AI Seminar: A stability upgrade for data-driven models
Speaker
Jonas Martin Peters, Associate Professor, Department of Mathematical Sciences University of Copenhagen.
Abstract
When we learn models from data, we traditionally focus on predictability, that is, we prefer models that predict the observed data well (e.g., in a cross-validation scheme). Such models, however, often fail to generalize to unseen experiments. In this work, we propose to exploit ideas from causality for model learning. In particular, we suggest to explicitly take into account stability across different experiments as a fitting criteria. As an example, we discuss CausalKinetiX and its application to data from metabolic networks. Many real world systems can be described by a set of differential equations. For a lot of complex systems, the differential equations are unknown and data science methodology is used to learn them from data. Unlike traditional approaches, CausalKinetiX trades off predictability and stability. It thereby infers models that generalize better to unseen experiments and that draw a more realistic picture of the system's underlying structure. CausalKinetiX is joint work with Niklas Pfister and Stefan Bauer, see https://arxiv.org/abs/1810.11776. No prior knowledge about causality is required.
This seminar is a part of the AI Seminar Series organised by SCIENCE AI Centre. The series highlights advances and challenges in research within Machine Learning, Data Science, and AI. Like the AI Centre itself, the seminar series has a broad scope, covering both new methodological contributions, ground-breaking applications, and impacts on society.