Reproduce “On Deep Multi-View Representation Learning: Objectives and Optimization” 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

Weiran Wang, Raman Arora, Karen Livescu, Jeff Bilmes

Researchers at Toyota Technological Institute at Chicago, Johns Hopkins University, and the University of Washington systematically analyze deep multi-view learning approaches and introduce Deep Canonically Correlated Autoencoders (DCCAE). The work demonstrates that CCA-based methods, especially DCCAE, consistently achieve superior performance on downstream tasks by effectively leveraging cross-view information, while also providing practical insights into the stochastic optimization of correlation objectives.

Key insights

  • Canonical Correlation Analysis (CCA) objectives, particularly their deep nonlinear variants, consistently outperform autoencoder-centric and distance-minimizing approaches for multi-view representation learning across diverse data types.
  • The full whitening constraints of CCA (decorrelating dimensions within each view) are critical for learning effective and discriminative features.
  • Combining canonical correlation with reconstruction loss, as in DCCAE, offers a robust strategy that often yields strong performance by balancing inter-view dependence maximization and intra-view information preservation.

Summary by alphaXiv.

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