Curriculum learning progressively trains neural networks on increasingly complex hysteresis patterns, starting from simple linear relationships and advancing to full magnetic hysteresis behavior.
Reference: https://en.wikipedia.org/wiki/Curriculum_learning
Training Progression
The strategy involves learning simple patterns first, then gradually introducing complexity:
- Linear relationships - Curriculum Learning Stage 1
- Nonlinear responses - Curriculum Learning Stage 2
- Time-lag effects - Curriculum Learning Stage 3
- Weak hysteresis - Curriculum Learning Stage 4
- Full hysteresis - Curriculum Learning Stage 5
Implementation
Each stage builds upon the previous, with detailed implementations in individual stage notes. See Neural Network Training Strategies for broader context on training methodologies.
Future Development
- Train TFT on curriculum 5 and compare against curriculum learning [priority:: lowest] [completion:: 2025-09-01]
Further steps will be to add dynamic effects, and perhaps rate-dependent hysteresis. Finally we will try to transfer the model to measured data.
Related Concepts
- Neural Network Training Strategies - Broader training methodologies
- Magnetic Hysteresis - Underlying physical phenomenon
- Temporal Fusion Transformer - Primary model architecture
- Transfer Learning and Fine-tuning - Next step for real data application