This ICML workshop seeks to broaden the role of geometric models and thinking in machine-learning research.

The interplay between machine learning and geometry is an active field of research drawing the attention of researchers from many fields as it offers not only beautiful mathematical and statistical theory but also substantial impact on important real-world problems in machine learning. Examples of this interplay include, but are not limited to:

  • geometric insights into deep learning (helping us to better understand and improve these models);
  • statistical models that respect and exploit the constraints of geometric data ("statistics on manifolds");
  • (Riemannian) manifold learning;
  • viewing probability distributions as points on a nonlinear manifold ("information geometry");
  • optimization over nonlinear manifolds;
  • Wasserstein spaces and their applications;
  • understanding the effect of encoding group invariances (e.g. rotation invariance) on the intrinsic geometry of the data space, which becomes a quotient space.

This ICML workshop is co-located with ICML, IJCAI and AAMAS in Stockholm.

The Bosch center for AI generously sponsors a prize for the best abstract, which will be presented at the workshop.

The meeting is now passed -- thanks to all attendees for fantastic discussions. Also, congratulations to William Guss for winning the Best Contribution Award (sponsored by Bosch Center for Artificial Intelligence)!

Confirmed Speakers

Justin Solomon

Justin Solomon
Assistant Professor, MIT

Speaker name

Nina Miolane
Research Fellow, Stanford University

David Rosen

David Rosen
MIT LIDS, formerly at Oculus Research

John Skilling

John Skilling
Research Director, Maximum Entropy Data Consultants Ltd.

Frank Nielsen

Frank Nielsen
Sony Computer Science Laboratories Inc, Japan

Stefano Soatto

Stefano Soatto
Professor, UCLA

Tentative Program

Date: Sunday July 15, 2018
Session 1. Chair: TBD
8:30-9:15 Justin Solomon
Correspondence and Embedding for Geometric Data
9:15-10:00 Nina Miolane
Geometric Statistics: Learning from Medical Images?
10:00-10:30 Coffee
Session 2. Chair: TBD
10:30-11:15 David Rosen
Certifiably Correct SLAM
11:15-11:45 Contributed talk: Cyrus Mostajeran
Homogeneous Cone Fields: Affine-Invariant Orders
11:45-12:30 John Skilling
Failures of Information Geometry
12:30-14:00 Lunch
Session 3. Chair: TBD
14:00-14:45 Frank Nielsen
Information geometry: Structures and their uses
14:45-15:30 Stefano Soatto and Pratik Chaudhari
High-dimensional Geometry and Dynamics of Stochastic Gradient Descent for Deep Networks
15:30-16:00 Coffee (poster session starts)
16:00-18:00 Poster Session (public abstracts)

A list of accepted abstracts are available here.

Call for Contributions

Submission deadline: May 7, 2018
Author notification: May 21, 2018

We are looking for poster/oral contributions to the second instalment of this workshop which explores the intersection between geometry and machine learning. Submissions will be reviewed based on short abstracts. Maximum length is one page (plus at most one additional page which can contain only references) using the ICML 2018 format. See ICML website for the template files. This is not a "blind" submission; i.e., the authors' identities should be included in the PDF. Submission is open to both unpublished and previously-published work. There will be no proceedings for this workshop; however, upon an author’s request, an accepted contribution will be made available on the workshop website. This is a great way to share recent and ongoing work with the community.
Submissions should be e-mailed to gimli.meeting@gmail.com
There will a sponsored prize for the best abstract, which will be presented at the workshop.

Venue

The workshop takes place at the ICML conference venue, room A4.

Organizers

The conference is jointly organized by:

For matters regarding the conference, you can contact the organizers at gimli.meeting@gmail.com.

We are grateful for funding from the Bosch Center for Artificial Intelligence, the Villum Fonden Young Investigator program and the European Research Council (ERC) through a starting grant.