Robustness in Learning and Statistics: Past and Future

Overview:

Robust statistics and related topics offer ways to stress test estimators to the assumptions they are making. It offers insights into what makes some estimators behave well in the face of model misspecification, while others do not. In this summer school, we will revisit classic topics in robust statistics from an algorithmic perspective. We will cover recent progress on provably robust and computationally efficient parameter estimation in high-dimensions. We will compare this to other popular models, like agnostic learning and outlier detection. With the foundations in hand, we will explore modern topics like federated learning, semi-random models and connections to decision theory where being robust is formulated in alternative ways. We hope to have time for discussion about open questions like adversarial examples in deep learning, and invite the audience to help us muse about the right definitions to adopt in the first place.

Target audience:

The school is oriented towards graduate students, postdocs and faculty, who either work in robust statistics or want to understand the questions at the forefront of current research.

Topics:

• 1d robust statistics: Huber contamination model, robust estimates of location and scale, empirical mean vs. median, empirical variance vs. MAD, breakdown point
• High-d robust statistics: high break-down point for regression, Tukey median, computational difficulties
• Comparison to supervised learning, outlier detection
• Modern algorithms: filtering, Sum-of-squares
• Broader view of robustness: list-decodable learning, semi-verified learning, federated learning, semi random models, selective prediction
• Connections to modern decision theory

Schedule:

Full schedule (including slides)

Registration:

Registration is free but is required. The deadline was July 31, 2019. Registration is now closed.

Speakers:

David Donoho Stanford	Ankur Moitra MIT	Gregory Valiant Stanford

Accommodations:

Supported by:

Questions? please email Shachar Lovett.