Publication
Causality in Flux: Continual Adaptation of Causal Knowledge via Evidence Matching
Jonas Seng; Florian Peter Busch; Kristian Kersting
In: Martin Mundt; Keiland W. Cooper; Devendra Singh Dhami; Tyler L. Hayes; Rebecca Herman; Adéle Ribeiro; James Seale Smith (Hrsg.). Proceedings of The Second AAAI Bridge Program on Continual Causality. AAAI Bridge Program on Continual Causality, February 20-21, Vancouver, Canada, Pages 11-20, Proceedings of Machine Learning Research (PMLR), Vol. 268, PMLR, 2024.
Abstract
Utilising causal knowledge in machine learning (ML) systems yields more robust models
with the capability of performing certain extrapolations. However, much of current causal-
ity research focuses on deriving causal models in isolation, hence current systems are not
capable of updating and improving causal knowledge when new observations arrive. Draw-
ing inspiration from human learning, Continual Learning (CL) aims at updating models
given a sequential stream of evidence. Leveraging common patterns and past experiences
to gradually improve causal knowledge in ML models is a crucial step towards more robust
CL systems. In this work, we propose to learn and update causal models in a lifelong learn-
ing setting where causal knowledge explaining newly arriving observations is inferred from
similar previously seen observations. We call this framework evidence matching. Further,
an analysis of real world data supporting our motivation is provided