🍐Lecture 10

● Improve test error

Techno

Model assumpling：train multi model（不同初始化

Test time ： 取average of the prob

○ Early stopping, model ensembles

○ Regularization: dropout, data augmentation

Dropout：randomly drop activation

//Data Augmentation

可以：

Horizontal flip

○ Choosing hyperparameters

//LR maybe too low

//maybe try decay LR

//decaly LR too early

//nice let it continue

//可以加更多data；也可以加R（W）