First ICLR held in Scottsdale with open peer review

The gap it was meant to fill

When ICLR was set up, researchers working on representation learning had no common venue of their own. The stated goal of the conference was to fill that absence, treating representation learning as a single field spanning deep learning, metric learning, kernel learning and non-convex optimization ^[2]. Representation learning is the part of machine learning in which a model learns the features it uses from data, instead of having a person design those features by hand. Metric learning studies how similarity or distance between examples should be measured. Kernel learning is about the similarity functions that many classical methods depend on. Non-convex optimization is the mathematics of training objectives that have many local minima, which is the normal situation for deep networks. Putting all of these under one roof was itself a claim about how the field was developing.

A conference built around open review

ICLR was founded by Yann LeCun and Yoshua Bengio in 2012 ^[1]. Its most distinctive feature is its publication process, built on open peer review following models proposed by LeCun ^[1]. In the usual arrangement, reviewers are anonymous and their comments are read only by the authors and the program committee. In an open process, the reviews and the discussion they provoke are public, so a reader can see what other researchers made of a paper at the moment it was judged, and the reviews stay attached to the decision instead of disappearing after it. For the first edition, paper decisions were made available on the OpenReview website ^[2], which put the outcome of reviewing on the record next to the papers themselves. The intent was not only to judge work but to leave a trace of how the judgment was reached.

The first meeting

The first ICLR was a three-day event held from May 2 to May 4, 2013, in Scottsdale, Arizona, co-located with AISTATS2013 ^[2]. Co-location meant that attendees of the two meetings could move between their sessions without travelling in between. The two founders, Yoshua Bengio and Yann LeCun, served as general chairs of the first conference ^[2], the role that oversees the meeting as a whole rather than a single track of papers.

Papers and the link to a journal

The first edition also built a path from conference to journal. Journal versions of the conference's oral papers were accepted for a special topic of JMLR, with the deadline set at June 29 ^[2]. JMLR, the Journal of Machine Learning Research, is a long-running journal in the field; a special topic collects work around one subject. The arrangement gave the strongest papers two lives: a talk at the meeting, and then an extended written version in a journal. It connected a fast, public review cycle with slower archival publication, so that the paper a reader cites and the talk a researcher heard are recognisably the same work.

What the first edition established

The first edition therefore fixed the elements the conference is known for. It had two named founders, LeCun and Bengio, who also chaired it ^[1][2]. It had a scope deliberately drawn across several subfields rather than one ^[2]. It had a review process whose decisions were published on OpenReview ^[2]. And it had a route for its oral papers into a journal special topic ^[2]. A venue is more than a date and a city: what a meeting does with its reviews, and where its best papers go afterwards, shapes what people are willing to send it. The first ICLR answered both questions in public, and that is what the meeting has been recognised for since.