AutoML 2023 - The second international conference on automated machine learning brings together researchers and users, with the goals of developing automated methods for speeding up the development of machine learning applications, obtaining improved performance, and thereby democratizing machine learning.
See Dates for all deadlines; submission will be via OpenReview.

Topics

Topics of interest include but are not limited to:

• Neural Architecture Search (NAS)
• Hyperparameter Optimization (HPO)
• Meta-Learning and Learning to Learn
• Automated Data Mining
• Automated Reinforcement Learning (AutoRL)
• Bayesian Optimization for AutoML
• Evolutionary Algorithms for AutoML
• Multi-Objective Optimization for AutoML
• Combined Algorithm Selection and Hyperparameter Optimization (CASH)
• AutoAI (incl. Algorithm Configuration and Selection)
• Trustworthy AutoML (e.g., wrt fairness, robustness, sustainability, uncertainty quantification and explainability)
• Automated data wrangling
• AutoML for data types understudied by AutoML (relational, multimodal, …)
• End-to-end learning for tabular / featurized data
• Joint AutoML for multiple parts of the ML/DL/data science pipeline
• AutoML for the sciences
• AutoML for high-importance societal problems (climate change, medicine, misinformation, sustainable development goals, …)
• AutoML for social good
• Green AutoML / Making AutoML more efficient
• Applications of AutoML (yielding scientific insights, or application of knowledge discovered by AutoML)
• Reproducibility

Also see the special track (coming soon) for systems, benchmarks and challenges

Submission Guidelines

If a submission should violate any of the following rules (especially, double-blind, formatting, reproducibility and dual submission), it will get (desk-)rejected.

Double-blind Reviewing

All submissions will undergo a double-blind review, that is (i) the paper, code and data submitted for reviewing have to be anonymized to make it impossible to deduce the authors, and (ii) the reviewers will also be anonymized.

Formatting Instructions

The paper has to be formatted according to the LaTeX template available at https://github.com/automl-conf/LatexTemplate. The page limit for the main paper is 9 pages; this does not include the broader impact statement, submission checklist, references and appendix. The broader impact statement and checklist are mandatory (please see the LaTeX template for details) for both submission and camera-ready but do not count into the page limit. References and appendix are not limited in length. Accepted papers are allowed to add another page to the main paper to react to reviewer feedback. (This should already be done during the rebuttal phase if needed, to ensure acceptance decisions to be made about versions close to the camera ready.)

Submission platform

We use OpenReview to manage submissions (link). Shortly after the authors’ notification, the de-anonymized paper and anonymous reviews of all accepted papers and opt-in rejected papers will become public in OpenReview, and open for non-anonymous public commenting. Authors of rejected papers will have until two weeks after the notification deadline to opt in to make their de-anonymized papers (including anonymous reviews) public in OpenReview. Otherwise, there will be no public record that the rejected paper was submitted.

Ethics review

We ask that authors think about the broader impact and ethical considerations of their work. For example, authors may consider whether there is potential use for the data or methods to create or exacerbate unfair bias. Reviewers cannot directly reject papers based on ethical considerations but can flag papers for ethical concerns by the conference organizers, who may decide to reject papers based on these grounds (e.g., if the primary application directly causes harm or injury).

Dual submissions

The goal of AutoML 2023 is to publish exciting new work for the first time while avoiding duplicating the effort of reviewers. Papers that are substantially similar to previously published papers, accepted for publication, or submitted in parallel for publication, may not be submitted to AutoML 2023. Here we define “publication” as a paper that appears in a venue that is (i) archival, and (ii) the paper is 5 or more pages, excluding references. (This does not include non-archival workshops, or papers with up to 4 pages.)

For example:

• Allowed to submit: A manuscript on arxiv; a paper that appeared in a NeurIPS/ICML/ICLR workshop; a short CVPR/ICCV/ECCV workshop paper (<= 4 pages).
• Not allowed to submit: A conference paper published at NeurIPS/ICML/ICLR; a published journal paper.

The dual submissions policy applies for the duration of the review process. We also discourage slicing contributions too thinly.

Posting non-anonymized submissions on arxiv, personal websites, or social media is allowed. However, if posting to arXiv prior to acceptance using the AutoML style, we ask that authors use the “preprint” rather than the “final” option when compiling their document.

Reproducibility

We strongly value reproducibility as an integral part of scientific quality assurance. Therefore, we require that all submissions are accompanied by A link to an open source repository providing an implementation (if empirical results are part of the paper). To abide by double-blind reviewing, we have hosted our own version of anonymous GitHub which supports anonymization and full download of repositories: https://anon-github.automl.cc/.

The submission checklist, which is part of the LaTeX template and does not count as part of the 9-pages (incl. details of repeated measurements, tuned hyper(-hyper-)parameters, …). We recommend implementing new ideas in existing packages instead of re-implementing the basics from scratch and thereby introducing many confounding factors. For example, a new acquisition function could be implemented in one of the many Bayesian optimization packages. (A side effect of this can be easier usability and thus increased impact.)

Author responses

During the rebuttal phase, the authors are allowed to update their papers based on the questions and the feedback of the reviewers. Adding more co-authors is not possible.

Commitment to review

We ask that for each submission, at least one author volunteers to review for AutoML 2023 via the volunteer form.

Changing the Author List

New authors cannot be added after the abstract deadline. Changing the ordering and removing authors is allowed, but further authors cannot be added after the abstract deadline. This is because reviewer bidding happens right after the abstract deadline, and the full set of authors is needed to correctly identify conflicts of interest

Publication of accepted submissions

The publication of accepted submissions will be done via OpenReview. Feedback by the reviewers can be incorporated into the final version of the paper; other major changes are not allowed. Furthermore, each paper has to be accompanied by a link to an open-source implementation (if there are any empirical results in the paper) to ensure reproducibility. As mentioned above, accepted papers will be made available alongside with their reviews and meta-reviews.

Attending the Conference

The conference is scheduled for September 12-15, 2023 in Berlin, Germany. We are planning for a hybrid conference, with an active integration of remote attendees. The conference’s main objective is to allow for in-depth discussions and networking. It is not mandatory for authors of accepted papers to attend the conference in person, but we encourage it. Nevertheless, for all accepted papers, we require a short video. This will allow attendees to watch videos ahead of time at their own pace and plan ahead to use the conference most effectively for in-depth discussions and networking.

Program Chairs

• Aleksandra Faust
• Roman Garnett
• Colin White

For any questions, please don’t hesitate to contact us: pc-chairs-2023@automl.cc