Learning Independent Causal Mechanisms

Giambattista Parascandolo, Mateo Rojas Carulla, Niki Kilbertus, Bernhard Schoelkopf

Abstract:

Independent causal mechanisms are a central concept in the study of causality
with implications for machine learning tasks. In this work we develop
an algorithm to recover a set of (inverse) independent mechanisms relating
a distribution transformed by the mechanisms to a reference distribution.
The approach is fully unsupervised and based on a set of experts that compete
for data to specialize and extract the mechanisms. We test and analyze
the proposed method on a series of experiments based on image transformations.
Each expert successfully maps a subset of the transformed data
to the original domain, and the learned mechanisms generalize to other
domains. We discuss implications for domain transfer and links to recent
trends in generative modeling.