NeurIPS is an annual conference on artificial learning focused on neural systems. The event will be held in Vancouver, December 8-14. Florence d’Alché-Buc, Program Co-chair at NeurIPS 2019, answered our questions.
What happens once an article has been sent to the reviewers?
Three reviewers first provide a detailed assessment of the articles submitted. Afterwards, there is a period during which authors are able to access the assessment of their article and to respond to points raised by the reviewers. In case of disagreement, the reviewers exchange views to arrive at a consensus, under the guidance of the Area Chair. If disagreement persists, the article is dealt with at a higher level.
The review committee then sends its reports to the Area Chair, who discusses the acceptability of the article, and any difficult cases, with the Senior Area Chair. Additional reviewers may be involved if required. The Senior Area Chair then informs the program committee of their decision and of any difficult cases.
What is your role in the selection process and what are the key criteria for an article to be published?
I am one of the three Program Co-chairs selected by Hugo Larochelle at Google Brain, who is chairing the Scientific Program Committee this year. Our most important task is to decide, with the Senior Area Chairs, whether to accept or to reject articles in difficult or conflictual cases.
Selection criteria include the article’s relevance to the conference, its impact on the community, its clarity, its scientific quality in terms of technical soundness and its originality. High-quality articles can be rejected if one of these criteria is not met.
We also have a number of other duties such as schedule management, replying to a large number of questions via email and generally ensuring the scientific quality of the program.
What can we expect at the 2019 event?
This year, we have paid particular attention to the reproducibility of the work submitted. The field of artificial learning is growing extremely rapidly and producing a great number of methods. It is therefore essential that any new contribution conform to the usual standards in scientific disciplines.
This is why authors were required to complete a questionnaire on the article they submitted, to ensure that they were aware of potential flaws in it. They were asked questions on the algorithms they presented: the level of detail in their descriptions, the conditions for their use and the hyperparameters. The idea was to encourage authors to share as much information as possible on their work. The strategy has borne fruit, as 75% of the articles provide the code for their method.
Six articles submitted by the DSAIDIS Chair were accepted this year, you can read them on the event website: https://datascienceandai.wp.imt.fr/fr/agenda/561/