Abstract

Discovering semantic meaningful latent factors and the causal relations among them is an emergent topic in representation learning with notable impacts on real-world applications. However, many existing Causal Representation Learning (CRL) methods are hindered by strong assumptions, such as full data annotation, the need for counterfactual data, and/or prior knowledge of the causal structure. To address these limitations, we introduce Causal-Macro, a weakly supervised architecture that effectively discovers semantic causal factors and learns their causal relations. We theoretically show that Causal-Macro is identifiable in the sense that the marginalized posterior distribution of learned factors can be identified up to coordinate-wise reparameterization of groundtruth factors. Additionally, we show that existing CRL datasets are limited to simple causal graphs with a small number of generative factors. Thus, we propose two new datasets with a larger number of generative factors and more sophisticated causal graphs. Our comprehensive evaluations and detailed ablation studies demonstrate the superior performance of Causal-Macro over existing methods.

Date

October 31, 2025

Authors

Jiageng Zhu, Hanchen Xie, Jianhua Wu, Mohamed E Hussein, Mahyar Khayatkhoei, Jiazhi Li, Wael AbdAlmageed

Journal

35th British Machine Vision Conference 2024, BMVC 2024, Glasgow, UK, November 25-28, 2024

Publisher

BMVA