Robust Machine Learning for Usable Decision Systems
Practical decision systems require much more than end-to-end learned models. This talk will focus on research and engineering questions on machine learning robustness and the two broad categories of work in that area: [A] model correctness analysis, verification & validation; and [B] bounding the risk of handling observations for which the system is not qualified or competent. The first part of the talk will focus on the latter set of questions: how do we learn models that ‘know when they don’t know’, what are the formalisms that we need, what is practically doable-both, for supervised learning and for reinforcement learning. First, we will explore methods for learning self-competence models in addition to the predictions. Next, we will discuss self-competence estimation approaches for reinforcement learning and decision-making. Finally, we will explore multi-faceted learning and related techniques, that can be applied in decision systems with the goal of increased robustness.
Session ID: KEY1316 Presentation Type: Live Keynote Session (Replay Available)
Date / Time: [Day 3] Wed. Sep. 16, 2020 @ 13:00 ET (US)
To view this session, register for the conference.