NASA’s new dark energy space telescope can also detect killer asteroids
technologyreview.com·2h ago
AI Research
arXiv CS.AI·Jun 11·5 min read
CoVar: Confidence-Variance-Guided Pseudo-Label Selection for Semi-Supervised Learning
TL;DR
Semi-supervised learning often struggles with selecting reliable pseudo-labels, which are labels assigned to unlabeled data based on model predictions. CoVar, a new method, uses a confidence-variance approach to improve the selection of these pseudo-labels.
✦ Why It Matters
Engineers can implement CoVar to improve the reliability of pseudo-labels in their semi-supervised learning projects.
Key Takeaways
- Utilize CoVar for improved pseudo-label selection in semi-supervised learning.
- Integrate confidence and variance metrics for better label reliability.
- Apply SVD-based spectral relaxation to distinguish reliable predictions.
- Achieve competitive performance on standard classification benchmarks.
How It Works
CoVar operates by assessing pseudo-label reliability through a combination of Maximum Confidence (MC) and Residual-Class Variance (RCV). It starts with entropy minimization and derives a second-order cross-entropy approximation, favoring low-loss pseudo-labels when MC is high and RCV is low.
By embedding predictions into a two-dimensional space, it uses SVD-based spectral relaxation to separate reliable from unreliable predictions, which are then converted into training weights.