Integreat will integrate causal knowledge into ML models by imposing a causal backbone structure represented, for example, by a graph. A causality-aware ML model will be able to transfer some of its parts to new domains and tasks without re-training, and it will also ultimately enable causal reasoning using the framework of do-calculus to improve itself. Causal inference is a means towards sustainable ML, because causal models are often more parsimonious, and thus transparent and cost-efficient in training.
For situations where causal knowledge is lacking, Integreat will develop new methods for extracting such knowledge from data, building upon structural equation models and causal structure learning, a technique that has recently shown great potential when approached from the Bayesian perspective. A further new idea is to introduce synthetic causal knockoffs of a known form, to help inferring new causal relations. To acquire knowledge, intervention experiments are also very useful; we will exploit the availability of a knowledge-based probabilistic model of the system to propose adaptive designs. In reinforcement learning, the agent interacts with the environment to learn causal relations, thus effectively performing intervention experiments for decision making.
A key challenge is the amount of uncertainty involved in causal knowledge, especially when inferred from data. To this end, as a first step, we will consider the class of graphical probabilistic models for large systems of variables that are designed for efficient representation and inference. Invariant predictions across domains and data sets will allow uncertainty assessment of learned causal relations.
Key researchers in this research theme: