Recent Progress in Predictive Inference - Emmanuel Candes, Stanford University

Explore cutting-edge techniques in predictive inference with Stanford University's Emmanuel Candes in this 59-minute conference talk from the Alan Turing Institute. Delve into novel calibration methods addressing critical issues in machine learning, including Learn then Test for explicit finite-sample statistical guarantees and adaptive conformal inference for maintaining prediction coverage despite distribution shifts. Discover how these approaches can be applied to various domains, from multi-label classification and instance segmentation to economic forecasting during major world events. Gain insights into the latest advancements in trustworthy artificial intelligence, covering topics such as machine learning accountability, fairness, privacy, and safety. Follow along as Candes discusses data ethics, conformal inference, prediction intervals, and risk calibration, using real-world examples from medical service utilization, stock market volatility estimation, and county-level election result predictions.

Intro
Data ethics 101: convey uncertainty and reliable outcomes
Previous work on conformal inference
Prediction intervals
Setting with perfect knowledge
Formulate quantile estimation as a learning task
Validity for unseen data?
Calibrate: how?
Comparison with other implementations of conformal inference
Predicting utilization of medical services
Online methods?
Adapting conformal to distribution shift
Connections
Estimating volatility in the stock market
Distribution free theory
Hidden Markov model
Predicting county level election results
From tolerance region to PAC-learning
Learn then test: risk calibration via multiple hypothesis testing
Example: object detection
Summary