Reimagining service delivery in the agentic era with Google Public Sector
cloud.google.com·23h ago
AI Research
arXiv CS.AI·Jun 11·5 min read
Geometric Erasure by Contrastive Velocity Matching in Rectified Flows
TL;DR
In machine learning, existing methods struggle with effectively erasing unwanted information from data distributions. The authors introduce a technique called Contrastive Velocity Matching, which operates within Rectified Flows to achieve geometric erasure.
✦ Why It Matters
Engineers can implement Contrastive Velocity Matching to enhance data privacy and model performance in their applications.
Key Takeaways
- Adopt GEM for effective content suppression in generative models.
- Utilize teacher-guided signals to enhance model safety.
- Combine attraction and repulsion signals for targeted erasure.
- Ensure benign content generation while removing harmful concepts.
How It Works
GEM operates by integrating trajectory-based unlearning with teacher-guided flow-matching. The teacher model provides dual signals: attraction to desired outputs and repulsion from unwanted concepts.
These signals are mathematically combined into a single geometric guidance objective, which directs the generative model to suppress harmful content effectively while preserving the generation of safe outputs.