Basic recipe for machine learning: High bias? High variance?

How to check for High Bias, High Variance:

• High bias (underfit) means the model doesnt get trained enough
• High variance (overfit) means the model is overtrained

	Case 1	Case 2	Case 3	Case 4
Train error	1%	15%	15%	0.5%
Test error	11%	16%	30%	1%
	High variance: The model memorizes the training set. It performs badly on the test set	High bias: The model doesnt lean enough	High variance and High bias: The model doesnt learn enough, and performs badly on the test set	Low variance and Low bias: This is where we want the model to be

Treatments

• First, we check for high bias (underfit). If YES (high train error), we make bigger network, train more layers, etc, then retrain the model. If NO, we have low train error, we move to the next step.
• Next, we check for high variance (overfit). If YES (high test error), we add more data, dropout, regulations, etc, then retrain the model. If NO, we are done.