FreeMatch is a semi-supervised learning algorithm that uses a self-adaptive global threshold and class-specific thresholds — automatically adjusting confidence thresholds based on the model's learning status without any fixed hyperparameter for the threshold.

How Does FreeMatch Work?

• Self-Adaptive Threshold (SAT): $ au_t = lambda cdot au_{t-1} + (1-lambda) cdot frac{1}{B}sum_b max(p_b)$ (EMA of model confidence).
• Class-Fairness: Per-class threshold adjustment based on class-specific confidence statistics.
• No Fixed $ au$: Unlike FixMatch's fixed $ au = 0.95$, FreeMatch's threshold adapts to the model's current state.
• Paper: Wang et al. (2023).

Why It Matters

• Hyperparameter-Free: Removes the need to tune the critical confidence threshold hyperparameter.
• Adaptive: Early in training (low confidence), threshold is low. Late in training (high confidence), threshold is high.
• Robust: Works well across different datasets and label amounts without threshold tuning.

FreeMatch is FixMatch that tunes itself — automatically adapting the confidence threshold based on model's evolving capability.