- goal is to find model with both low [[Bias and Variance]] 1. Achieve low bias - achieve high training data performance - if still can't even fit training data well, try: - bigger network - train longer - new NN architecture search 2. Achieve low variance - now that model can fit training data, check if you overfit the training data - if you did overfit, try: - more data - regularization - new NN architecture search (restart from step 1)