Beyond Worst-case Guarantees for Sequential Prediction - Robustness via Abstention

Explore a 47-minute lecture on sequential prediction and robustness in machine learning presented by Surbhi Goel from the University of Pennsylvania. Delve into the challenges of handling stochastic sequences with adversarial examples and discover a novel approach that allows learners to abstain from predictions on uncertain data. Examine how this method maintains regret scaling similar to purely stochastic settings while accommodating any number of adversarial examples. Gain insights into the limitations of traditional algorithms and the potential of this new model for improving prediction accuracy and reliability. Investigate open questions posed by this innovative approach and its implications for the field of machine learning and optimization.