Reproduce “IndiSeek learns information-guided disentangled representations” 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

Yu Gui, Cong Ma, Zongming Ma

IndiSeek presents an information-guided method for disentangled multi-modal representation learning, which effectively separates shared and modality-specific features. The approach utilizes CLIP for shared features and introduces a novel objective combining cross-modal mutual information minimization and reconstruction-based information completeness, leading to superior interpretability and enhanced performance (e.g., 0.910 Spearman correlation on CITE-seq and up to 70.03% accuracy on MultiBench) compared to existing methods.

Key insights

  • Achieving truly disentangled representations requires strict criteria: minimal sufficient shared features, and cross-modal independence of modality-specific features from the other modality's shared information.
  • The reconstruction error based on combining shared and modality-specific features provides a principled, theoretically justified way to ensure complete information capture without relying on sensitive mutual information estimators.
  • Existing methods' limitations stem from approximations like InfoNCE sensitivity or restrictive assumptions like linearity, which IndiSeek's hybrid approach addresses by using a robust NCE-CLUB and reconstruction loss.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.