Energy Based Models
Current Perspectives, Challenges, and Opportunities

ICLR2021 Workshop - May 7, 2021

ICLR2021 Virtual Site: https://iclr.cc/virtual/2021/workshop/2140

About the Workshop

Energy-Based Models (EBMs) are a learning framework that assigns a quality score to any given input, its energy; contrary to probabilistic models, there is no a priori requirement that these scores be normalized (i.e. sum to one). Energies are typically computed through a neural network, and training an EBM corresponds to shaping the energy function such that data points nearby the underlying data manifold are associated with lower energies than data points that are far from it. Not imposing normalization affords a great power and flexibility to the modelling process, e.g. in terms of combining energies, on conditioning on certain variables, of computing global scores on complex structured objects, or on expressing prior knowledge. However, this freedom comes with significant technical challenges, in terms of learning and inference.

A strong comeback of EBMs is currently underway. The goals of this Workshop are:

To provide a forum for discussing cutting-edge research on different EBM approaches (e.g for continuous vs. discrete problems, strategies to shape the energy function, etc.), and to position these in the broader ML landscape (MCMC, GANs, RL, Variational Inference, etc.)
To bring together different communities, not only from core ML, but also from application domains such as vision, nlp, biology, neuroscience, etc.
To learn and exchange ideas about current and potential applications of EBMs, and to provide insights which may inspire novel EBM developments.

Invited Speakers & Panelists

Stefano Ermon
Stanford University

Will Grathwohl
University of Toronto

Yann LeCun
Facebook AI Research

Yingzhen Li (李映真)
Imperial College London

Debora Marks
Harvard University

Benjamin Scellier
Mila, Université de Montréal

Schedule

All Times in EST [New York Time]

Fri 8:50 a.m. - 9:00 a.m. Opening remarks
Fri 9:00 a.m. - 10:00 a.m. Invited Talk Live: The energy-based view of self-supervised learning Yann LeCun
Fri 10:00 a.m. - 10:02 a.m. Intro for Yingzhen Li
Fri 10:02 a.m. - 10:52 a.m.. Invited Talk: EBM inference & learning: A personal story Yingzhen Li
Fri 10:52 a.m. - 11:00 a.m. Q/A: on Yingzhen's talk
Fri 11:00 a.m. - 12:00 p.m. Panel: Probabilistic vs. Non-Probabilistic Approaches to EBMs Yann LeCun, Yingzhen Li, Will Grathwohl
Fri 12:00 p.m. - 12:25 p.m. Contributed Talk: Improved Contrastive Divergence Training of EBMs Yilun Du
Fri 12:25 p.m. - 12:30 p.m. Q/A: for Yilun Du
Fri 12:30 p.m. - 12:55 p.m. Contributed Talk: Conjugate EBMs Babak Esmaeili, Hao Wu
Fri 12:55 p.m. - 1:00 p.m. Q/A: for Babak Esmaeili, Hao Wu
Fri 1:00 p.m. - 2:00 p.m. Break (Gathertown)
Fri 2:00 p.m. - 3:30 p.m. Poster session (GatherTown)
Fri 3:30 p.m. - 3:32 p.m. Intro for invited speaker Stefano Ermon
Fri 3:32 p.m. - 4:00 p.m. Invited Talk Live: Generative Modeling by Estimating Gradients of the Data Distribution Stefano Ermon
~~Fri 4:00 p.m. - 4:02 p.m.~~ ~~Intro for invited speaker Debora Marks~~
~~Fri 4:02 p.m. - 5:00 p.m.~~ ~~Invited Talk Live:~~ ~~Can EBMs help solve important biological challenges?~~ ~~Debora Marks~~
Fri 4:00 p.m. - 5:00 p.m. Poster session (GatherTown)
Fri 5:00 p.m. - 5:02 p.m. Intro for invited speaker Benjamin Scellier
Fri 5:02 p.m. - 5:52 p.m. Invited Talk: A deep learning theory for neural networks grounded in physics Benjamin Scellier
Fri 5:52 p.m. - 6:00 p.m. Q/A on Benjamin Scellier's talk
Fri 6:00 p.m. - 6:25 p.m. Contributed Talk: Graph EBMs for Molecular Graph Generation Ryuichiro Hataya
Fri 6:25 p.m. - 6:30 p.m. Q/A for Ryuichiro Hataya
Fri 6:30 p.m. - 6:55 p.m. Contributed Talk: EBMs for Continual Learning Shuang Li
Fri 6:55 p.m. - 7:00 p.m. Q/A for Shuang Li
Fri 7:00 p.m. - 7:10 p.m. Concluding remarks