Statistics & Data Science Seminar

  • Remote Access - Zoom
  • Virtual Presentation - ET
Seminars

Underspecification Presents Challenges for Credibility in Modern Machine Learning

In this talk, I'll discuss a paper by the same name in which many colleagues and I probed real-world machine learning systems for poorly constrained behavior. The paper I'll be discussing is here, with the following abstract.

Machine learning (ML) models often exhibit unexpectedly poor behavior when they are deployed in real-world domains. We identify underspecification as a key reason for these failures. An ML pipeline is underspecified when it can return many predictors with equivalently strong held-out performance in the training domain. Underspecification is common in modern ML pipelines, such as those based on deep learning. 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 show that this problem appears in a wide variety of 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.

Some other relevant papers for discussion are here  and here , but the talk will focus mostly on the main paper.

Alexander D'Amour is a Senior Research Scientist at Google Research. His research is focused on the intersection of causality and machine learning: both how we can use machine learning to more effectively answer causal questions, and how we can use concepts from causal inference to make machine learning more robust, fair, and trustworthy. Alex has also done work in a number of application areas, including sports analytics, marketing, and epidemiology. Prior to joining Google, Alex was a Visiting Assistant Professor in the UC Berkeley Department of Statistics, and received his PhD from the Harvard University Department of Statistics.

Seminars consist of a 40-minute talk, followed by a 5-10 minute 'discussion' by the speaker's host, and then followed by Q&A.

Zoom Participation. See announcement.

For More Information, Please Contact: 
Keywords: