Towards Causal Deep Learning for Explainability, Robustness, and Extrapolation

In many applications, Machine Learning (ML) predictions are used to make downstream decisions. Acting on ML predictions however can change the distribution of features that the ML model relies on for predictions. The implication is that such downstream decisions procedures implicitly expect the ML model to generalize outside of the observational distribution. Unfortunately, this is often not the case, and ML models tend to be brittle outside of their training distribution. ML models will thus produce unreliable extrapolations, leading to poor downstream decisions based on wrong predictions.

Causal models, which aim to learn the structural causal models underlying the data generation process, are a natural fit for such use-cases. This is because causal mechanisms are more robust to superficial changes in the data distribution, and can be expected to extrapolate better to new environments. This project aims to combine deep learning and causal inference to develop causal forecasting models adapted to three important applications. Students will start by implementing existing approaches on three applications (demand forecasting, health prediction, and ML model understanding), before working on new techniques for causal ML in three directions.

Faculty Supervisor:

Mathias Lécuyer

Student:

Partner:

Institut Polytechnique de Paris

Discipline:

Computer science

Sector:

Education

University:

The University of British Columbia

Program: