Generative flow-based models in application to uncertainty modeling for machine learning tasks

The modern machine learning models achieve incredible results in various tasks, including classification, regression, object detection, semantic segmentation, clustering, learning representations, generating multiple types of data, and many others. For the major of the mentioned applications, the goal is rather to provide the deterministic prediction, not focusing much on modeling the uncertainty of the predictions. While working on the project, we will develop, implement, and validate a set of machine learning approaches that will solve the problem of modeling uncertainty in various tasks, including regression, clustering, and modeling the sequential data. Standard baseline approaches usually use simple normal distributions or mixtures to model the uncertainty. We are going to overcome the limitations by proposing the set of architectures that utilize the generative flow-based models - the flexible generative approaches that make use of the change-of-variable formula. We plan to show how to incorporate those models into well-known structures, like deep neural networks and tree-based learners of Hidden Markov Models, and train them in an end-to-end fashion. We will implement and validate the various types of such models and validate them on many machine learning problems from different domains. We will design and validate the tools that will help better understand the distributions modeled by the flow-based components. Further, we will work on the approaches that enrich the base distribution of the flows. All of the output solutions of the project are going to be evaluated quantitatively and qualitatively using the methodology characteristic for the domain. We plan to publish the results of our work at top machine learning conferences, including AAAI, NeurIPS, ICML, ICLR.