Seminar by Blair Bilodeau

Please Note: This seminar will be given in person.

Adaptive Sequential Decision Making and Uncertainty Quantification

One way to quantify the risks of deploying complex statistical methods is theoretical guarantees, yet statistical theory often relies on unverifiable assumptions and can therefore fail to explain performance in real-world settings. My research seeks out guarantees without such limitations across a wide range of statistical tasks, including inference, prediction, and decision making. In this talk, I will present two papers from this research program.

First, I will present https://arxiv.org/abs/2202.05100 (awarded an Oral designation at NeurIPS 2022, reserved for only 2% of submissions), where we study how to most efficiently select interventions in sequence to learn causal effects. We provide an adaptive method and corresponding guarantees: simultaneously optimal performance when benign causal structure exists and consistent estimation even when all causal assumptions fail. Second, I will present https://arxiv.org/abs/2109.10461, where we resolve the minimax rates for conditional density estimation in parametric and nonparametric classes. I will particularly focus on consequences of our results, including the first dimension-free KL risk bounds for generalized linear models, the first fast rates for KL risk with arbitrarily unbounded covariate spaces, and the first characterizations of KL risk for natural extensions of smoothness to conditional densities. Finally, I will discuss how these advances form a foundation of my future research: general adaptivity in non-stationary and partial-feedback settings.