curriculum learning

Curriculum learning trains models on easier examples first, gradually increasing difficulty like human education. **Intuition**: Start with clear patterns, build up to complex cases. Avoids early confusion from hard examples. Better optimization trajectory. **Difficulty metrics**: Loss value (lower = easier), prediction confidence, human-defined complexity, data-driven scoring. **Strategies**: **Predetermined**: Fixed difficulty ordering based on metrics. **Self-paced**: Model selects examples it can currently learn. **Teacher-guided**: Separate model determines curriculum. **Baby Steps**: Multiple difficulty levels, progress when mastered. **Implementation**: Sort dataset by difficulty, start with easy subset, gradually expand, or weight examples by curriculum. **Benefits**: Faster convergence, better final performance on some tasks, more stable training. **Challenges**: Defining difficulty, computational overhead for scoring, may not help all tasks. **When most effective**: Noisy data (easy examples often clean), complex tasks with learnable substructure, limited training time. **Negative results**: Not always beneficial, random ordering sometimes competitive. Useful technique for specific scenarios requiring training stability.