Learning Causally Disentangled Representations via the Principle of Independent Causal Mechanisms
This work addresses the challenge of extracting meaningful causal information for downstream tasks in machine learning, representing an incremental advancement in causal representation learning.
The paper tackles the problem of learning causally disentangled representations by introducing a new notion based on independent causal mechanisms and proposing the ICM-VAE framework, which theoretically ensures identifiability and empirically shows improved disentanglement, interventional robustness, and compatibility with counterfactual generation.
Learning disentangled causal representations is a challenging problem that has gained significant attention recently due to its implications for extracting meaningful information for downstream tasks. In this work, we define a new notion of causal disentanglement from the perspective of independent causal mechanisms. We propose ICM-VAE, a framework for learning causally disentangled representations supervised by causally related observed labels. We model causal mechanisms using nonlinear learnable flow-based diffeomorphic functions to map noise variables to latent causal variables. Further, to promote the disentanglement of causal factors, we propose a causal disentanglement prior learned from auxiliary labels and the latent causal structure. We theoretically show the identifiability of causal factors and mechanisms up to permutation and elementwise reparameterization. We empirically demonstrate that our framework induces highly disentangled causal factors, improves interventional robustness, and is compatible with counterfactual generation.