2.0 机器学习任务攻略
1 Framework of ML
2 General Guide – How to improve your model performance
–
2.1 training loss large: Model Bias
2.2 training loss small: Optimization
which one?: Model Bias v.s. Optimization Issue
可以看出,在训练集上, 56层network的training error甚至比20层network的还大,则说明56层network的optimization没有调好
那怎么确定是否是optimization的问题呢?
可以先用一些简单的模型,试试训练集上的error,再和复杂模型的training error对比
–
2.3 training loss small; testing loss large: overfitting
如何解决overfitting呢?
- more training data (not suggested)
- Data Augmentation
- constrain your model to be less complex
2.4 training loss small; testing loss large: mismatch
是否判断出现mismatch:需要我们对训练数据、测试数据有一定的了解,它们是如何产生的?
3 Bias - Complexity Trade-off
4 how to pick the right model
4.1 用Cross Validation选model
通过validation set的error挑选model