Causality-Inspired Machine Learning
Contact person: Johan Pensar
Keywords: Causality; Machine Learning; Domain Generalization
Research group: Statistics and Data Science
Department of Mathematics
There has lately been an increasing interest towards leveraging ideas and principles from causality in machine learning. When a machine learning model is trained to minimize standard empirical risk, it typically performs well within its training domain, yet the performance can degrade considerably under various shifts in the domain (or environment). One of the key characteristics of causal relationships is that they are thought to be stable (or even invariant) across domains. Thus, when a model is guided to capture causal relationships, it is more likely to generalize well to different contexts and scenarios outside its specific training domain. Moreover, causality-informed models are also thought to be more parsimonious than pure correlation-based models, enjoying an improved interpretability and sample efficiency. By drawing inspiration from the field of causality, the main goal of this research theme will be to develop new machine learning methods with improved robustness, interpretability and/or sample efficiency.
External partners:
- Norwegian Computing Center (NR)
Mentoring and internship will be offered by a relevant external partner