Calibration and Generalizability of Probabilistic Models on Low-Data Chemical Datasets
Explore a comprehensive conference talk on calibration and generalizability of probabilistic models for low-data chemical datasets. Delve into the DIONYSUS study, which examines various molecular representations and models for predicting molecular properties. Learn about three key experiments: performance analysis, Bayesian optimization for molecular design, and out-of-distribution inference using ablated cluster splits. Gain practical insights into model and feature selection for small chemical datasets, a common scenario in new chemical experiments. Discover the open-source DIONYSUS repository, designed to aid reproducibility and extension to new datasets. Follow along with the speaker's in-depth analysis, covering motivations, experimental overviews, and practical recommendations, concluding with a Q&A session.