Reproduce “Weakly Supervised Disentangled Generative Causal Representation Learning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Xinwei Shen, Furui Liu, Hanze Dong, Qing Lian, Zhitang Chen, Tong Zhang
Researchers from HKUST, Zhejiang Laboratory, and Huawei Noah's Ark Lab developed DEAR, a framework that learns disentangled generative causal representations by integrating Structural Causal Models into a bidirectional generative framework. The approach, which utilizes weak supervision, demonstrated successful disentanglement of causally related factors and led to improved sample efficiency and distributional robustness in downstream tasks.
Key insights
- Existing disentanglement methods that assume independent latent priors inherently fail to disentangle underlying factors when those factors are causally related, even with supervision.
- Integrating a Structural Causal Model (SCM) as the prior distribution for latent variables within a bidirectional generative model enables the identification and disentanglement of causally related factors under weak supervision.
- Learning disentangled causal representations yields substantial improvements in sample efficiency and distributional robustness for downstream prediction tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.