Reproduce “Disentangling Factors of Variation by Mixing Them” 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
Qiyang Hu, Attila Szabó, Tiziano Portenier, Matthias Zwicker, Paolo Favaro
A feature-chunk autoencoder learns to separate image variation without labels, paired examples, or domain-specific assumptions. By repeatedly mixing and unmixing latent chunks, enforcing realistic generated images, and requiring each chunk to leave a detectable visual signature, the method enables attribute transfer and improves Sprites retrieval mAP from 0.32 for a random baseline and 0.62 with mixing alone to 0.70 with the full objective.
Key insights
- The method treats disentanglement as invariance: a chunk representing one factor should remain usable when unrelated chunks are replaced.
- The three objectives serve different roles: mixing encourages factor separation, adversarial training improves realism, and classification prevents information from collapsing into only a few chunks.
- Disentanglement is partial rather than guaranteed; correlated attributes may share chunks, some chunks remain uninterpreted, and the number of chunks must be chosen beforehand.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.