The Thirty-second Annual Conference on Neural Information Processing Systems
Highlights of what we saw at this year’s conference – Part 1/4
Apart from the absence of coffee mugs and the last-minute name change, NeurIPS 2018 in Montreal was most noted for its 11’28” ticket sell-out time. There were 4854 submissions (a 50% increase on 2017 and roughly double the submissions to ICML 2018) resulting in 1010 accepted papers (a 49% increase on 2017). The conference had three parallel tracks (vs. two last year) so it was a lot easier to choose what to do than at ICML 2018 which had about 10 parallel tracks. Poster sessions were in the day instead of the evening, which is a welcome change for presenters who’ve travelled West!
Apart from the absence of coffee mugs and the last-minute name change, NeurIPS 2018 in Montreal was most noted for its 11’28” ticket sell-out time. There were 4854 submissions (a 50% increase on 2017 and roughly double the submissions to ICML 2018) resulting in 1010 accepted papers (a 49% increase on 2017). The conference had three parallel tracks (vs. two last year) so it was a lot easier to choose what to do than at ICML 2018 which had about 10 parallel tracks. Poster sessions were in the day instead of the evening, which is a welcome change for presenters who’ve travelled West!
EXPO day
This year, NeurIPS added an “EXPO” to the programme which gave sponsors the opportunity to showcase some of their work in the form of either a demo, a workshop, a talk or a panel on Sunday, the day before the start of the conference. The content of these events was subject to a submission and selection process and the technical level was in general very high. Naver was present through a demo (by us) and a workshop by the AI CLOVA team. Jackie C. K. Cheung kicked off the afternoon workshop showing ways of including external knowledge (“common-sense”) inside an NLP architecture, either by using definitions or a web-scale information retrieval engine.
The rest of the presentations were about research at Naver including the work on Scene Text Recognition, the problem of detecting text in an image. Seong Joon Oh showed a 4-stage framework which generalises many previously proposed approaches. He discussed how this allows to compare those approaches with respect to accuracy, time and memory and to detect “blind spots”: images that are hard for any of the combinations. Those include for instance calligraphic text and vertical text. CLOVA recently released a tech demo web site to show what they’re doing in the AI space. During the workshop Julien Perez and Matthias Gallé from our European LABS presented our research in machine comprehension and core NLP.
In parallel to the Clova session, Facebook was showcasing the release of PyTorch 1.0 and its tracing capability which promises to bridge the gap between dynamic and static graph auto-differentiation paradigms.
In another room, a panel organised by Element AI about “Governance of AI” included Eimear Farrell, technology advisor and representative from Amnesty International and Human Rights Commissioner, Ed Santow. One of the points the panellists agreed on is that human rights is a good framework to address the potential threats of AI, and that there are other risks beyond privacy, the human right that is most visibly attacked these days.
The main conference hadn’t yet begun but, based on the EXPO and, despite the Montreal rain, the event was already looking promising with a large panel of industrial and academic contributions.
Julien Perez and Quentin Grail at the NAVER LABS Europe expo stand.
Highlights of what we saw at this year’s conference – Part 2/4
Highlights of what we saw at this year’s conference – Part 3/4
Highlights of what we saw at this year’s conference – Part 4/4