Disentangled representation learning is crucial for robust attribute prediction. While recent methods have begun addressing attribute correlations, hidden correlations remain underexplored. Data under certain attribute values may exhibit underlying modes correlated with other attributes. To preserve mode information and achieve disentanglement, we introduce a method that jointly discovers modes and enforces mode-based conditional independence. However, the interdependency between these two modules may lead to error amplification during naive iterations.
We propose Coordinated Disentanglement with Iterative Mode Discovery (CoDID), an end-to-end framework featuring a dynamic architecture that adapts to the evolving number of modes, along with a coordination mechanism that mitigates error amplification through meta-optimization. Empirical results demonstrate state-of-the-art performance across diverse tasks.
Blogger's Review: This study introduces a novel approach to disentangled representation learning by effectively addressing hidden correlations through a coordination mechanism. Its adaptability and flexibility show promise for future applications, making it a noteworthy contribution in this domain.