【新智元导读】The top level of artificial intelligence ICML 2019 released the results of this year's paper admission. Of the 3,424 papers submitted, 774 were accepted, with an acceptance rate of 22.6%, which was lower than last year. From the number of accepted papers, Google became the biggest winner this year, followed by MIT and Berkeley.
ICML 2019 The results of the papers were released. Are you in the middle?
For a time, netizens were happy and broke, and they all said that they were in the middle of it! I am in the middle! ”
It is understood thatThis year, ICML submitted a total of 3424 papers, of which 774 were accepted, and the paper acceptance rate was 22.6%.. The acceptance rate is lower than the 25% of ICML 2018 last year.
At present, all accepted papers have been published on the official website. If you have a submitted reader, you can go to the link below to see if the paper has been accepted:
Google is the biggest winner, followed by MIT and Berkeley
Reddit netizens published statistics on the admission of this ICML 2019 paper by him and his company.
This year, Google has undoubtedly become the biggest winner in all accepted papers.
Ranking of the total number of accepted papers (by research institute)
The above table shows the ranking of the total number of papers accepted by the Institute (including industry and academia). At least one author in this statistic is affiliated with a research institute, so a paper can appear multiple times and belong to multiple research institutes.
among them,Blue represents the total number of papers, green and red represent the number of papers accepted by the first author and correspondent authors respectively.. Also, affiliates are manually incorporated into the Institute, such as Google Inc., Google AI, and Google UK, all mapped to Google.
It can be seen that the number of Google admissions papers far exceeds that of other research institutes, ranking first; followed by MIT, Berkeley, Google Brain, Stanford, Carnegie Mellon, and Microsoft.
The author also conducted a statistical Top 50 ranking based on academic and industrial circles, respectively.
Ranking statistics visualization address:
In the academic rankings, MIT, California Berkeley, Stanford and Carnegie Mellon won the top four, becoming the first echelon of the number of accepted papers, and opened a certain gap with the second echelon.
Ranking visual address:
In the Top 50 ranking of the Enterprise Research Institute, Google is undoubtedly the biggest winner: Google, Google Brain and Google DeepMind scored first, second and fourth respectively. Microsoft, Facebook and IBM also performed well, ranking third, fifth and sixth.
For domestic companies, Tencent scored better and ranked eighth.
In addition, from the current ICML 2019 admissions paper, the following statistics can also be obtained:
452 papers (58.4%) are purely academic research;
60 papers (7.8%) from industrial research institutions;
262 papers (33.9%) were authored by academia and industry.
Summarizing the above statistics, we can get the following results:
77% of contributions come from academia;
23% of the contribution comes from the industry.
The Berkeley researchers were admitted to 7 and 6 papers respectively!
According to the situation of the accepted papers this year, the ranking statistics based on the authors of the papers are also obtained.
According to the above pictureUnique author (yellow), first author (orange), correspondent (green), and total number of papers (blue)Statistics were made.
From the University of California at BerkeleyMichael JordanAdmitted by ICML 20197 articlesText, the results are quite amazing! Followed byVolkan Cevher (EPFL)withSergey Levine (University of California, Berkeley)Each being admitted6 articles.
Equally impressive is the fact that a large number of researchers have two or more unique authors or first author papers. In this ranking, Ashok Cutkosky from Google is far ahead. He has a single author paper, a first author paper and a general author paper.
The results of Simon Du (CMU) and Jayadev Archarya (Cornell) were also more eye-catching, and they published three first author papers.
According to the ranking of the correspondent authors, perhaps we can find more senior people.
Relative contribution ranking
The final ranking is based on relative contributions, ie how many authors in a paper actually came from a research institute.
How does ICML 2019 review papers?
When Xinzhiyuan made the ICML 2019 review results, it was reported by the netizens. Let us first look at the paper review mechanism of ICML 2019:
1. Double blind review
The ICML 2019 paper review uses a double-blind mode; that is, the reviewer does not know the identity of the author, and the author does not know the identity of the reviewer. Please quote your previous results as much as possible in a third person. Please do not add a thank you to the submitted paper and a link to the public github repository. If you need an anonymous reference, such as referring to the results you are reviewing elsewhere, please upload this result as a supplement. Anonymity is a mandatory requirement, and papers that reveal the identity of the author explicitly or implicitly may be rejected. Reviewers can use external resources to infer the identity of the author, such as technical reports posted online. This does not violate the double blind review policy.
2. About supplementary materials
This ICML conference supports the submission of two additional materials - supplementary papers and code/data, which can be used to provide additional evidence support. If the author makes an anonymous reference in the paper, be sure to upload the reference so that reviewers can quickly view it. Please be careful not to disclose the identity of the author in the supplementary materials.
Regarding code submission, we hope that the author can reasonably submit the code and organization in an anonymous form. The author name and license information are deleted. ICML 2019 allows code to be submitted via an anonymous github repository. However, these codes must be in a directory that cannot be modified after the submission deadline. Please enter the github link in a separate text file in the submitted zip file.
Please note that the submitted papers must be completely independent. The conference advocates reviewers to review the supplementary materials submitted by the authors, but the reviewers do not have this obligation, and we do not want to impose unnecessary burdens on the reviewers. Authors must not use auxiliary materials to extend the length of the paper. If the author believes that the material is critical to the evaluation of the paper, it needs to be included in the paper. Supplemental materials can be submitted in zip file or pdf format.
Supplementary materials are not published or archived. If you want to include it in the final version of the paper, you must place the material on the website and reference the website in the final version of the paper.
In order to find the right reviewer for the author to submit the paper, we will use the Toronto Paper Matching System to assist in screening.
As part of the submission process, authors are required to allow permission to use the “paper matching service” license, and the author of the paper submitted to ICML2019 should be allowed to use the service.
About multiple drafts
It is not appropriate to submit a paper that is the same (or substantially similar) that was previously published or received, or to submit the same paper to another meeting. Such behavior violates our multi-investment policy.
There are a few exceptions to this rule:
Allows a streamlined version of a paper that has been submitted to a journal but has not yet been published in the journal.
It is the responsibility of the author to ensure that relevant journals are allowed to submit relevant papers to the conference at the same time.
Papers in the following cases are allowed to be submitted: papers submitted or pending at a conference or seminar (such as ICML or NeurIPS seminars) without Proceeding, and only those papers in the abstract section have been published.
Papers submitted as technical reports (or preprints submitted in arXiv) are allowed to be submitted to this ICML. In this case, we recommend that the author not quote the report to maintain anonymity.
The author's reply to the review comments
Currently, between March 9 and March 14, the author can review the comments and post a reply.
The purpose of the author's response is to change the reviewer's judgment on the paper, with a maximum of 5,000 characters. Any author of the paper can enter/edit the author's response, and the response can be edited or restored until the deadline.
The review of the paper is double-blind. Do not include any information in the response that identifies the author/co-author. Do not include any URLs in your responses.
The author should make a judgment when preparing the response. There is no need to reply to each small question or improvement suggestion one by one. These responses should be seen as a good opportunity to solve problems, such as explaining the reviewer's uncertainty about the paper, the wrong hypothesis made by the reviewer, or the reviewer misunderstanding a part of the paper. . We recommend maintaining politeness and professionalism in the author's response.
ICML 2018: Acceptance rate 25%, review of two best papers
The Machine Learning Summit ICML 2018 was held in Stockholm, Sweden from July 10th to 15th. In 2018, ICML received 2,473 submissions, an increase of 47.6% from 1676 in 2017. A total of 621 papers were selected, with an acceptance rate of 25%.
The ICML 2018 Best Paper Award has two award-winning papers. The first authors are Anish Athalye from MIT and Nicholas Carlini and David Wagner from UC Berkely, and the other authors are all from UC Berkely.
Best Paper One:
Obfuscated Gradients Give a False Sense of Security: Circumventing Defenses to Adversarial Examples
Anish Athalye (MIT), Nicholas Carlini (UCB), David Wagner (UCB)
Thesis address: https://arxiv.org/pdf/1802.00420.pdf
Neural network-based classifiers are often used for image classification, and their levels are usually close to humans. However, these similar neural networks are particularly susceptible to anti-sample and subtle interference inputs. The picture below is a typical confrontation sample. After adding some invisible disturbances to the original image, the InceptionV3 classifier classifies the cat into a guacamole. According to a 2013 study by Szegedy et al., this “stupid image” can be synthesized using only the gradient descent method. This type of discovery sounds an alarm for object detection research.
In this paper, the authors evaluated nine papers received by ICLR 2018 and tested their robustness against the sample. The experimental results confirmed that in the eight papers on the defense mechanism against the sample, seven defense mechanisms could not withstand the new attack techniques proposed by the paper, and the defense level was limited.
Best Paper 2:
Delayed Impact of Fair Machine Learning
Lydia Liu, Sarah Dean, Esther Rolf, Max Simchowitz, Moritz Hardt (full UCB)
Paper address: https://arxiv.org/pdf/1803.04383
One of the fundamental concepts of machine learning is to use training to reduce errors, but such systems often produce discriminatory behavior due to sensitive characteristics such as race and gender. One of the reasons may be bias in the data. For example, in applications such as loans, recruitment, criminal justice, and advertising, machine learning systems can harm real-world vulnerable groups by learning historical biases that exist in the data. Therefore, it was criticized.
ICML 2019 will also be held on June 9th, and we can look forward to this year's best paper.
So, this year's ICML 2019, has your paper been accepted?