International Conference on Learning Representations (ICLR)

What ICLR is

The International Conference on Learning Representations (ICLR) is a machine learning conference, typically held in late April or early May each year^[1]. Along with NeurIPS and ICML, it is one of the three primary conferences of highest impact and reputation in machine learning and artificial intelligence research^[1].

Like other conferences in the field, it is a place where researchers present new work, argue about methods, and meet the people whose papers they have been reading. What distinguishes ICLR is both its subject matter — how data gets turned into useful internal representations — and the way it decides which papers to accept.

Origins

ICLR was founded by Yann LeCun and Yoshua Bengio in 2012^[1]. The first edition was a 3-day event from May 2nd to May 4th, 2013, co-located with AISTATS2013 in Scottsdale, Arizona^[2].

The stated goal of that first conference was to fill the absence of a common venue for researchers in representation learning^[2]. Representation learning is the branch of machine learning concerned with learning the features themselves rather than having a person hand-design them: instead of deciding in advance which measurements of the data matter, a model is trained to discover useful internal descriptions on its own. In 2013 that work was scattered across conferences devoted to other topics, and ICLR was created to give it a home.

Scope

The scope of the conference includes deep learning and feature learning, metric learning, kernel learning, compositional models, non-linear structured prediction, and non-convex optimization^[2]. These are related concerns. Metric learning is about learning a notion of distance that suits the task; kernel learning is about learning the similarity function a model uses; compositional models build complex representations out of simpler parts; non-linear structured prediction is about producing outputs with internal structure, such as sequences or graphs; and non-convex optimization is the mathematical machinery that training such models requires, since the objectives involved are not the well-behaved bowl-shaped functions of textbook optimization.

Open peer review

Since its inception in 2013, ICLR has employed an open peer review process to referee paper submissions, based on models proposed by Yann LeCun^[1]. In conventional peer review, reviewers are anonymous to the authors and the reviews are not published; in an open process, the reviews and the discussion around a paper are visible to a wider audience, so readers can see not only the verdict but the reasoning behind it. The first conference adopted a novel publication process, with paper decisions made available at the OpenReview website^[2].

The program

The conference includes invited talks as well as oral and poster presentations of refereed papers^[1]. Oral presentations are given a slot on the main stage, while poster presentations are shown in a session where attendees walk between boards and talk to authors directly — in practice the setting where most of the substantive exchange happens.

For the first edition, journal versions of the conference oral papers were submitted to a JMLR special topic^[2], giving the strongest accepted work a route into a journal.

Place in the field

The standing of ICLR rests on two things. The first is its subject area: representation learning and deep learning became central to machine learning, so a conference dedicated to them sat close to where much of the work was happening. The second is its review model. Because reviews and decisions are visible, the conference also functions as a public record of how a piece of research was received, which is useful to readers who encounter a paper long after it was published. Together with NeurIPS and ICML, it forms the trio of venues that define the publication calendar of the field^[1].