA of AI we are in is an era full of opportunities, and the progress of AI technology is the driving force of this era. Advances in AI technology are largely driven by researchers, for example, in the field of computer vision, we learned from the website paperswithcode that on the ImageNet1K image classification task, its Leaderboard Hero List has been refreshed every year, or even every month, since 2011, and almost every major breakthrough is based on researchers inventing new algorithms, training new models, having more data, and getting better results. We from the early SIFT, to AlexNet, to the later ResNet, etc., these models use different data, different algorithms, as shown in the following figure, each point in the figure is a good paper trying to hit a new high, whether it is a university, a research institution, or a major company, researchers from different angles to continuously carry out research, improve the capabilities of the entire AI. Great innovations across the AI space come from high-level AI research on a global scale.
So how is China doing with AI research? According to the data released by research institutions, the global proportion of China's papers in the field of artificial intelligence (below on the left) has increased from 4.26% in 1997 to 27.68% in 2017, far ahead of other countries. At the same time, the number of highly cited papers in China surpassed the United States to become the world's first in 2013. In addition, we can conclude from the LIST of AI 2000 scholars released by the AMiner team of Tsinghua University (the right in the figure below) that the distribution of scholars in the 20 subfields of artificial intelligence (person-time) is basically the United States and China from the distribution point of view, in addition to the multimedia and Internet of Things sub-fields China is slightly ahead, the United States is leading China in many fields.
In general, the current situation of AI scientific research in China is the first in the number of papers, the number of excellent AI scholars is second, the total number of scientific research talents is second, and the proportion of outstanding talents is relatively low. According to the data of 2017, China's artificial intelligence talents are about 18,000 people, accounting for 8.9% of the world's total, second only to the United States 13.9%, ranking second; in terms of enterprise talent investment, it is a high-intensity talent investment, basically concentrated in American enterprises, and only one company in China, Huawei, has entered the global top 20.
The third number is the high H factor, which is used to evaluate the academic impact of researchers, which is a very important indicator, this report counts the top 10% of high-H index scholars, 977 outstanding talents in China, about one-fifth of the United States, this gap is relatively large. We are now roughly ranked 6th in the world in terms of the H Index, with a low percentage of top talent.
Better tools make good work
We can't help but think, how to improve the research level of researchers? We have different solutions, I want to talk about the desire to do good things must first be used, do scientific research is best to have some easy to use scientific research tools, but the reality is that we do not have particularly easy to use tools, researchers use a lot of tools in their daily lives, such as reading papers, writing papers, publishing papers, reviewing papers, managing literature, meetings, etc. Lack of corresponding tools. When reading papers, most researchers use a more primitive method: print papers and take notes with a pen; the same is true for writing papers, although we have some paper editing tools, but they are not ideal; there are no useful tools for publishing papers, reviewing papers, managing literature, holding academic conferences, etc. All of the above are encountered by researchers every day, but today's tools are missing.
The second lack of communication platform "alone and no friends, then lonely and unheard", if we read the paper without someone to communicate with, it is a very painful thing, for example, we often encounter problems when reading papers, these problems or formulas, or inferences, etc., generally speaking, we do not have people who can consult at any time. In today's World, where the Internet is so developed, it is difficult to find a community focused on academia. If there is such a community for people to ask questions, discuss, and focus on academic content, it is a very good thing for many researchers, but today such an academic community is lacking.
The third academic achievement is not spread smoothly, there are now many search engines, we use more such as Baidu Academic, CNKI and other systems, which basically meet our needs for papers. But the problem is that there are too many papers, do not know what to read, where to start, for example, this year's CVPR 2022 included 2067 papers, this number is very large, imagine how it is possible to read all these papers, and even find out which papers need to be read is very difficult. Publishers also don't have particularly good tools to give readers better access to knowledge and more convenient dissemination of knowledge.
Build a thesis community: ReadPaper essay reading platform
Tools are needed to do scientific research, but tools are missing. I'm going to talk about some of the simple experiments we've made here, and last year we tried to build a paper reading community. We do this community mainly to solve three problems:
The first academic exchange is just need: everyone needs to communicate and can communicate in depth, but the academic community is missing. As a reader, you want to read a good paper, and if you have any questions, you can answer them; as a paper author, you want the article to be read by more people, and you want the reader to make suggestions for your article; the institution hopes to be able to find the research direction, whether the topic is set up in the right direction. But there is a lack of an active academic community at home and abroad.
The second paper itself is difficult to read, with millions of papers published around the world every year, but most are not easy to read, especially for researchers whose native language is not English.
The third community has insufficient influence on academic research, and the lack of academic community has led to insufficient community influence on academic research.
We have created a professional academic discussion community (url: readpaper.com), which was officially launched in November last year, with four major aspects: dissertation search, online notes, literature management, and academic discussion.
The first is the paper search, we collected about 200 million academic paper metadata in the back end, you can search in natural language, so that when you are looking for papers, speed reading papers, you can accurately and directly, and users can quickly browse relevant information.
The second intensive reading paper, users can enter the PDF file for detailed reading, can also take online notes, immersive reading experience, to achieve reading while noting.
The third literature collation, which is also a common problem encountered by many scholars and researchers, how to track, sort, classify too many literatures in an orderly manner, how to put the literature together for long-term tracking, and even share with you, the collation of papers is also what our community hopes to provide for everyone.
The fourth group reading thesis may be a study group assigned by the supervisor, or a group of people who do not know each other on the Internet because of the same paper spontaneously organized together, they comment on each other, discuss, and make progress together.
We start by reading papers to build a community of academic discussions, which is one of the things we want to do.
I'll briefly explain some of the features below, the first is the immersive reading experience, when we open the PDF to read the paper, we enter our paper super reader. We will use some natural language processing techniques to parse the PDF itself The table of contents is the table of contents of your literature presented to the reader for easy jumping. Citation resolution is the resolution of the index of the cited documents in the text. When we read papers, we often encounter citations such as (23) and (36), before we print the paper to read, when we encounter (23) citations, we will turn to the back to see the specific content of the citation, and then return to the place where we read before, which is very inconvenient. We have the citation parsing function, after the PDF is parsed, when you click (23) when you read, a box will pop up on the screen, the box shows who the author of the article is, what the title is, the time of publication of the article, etc., and even the abstract of the article can be displayed, so that you can stay where you are reading and focus on reading, this function is incomparable to the experience of printing on paper for reading, because you have no way to cite the abstract information of the paper. Our analysis of the list of cited documents will also be displayed in the super browser, you can also sort the list of citations, and we don't know which paper is a highly cited paper, but our reader can help you solve this problem.
Chart parsing is also, for example, sometimes the printed paper diagram is on page 8, the relevant annotation text is on page 10, click the parsed icon to fix the figure on the screen, and the picture and text are read in contrast, which is very convenient.
The third word translation, the vast majority of papers are currently in English, there are still obstacles for Chinese students and scholars, word translation is that the reader can define the word can be translated into Chinese, easy to read.
Text screenshots can be taken as notes, abstracts, and full-text searches, allowing readers to read papers faster and better.
ReadPaper Papers Ten Questions
Next, to show the classic ten questions, our Idea Founding Director Dr. Shen Xiangyang said "You are how you read". We can understand that the process of reading is how the content encoded by the author can be successfully decoded by the reader, and the process of information being decoded by the reader is the process of reading, so the decoding process is the process of forming its own cognitive model, so "You are how you read" How you read defines the cognitive model. According to years of experience in the research community and industry, Dr. Shen Xiangyang summarized the classic ten questions to help everyone read the paper, how to truly understand the paper by answering these questions, and reading the paper with questions can help readers decode the author's thoughts in a directional manner.
This function has been integrated in the readpaper.com, for example, the paper shown in the following figure has been answered ten times, and the person who answered the question has spent a lot of time reading the paper and then sharing it. By reading the answers to these ten questions, the reader can quickly understand the information of a paper, and if you want to read 50 CVPR papers a day, these ten questions should be of great help.
I just briefly introduced a simple attempt at our paper reading tool, we want to build a community for each article, because we believe that every article is excellent, there are creative ideas in the article, it deserves to have a community, we want to build a community for the article, so that there are no difficult papers in the world, so that everyone can read the paper more conveniently.
Because of the time reason, there are many functions I have not explained, such as the search, management, discussion of the paper, you can try to use it yourself, in addition, we released the PC client, the iPad client is about to be released, the system has a considerable number of users, the community is forming, we think that between the students and the thesis we do this little thing can help readers better read the paper, better learn the knowledge of the predecessors, and for the students who are about to graduate, how to collect information to make reports, Helping you write your own essay can help. At present, users still prefer our tools. Better tools make good work.
Rethink the academic ecology
How do we go from a big country in scientific research to a big country with outstanding talents in the era of AI? I hope that there are more sharp tools to help scientific talents. I am also thinking about whether the whole academic ecology is there something we can do, and whether it can provide better tools in various links, and such tools are helpful to everyone. We will do scholarship in the future will definitely deal with the entire ecology, the core of this ecology is a community, which has students, teachers, authors, readers, papers are a bridge to transmit knowledge, such communities and other parts of the ecology are inextricably linked, researchers in scientific research institutions are doing research topics, scientific research institutions have classes, read papers, write papers, researchers also have to publish papers, submissions, all kinds of review tools to be improved, participating in academic conferences is a complex process, From organizing academic conferences to follow-up, meetings, and exchanges, the optimization of every detail can help researchers improve the efficiency of communication, as well as how to obtain high-quality articles from publishers more effectively and make knowledge spread faster under the premise of protecting copyright. The commercialization of researchers' achievements is also a very important thing, such as the improvement of the ability of employees in the enterprise, how to transform scientific research technology from the core community to commercialization, the recruitment of talents, etc. Are things that we can optimize, and each step in the middle needs a good tool to help researchers, to enhance our research level, and improve the entire scientific research and academic ecosystem.
AI era of scientific research to take the lead, for the digital economy of the industrial intelligence continue to create new growth space, I think, the desire to do good, must first use its tools, we need to rethink the entire academic ecology, I also hope that we can create more tools to help the development of scientific research.