Thanks to my supervisor
@mersault.bsky.social!

Paper: openreview.net/pdf?id=JJYHb...
Code: github.com/OlgaOvcharen...

Cross-modal errors span across multiple modaltities, like mismatched images or incorrect text fields, and are are an overlooked but critical issue in domains like e-Commerce and healthcare. Even with modern tools, large orgs struggle to catch cross-modal errors.

Thanks to all co-authors @flobarkmann.bsky.social, Philip Toma, Imant Daunhawer, Julia Vogt, @mersault.bsky.social, and @valboeva.bsky.social.

📄 Full paper: openreview.net/pdf?id=jnPHZ...
💻 Code: github.com/BoevaLab/scS...

📊 Key findings:
- Specialized single-cell frameworks and scFMs excel at uni-modal batch correction
- Generic SSL methods like VICReg & SimCLR outperform on cell typing & multi-modal tasks
- Masking emerges as the most universally effective augmentation strategy

📄 Read the full preprint: biorxiv.org/content/10.1...

Thanks to all co-authors Philip Toma, Imant Daunhawer, Julia Vogt, @flobarkmann.bsky.social, and @valboeva.bsky.social

Furthermore, we find that masking leads to the biggest performance improvement for creating augmented views of a cell.

Our bechmark shows that SimCLR and VICReg are the best methods for single-cell data across batch integration, query-to-reference, and missing modality prediction tasks.