Title: Taking Uncertainty Seriously for Calibrated Forecasting and Decision Making

Abstract: As scientists, we want to understand the world. As engineers and policymakers, we need to make concrete decisions. Yet we are faced with convenience samples based on noisy measurements and limited knowledge of underlying processes. In this talk, I'll demonstrate how to formulate scientific models, couple them with measurement models, and then solve the resulting inverse problem to drive prediction and decision analysis. I'll illustrate with classical examples involving Laplace's model of birth sex-ratio, Galileo's model of gravity, Lotka & Volterra's model of population dynamics, and the Efron-Morris model of batting ability. I'll also discuss my own ongoing work on soil-carbon respiration, differential expression of gene splice variants, and crowdsourced corpus labeling for machine learning. In all cases, I'll focus on empirical evaluation of calibration and sharpness, which generalize bias and variance to probabilistic forecasts. To solve problems like these, we continue to develop Stan, an expressive probabilistic programming language, efficient differentiable mathematics library, and robust statistical inference engine. In the second part of the talk, I'll provide an overview of Stan's foundations in language theory, automatic differentiation, constrained variable transformations, adaptive Hamiltonian Monte Carlo sampling, diagnostic robustness checks, and general posterior predictive inference.

Bio: Bob Carpenter is a research scientist at Columbia University, where he founded the Stan project. He previously held a tenured faculty position at Carnegie Mellon University in computational linguistics and a research position at Bell Laboratories in speech recognition. He later worked as an industrial programmer at SpeechWorks building speech recognizers and at Alias-i building the LingPipe natural language processing toolkit. He has a Ph.D. in computer and cognitive science from the University of Edinburgh.