Timezone: »

 
Poster
Underspecification Presents Challenges for Credibility in Modern Machine Learning
Alexander D'Amour · Katherine Heller · Dan Moldovan · Ben Adlam · Babak Alipanahi · Alex Beutel · Christina Chen · Jonathan Deaton · Jacob Eisenstein · Matthew Hoffman · Farhad Hormozdiari · Neil Houlsby · Shaobo Hou · Ghassen Jerfel · Alan Karthikesalingam · Mario Lucic · Yian Ma · Cory McLean · Diana Mincu · Akinori Mitani · Andrea Montanari · Zachary Nado · Vivek Natarajan · Christopher Nielson · Thomas F. Osborne · Rajiv Raman · Kim Ramasamy · Rory sayres · Jessica Schrouff · Martin Seneviratne · Shannon Sequeira · Harini Suresh · Victor Veitch · Maksym Vladymyrov · Xuezhi Wang · Kellie Webster · Steve Yadlowsky · Taedong Yun · Xiaohua Zhai · D. Sculley

Thu Jul 27 01:30 PM -- 03:00 PM (PDT) @ Exhibit Hall 1 #204
in Poster Session 5 »

Machine learning (ML) systems often exhibit unexpectedly poor behavior when they are deployed in real-world domains. We identify underspecification in ML pipelines as a key reason for these failures. An ML pipeline is the full procedure followed to train and validate a predictor. Such a pipeline is underspecified when it can return many distinct predictors with equivalently strong test performance. Underspecification is common in modern ML pipelines that primarily validate predictors on held-out data that follow the same distribution as the training data. Predictors returned by underspecified pipelines are often treated as equivalent based on their training domain performance, but we show here that such predictors can behave very differently in deployment domains. This ambiguity can lead to instability and poor model behavior in practice, and is a distinct failure mode from previously identified issues arising from structural mismatch between training and deployment domains. We provide evidence that underspecfication has substantive implications for practical ML pipelines, using examples from computer vision, medical imaging, natural language processing, clinical risk prediction based on electronic health records, and medical genomics. Our results show the need to explicitly account for underspecification in modeling pipelines that are intended for real-world deployment in any domain.

[ JMLR] 

Author Information

Alexander D'Amour (Google DeepMind)
Katherine Heller (Google)
Dan Moldovan
Ben Adlam (Google Brain)
Babak Alipanahi
Alex Beutel (OpenAI)
Christina Chen
Jonathan Deaton
Jacob Eisenstein (Google AI)
Matthew Hoffman (Google)
Farhad Hormozdiari
Neil Houlsby (Google)
Shaobo Hou
Ghassen Jerfel (Waymo Research)
Alan Karthikesalingam (Google Health)
Mario Lucic (Google Brain)
Yian Ma (UCSD)
Cory McLean (Google LLC)
Diana Mincu
Akinori Mitani (Google)
Andrea Montanari (Stanford University)
Zachary Nado (Google Research, Brain Team)
Vivek Natarajan
Christopher Nielson
Thomas F. Osborne (VA Palo Alto Healthcare System)
Thomas F. Osborne

Thomas Osborne, MD, is the Director of VA’s National Center for Collaborative Healthcare Innovation (NCCHI) and Executive Director of VA Convergence Center (VC2). He is leading the development, assessment, and deployment of pioneering healthcare solutions throughout VA, with other government agencies, and with industry. His teams mission is to deliver the best and most advanced healthcare solutions to our Veterans. Dr. Osborne is the inaugural recipient of the VA Under Secretary for Health Robert L. Jesse Award for Excellence in Innovation, the Arthur S. Flemming Award for exceptional public service, as well as multiple other national awards. His work has been published in numerous medical journals and textbooks on topics such as health equity, predictive analytics, artificial intelligence, sensor technology, augmented reality, virtual reality, and the future of health care. Dr. Osborne received his medical degree from Dartmouth Medical School before completing residency and fellowship at Harvard hospitals.

Rajiv Raman
Kim Ramasamy
Rory sayres (Google)
Jessica Schrouff (Google Health)
Martin Seneviratne (Google Health)
Shannon Sequeira (Google)
Harini Suresh
Victor Veitch
Maksym Vladymyrov (Google)
Xuezhi Wang (Google Deepmind)
Kellie Webster (Google)
Steve Yadlowsky (Google Brain)
Taedong Yun (Google Research)
Xiaohua Zhai (Google Brain)
D. Sculley (Google)

More from the Same Authors