SoftMatch is a semi-supervised learning algorithm that replaces the hard confidence threshold with a soft, continuous weighting function — assigning a sample weight between 0 and 1 based on confidence, rather than the binary keep/discard decision used in FixMatch.

How Does SoftMatch Work?

• Weight Function: $w(x) = exp(- ext{confidence\_deviation}^2 / 2sigma^2)$ (Gaussian weighting).
• No Threshold: Instead of $mathbb{1}[max(p) > au]$ (hard), use a smooth weight $w(x) in [0, 1]$.
• Truncation: Optionally truncate weights below a minimum to completely ignore very uncertain samples.
• Paper: Chen et al. (2023).

Why It Matters

• Soft Transition: No abrupt cutoff at $ au$ — samples near the threshold contribute partially.
• More Data: Moderate-confidence samples contribute to learning instead of being discarded entirely.
• Stability: Smoother loss landscape -> more stable training dynamics.

SoftMatch is the gentle version of FixMatch — using smooth weights instead of hard thresholds to extract value from all confidence levels.