Bias - Variance
Bias-variance trade-off
training set error low
| Train set error | Dev Set Error | Type of Issue | solution | ||
|---|---|---|---|---|---|
| low （1%） | high （10%） | overfitting | high variance | regularization/cross validation | train more data in |
| high (15%) | hight (>15%) | underfitting | high Bias | bigger network | |
| low （0.5%) | low (1%) | fit | |||
当训练完算法、模型后，利用可视化工具区分出 bias - variance ,这将决定的优化算法性能的方向，和下一步的策略。