Why is calibration especially important when self-training selects pseudo-labels above a probability threshold?
Classic MLLevel: IntermediatePseudolabel confidence
Why is calibration especially important when self-training selects pseudo-labels above a probability threshold?
Hints
- The model supplies its own new supervision.
- Confidence errors can create a feedback loop.
Answer
Selection trusts reported confidence. An overconfident model can promote incorrect pseudo-labels into training and reinforce mistakes; verify calibration and held-out performance.
Source and licence: scikit-learn 1.7.2 — semi supervised · BSD-3-Clause · answer: Quantoros editorial