Sharpening

Sharpening is a technique that reduces the entropy of a probability distribution by raising it to a power (lowering the temperature) — making confident predictions more confident and encouraging the model to commit to a single class rather than spreading probability mass.

How Does Sharpening Work?

• Temperature Scaling: $\text{Sharpen}(p, T)_i = p_i^{1/T} / \sum_j p_j^{1/T}$ where $T < 1$ sharpens.
• Low $T$: Approaches one-hot (argmax). High $T$: Approaches uniform. $T = 1$: No change.
• Typical $T$: 0.5 in MixMatch, 0.3-0.7 in general semi-supervised learning.

Why It Matters

• Entropy Minimization: Encourages the model to make confident predictions on unlabeled data (cluster assumption).
• MixMatch: Sharpening is a core component of MixMatch, applied to averaged pseudo-label predictions.
• Soft Labels: Unlike hard pseudo-labels, sharpened soft labels preserve uncertainty ranking among classes.

Sharpening is turning up the contrast on predictions — making the model's soft predictions crisper and more decisive.