ML Interpretability for Scientific Discovery
ICML 2020 Workshop
Update
Registered ICML attendees please join us at the webinar link here.
Event will be streamed live on Youtube here: https://www.youtube.com/watch?v=0Q-4EQriYJs
Overview
ML has shown great promise in modeling and predicting complex phenomenon in many scientific disciplines such as predicting cardiovascular risk factors from retinal images, understanding how electrons behave at the atomic level, identifying patterns of weather and climate phenomena, etc. Further, models are able to learn directly (and better) from raw data as opposed to human selected features. The ability to interpret the model and find significant predictors could provide new scientific insights.
Traditionally, the scientific discovery process has been based on careful observations of natural phenomenon, followed by systematic human analysis (of hypothesis generation and experimental validation). ML interpretability has the potential to bring a radically different yet principled approach. While general interpretability relies on ‘human parsing’ (common sense), scientific domains have semi-structured and highly structured bases for interpretation. Thus, despite differences in data modalities and domains, be it brain sciences, the behavioral sciences, or material sciences, there is a need for a common set of tools that address a similar flavor of problem, one of interpretability or fitting models to a known structure.
This workshop aims to bring together members from the ML and physical sciences communities to introduce exciting problems to the broader community, and stimulate the production of new approaches towards solving open scientific problems.
Topics
Topics of interest include but are not limited to
Applications of deep learning based interpretability techniques to scientific domains.
Causality - causal models.
Combining structure with ML for discovery.
Interpretability techniques (or problems requiring interpretablity) for different data modalities: images, time-series, audio/speech, text, multi-modal data, voxels as images, point clouds, very high definition images e.g. MRI, etc.
GANs / generative modeling for representation understanding, visual interpretation etc.
Simulations and or synthetic experiments to evaluate or enable interpretation.
Representation learning in a world where there’s a rich, structure.
Interpretable / Disentangled representation learning
Visualizations for model or data explanation.
New datasets , challenges, benchmarks.
This is not an exhaustive list. We welcome submissions from a wide range of sciences including but not limited to brain sciences, behavioral sciences, weather/climate science, physics, chemistry, biology, medical applications, and others. We also invite folks from the ML community with suggestions on tools and members of the science communities with problems.
Invited Speakers
Participate
ICML 2020 will be a fully virtual conference. We will follow the guidance laid out by the ICML 2020 organizers with regard to the workshop participation.
We invite 2-4 page extended abstracts of unpublished works and previously published works that are in the theme of ML interpretability for scientific discovery. More details on how to submit are here.
Program Committee
Program committee members and reviewers:
Akinori Mitani (Google)
Amir Feder (Technion - Israel Institute of Technology)
Amirata Ghorbani (Stanford University)
Arunachalam Narayanaswamy (Google)
Avinash Varadarajan (Google)
Awa Dieng (Google)
Benjamin Sanchez-Lengeling (Google)
Bo Dai (Google Brain)
Chih-Kuan Yeh (Carnegie Mellon University)
Hanna Levitin (Columbia)
Katy Blumer (Cornell)
Kevin Wu (Stanford University)
Martin Forsythe (Lightmatter Inc.)
Miles Cranmer (Princeton University)
Pang Wei Koh (Stanford University)
Ramin Ansari (University of Michigan)
Stephan Hoyer (Google)
Subham Sekhar Sahoo (Google)
Suhani Vora (Google)
Wesley Wei Qian (University of Illinois at Urbana-Champaign)