Title: Uncertainty quantification in semi-structured distributional regression models
Authors: Daniel Dold - Institute for Optical Systems (IOS) Hochschule Konstanz (Germany) [presenting]
Oliver Duerr - Hochschule Konstanz (Germany)
Beate Sick - University of Zurich Zurich University of Applied Sciences (Switzerland)
David Ruegamer - TU Dortmund (Germany)
Abstract: Many applications require predicting a conditional outcome distribution based on semi-structured input data, such as the combination of images and tabular data. Recent deep semi-structured distributional regression models combine deep neural networks employed for complex data and statistical regression models for structured data. This approach combines the high predictive power of neural networks and the interpretability of statistical models. While deep semi-structured distributional regression models enable to account for the aleatoric uncertainty, it is still not clear how to capture the model uncertainty (epistemic uncertainty), which is of pivotal interest in many applications like out-of-distribution detection or active learning. A natural way of capturing epistemic uncertainty is to use Bayesian approaches. Many Bayesian methods exist for the two individual parts of semi-structured deep regression models. For statistical models, variants of Markov Chain Monte Carlo (MCMC) are usually successfully used. However, MCMC is not efficient for a deep neural network because of the high dimensional parameter space, and therefore, an approximate approach like variational inference needs to be used. We propose an efficient Bayesian approach for the whole semi-structured model.