Compare training and validation loss curves to tell overfitting from underfitting, and get targeted regularization and early-stopping advice.
准确率差距 >10% 或损失差距 >0.3 通常视为过拟合信号。
| Item | Formula / Note | Source / Standard |
|---|---|---|
| Generalisation gap | training loss falling while validation rises signals overfitting | Hastie, Elements of Statistical Learning |
| Diagnosis | compare training and validation curves and inspect learning curves | Machine learning practice |
| Mitigation | regularisation, early stopping, augmentation and smaller models help | Machine learning practice |