Accelerating the search for antimicrobial peptides with Google Antigravity
deepmind.google·2h ago
AI Research
arXiv CS.AI·1d ago·5 min read
RINO: Renormalization Group Distillation with No Labels
TL;DR
RINO (Renormalization Group Distillation with No Labels) introduces a self-supervised learning approach for high-energy physics, achieving significant accuracy improvements in classifying particle jets. It matches fully supervised models with only 1% of labeled data and enhances out-of-distribution classification accuracy by up to 5.9%.
✦ Why It Matters
Consider adopting RINO for your next high-energy physics project to improve model accuracy with less labeled data.
Key Takeaways
- Explore self-supervised learning techniques for unlabeled data.
- Utilize RINO for efficient training with minimal labeled data.
- Implement hierarchical clustering for multi-scale data representation.
- Evaluate model performance on out-of-distribution datasets.
How It Works
RINO leverages the multi-scale structure of high-energy jet formation by defining teacher-student representations across energy scales. It uses hierarchical clustering to create these representations and aligns them through DINO-style self-distillation, which helps retain important scale-dependent features during training.