Agentic data science with the Data Agent Kit
cloud.google.com·6h ago
AI Research
arXiv CS.AI·27d ago·5 min read
Follow the Geometry, Not the Model: Cold Start Semi-Supervised Learning
TL;DR
VAST (Veracity-Aware Semi-Supervised Training) decouples pseudo-label generation from classifier training, addressing cold-start challenges in semi-supervised learning. It outperforms traditional graph-based methods across multiple datasets, achieving significant performance gains.
✦ Why It Matters
Implement VAST to improve your semi-supervised learning models, especially in cold-start situations.
Key Takeaways
- Use VAST for better performance in cold-start scenarios.
- Leverage the Veracity Matrix for effective label aggregation.
- Implement Veracity Propagation to enhance label coverage.
- Avoid reliance on end-to-end confidence-gated SSL methods.
How It Works
VAST operates by first using a frozen self-supervised embedding to infer probabilistic beliefs about unlabeled data. This geometric approach allows for the construction of the Veracity Matrix, which aggregates evidence of labels across the data manifold.
After establishing these beliefs, they are distilled into an inductive classifier, effectively decoupling the pseudo-label generation from the classifier training process.