Curriculum Pseudo-Labeling is a semi-supervised learning strategy that progressively introduces pseudo-labeled samples in order of difficulty — starting with the most confident (easiest) predictions and gradually including less certain samples as the model improves.

How Does It Work?

• Easy First: Initially, only use pseudo-labels with very high confidence.
• Progressive Relaxation: As training progresses, lower the confidence threshold to include harder samples.
• Schedule: Threshold decreases linearly, cosine, or based on model performance metrics.
• Self-Paced: The curriculum naturally adapts to the model's learning stage.

Why It Matters

• Error Prevention: High-confidence-first avoids early training on incorrect pseudo-labels.
• Curriculum Learning: Follows the proven curriculum learning paradigm (easy to hard).
• Used In: FlexMatch, Dash, and other modern semi-supervised methods incorporate curriculum ideas.

Curriculum Pseudo-Labeling is learning from the easiest examples first — gradually building confidence before tackling harder unlabeled samples.