ASSET Seminar: “What Should We “Trust” in Trustworthy Machine Learning?” (Aaron Roth, University of Pennsylvania)

“Trustworthy Machine Learning” has become a buzz-word in recent years. But what exactly are the semantics of the promise that we are supposed to trust? In this talk we will make a proposal, through the lens of downstream decision makers using machine learning predictions of payoff relevant states: Predictions are “Trustworthy” if it is in the interests of the downstream decision makers to act as if the predictions are correct, as opposed to gaming the system in some way. We will find that this is a fruitful idea. For many kinds of downstream tasks, predictions of the payoff relevant state that are statistically unbiased, subject to a modest number of conditioning events, suffice to give downstream decision makers strong guarantees when acting optimally as if the predictions were correct — and it is possible to efficiently produce predictions (even in adversarial environments!) that satisfy these bias properties. This methodology also gives an algorithm design principle that turns out to give new, efficient algorithms for a variety of adversarial learning problems, including obtaining subsequence regret in online combinatorial optimization problems and extensive form games, and for obtaining sequential prediction sets for multiclass classification problems that have strong, conditional coverage guarantees — directly from a black box prediction technology, avoiding the need to choose a “score function” as in conformal prediction.

This is joint work with Georgy Noarov, Ramya Ramalingam, and Stephan Xie