Shulou(Shulou.com)06/02 Report--

This year's 33rd Neural Information processing Systems Conference (NeurIPS 2019) seems to be a little more violent than in previous years.

First of all, there is a sharp increase in the number of participants. The number of participants in 2018 reached about 9000, which has made the organizers happy from ear to ear. Unexpectedly, 13000 people came directly this year, an increase of 40% over last year, and the number of applicants for the lucky draw reached 15000, which paralyzed the NeurIPS website for a time.

Secondly, there is a bumper harvest of the thesis. The number of papers contributed reached 6743, an increase of nearly 40% compared with 4856 last year; 1429 papers were included, an increase of 41% over last year, and the proportion of papers included was about 21%. Both the number of contributions and the collection of papers have set a new historical record. In contrast, ICML, which is also the top meeting of AI, only contributed 3424 papers, and finally collected 774 papers, which were only about half of those of NeurIPS.

During the conference, casually search for keywords about NeurIPS on social networking sites, and it can be seen everywhere that a paper has become the biggest dream of relevant professional students.

So, if anyone again broadcasts the nonsense of AI winter this year, and nothing else, as long as throw away the NeurIPS data, they will be able to slap in the face.

Of course, the bright light of all aspects of the data is only one aspect of the conference, what we pay more attention to is, from the results of this conference, can we see some changes in the players and flag direction of AI? So, today, let's abandon the cold theoretical self-research and complex formulas to see what changes and no changes exist in this Congress than in the past, and what is the development trend of AI in the real world?

The same is the player: thousands of years of running water, iron Google

I believe that needless to say, readers can guess that this year's NeurIPS has been dominated by Google again.

It is not unusual for Google to dominate the list at various AI academic meetings. Compared with 137 papers last year, the number of papers included by Google this year has reached 180, an increase of nearly 31%, and the growth rate is very stable, which also greatly increases the difficulty for other institutions to catch up. What's more, there is a Deepmind with dazzling skills regardless of cost, Google is probably going to continue to enjoy the loneliness of being king for a long time.

Microsoft is probably the only one that can wrestle with Google in industry, and it has done well this year, making it into 76 articles.

The strength of the top three academics in the United States is still good. MIT was selected for 86, followed by Stanford with 85, which is a great improvement from 11 last year, which also reflects Stanford's efforts to improve in the field of artificial intelligence this year. CMC was shortlisted for 79, worthy of being among the top three in the United States.

In addition, Berkeley, facebook, Princeton, IBM and other six players have more than 40 papers selected. American universities and companies accounted for eight of the top 10 seats, leaving ninth and tenth to Cambridge and Oxford in the UK.

This also tells us that when faced with various voices and analyses that are pessimistic about the United States, and when the mentality is a little drifting away, we still have to look at these research data: the scientific research strength of the United States is alone, and this is a reality that it is difficult to make major changes in the coming decades. Only by recognizing the reality can we catch up step by step.

Of course, China's performance this year is also very eye-catching. Tsinghua, the king of engineering, was selected in 33 articles, the overall ranking came to 13th, Peking University was selected in 23 articles, and Chinese University of Science and Technology, Zhejiang University, Chinese University of Hong Kong, Fudan and other universities also made achievements. In terms of enterprises, Tencent, Ali, Baidu and Huawei have also performed well. Unfortunately, compared with one of the best papers picked by Huawei's Ark Lab last year, this year's best paper has lost its shadow in China. Of course, the best paper strength is on the one hand, luck is also on the other hand, only our overall strength has been improved, the best paper is also a natural thing. Therefore, it is once again expected that the Chinese Legion can jump again in terms of the quantity and quality of papers in future meetings.

In addition, NeurIPS official data also mentioned that the overall proportion of female authors submitting this paper has reached 13%, and the female group is increasingly participating in the transformation of AI and the world as an indispensable and important force.

What has changed is the standard: there are more "hydrology" and strict manuscripts review.

The meeting is becoming more and more popular, and more and more contributions are not necessarily a good thing. This year, the NeurIPS committee gave 20000 papers alone, and the reviewers were so busy that they were once criticized for being "too lame" this summer.

In fact, it is totally understandable. After all, with the sharp increase in the number of contributions and recruitment of NeurIPS2019 papers, it is inevitable that a large number of "hydrology" will appear, and there are not a few people who want to muddle through by changing the model and adding a few data. In this case, some new changes have taken place in this year's paper review compared with previous years.

For example, the judging committee encourages the creativity of paper research. There is a special emphasis on looking at the problem in a creative new way, and under the premise of this kind of research, we can come up with a result that can really surprise the reader. This conveys the direction of readability of the paper. Uniform convergence may be unable to explain generalization in deep learning, which won the Outstanding New Direction thesis Award, was jointly completed by Vaishnavh Nagarajan and J. Zico Kolter of Carnegie Mellon University and Bosch artificial Intelligence Center, showing some exaggerated problems in the achievements of some deep learning researchers. In other words, it mercilessly uncovers the false cloak of some researchers' research results, so that readers and even ordinary people can maintain a more objective understanding of the effects of in-depth learning.

For example, the paper is required not to emphasize the importance or significance of the theory as much as possible, but to pay more attention to its realistic pertinence. In other words, there should be a certain point of view. To a certain extent, this is consistent with the more profound reality of the industrialization practice of artificial intelligence.

At the same time, the judging committee made it clear that there were not three kinds of papers: the first is the "resource-intensive" paper research, that is, the results obtained by using a lot of resources; the second is that the conclusion can be reached in a quick and effective way, but chose the tortuous "sideways", or "dazzling skills". The third is to emphasize the simplification of the paper, that is, if we can express it succinctly, we should shorten the length as much as possible. throughout the excellent papers of this session, many of them are no more than 10 pages long, compared with dozens of pages of "hydrology" that people often read. You don't have to be too conscientious.

In short, there are all kinds of birds in Linzi, and the sharp increase in the number of NeurIPS papers this year has caused the review committee to control the quality and revise the review standard, which is also a reasonable thing, and it is wise and inevitable to maintain the authority of the papers included in the conference.

The trend is application: the more popular the algorithm is, the more the practice becomes the target.

On the whole, the achievements of the papers in the three fields of algorithm, deep learning and application occupy the top three, which has not changed much compared with previous years. However, there are some subtle changes in the three areas themselves, from which we may also see a little trend in today's AI research.

The obvious change should be the research of algorithm. The proportion of papers submitted in the field of algorithm research increased from nearly 25% in 2018 to about 27% in 2018, and the admission ratio was about 21% in 2018 and about 28% in 2019. It is an area where both the number of papers submitted and final admissions have increased in all research fields.

In contrast, the proportion of papers on deep learning, whether submitted or finally admitted, shows a certain downward trend, although this trend is not obvious. In the field of applied research, under the condition that the proportion of papers submitted is basically unchanged, the admission rate has decreased by about 2 points, which may have something to do with the fact that hydrology will inevitably become a demon under the large total amount.

On the other hand, in the aspects of reinforcement learning, theory, probability and statistics, and neural network, most of the paper acceptance rates show signs of decline.

This trend may not be difficult to explain. With the continuous acceleration of artificial intelligence industrialization, the worldwide landing of AI is in full swing, and it has already moved from pure theoretical research in the laboratory to industrial practice, so it is bound to face more practical problems. Therefore, continuous innovative research at the algorithm level is obviously in line with the overall trend of the current AI.

Under the trend of deepening AI industrialization in 2019, the practice direction of next year's NeurIPS paper may be more obvious.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.