Notes From Andrew NG - Nuts and Bolts of Applying DL
Training Error High?(Bias, e.g. human-level is 1% and Train error is 5%)
- Yes(try below and re-train):
- Bigger Model
- Train Longer
- New Model Architecture
- No
- Dev Error High?(Variance, e.g. HL is 1%, Train Error is 2% and Dev Error is 6%)
- Yes(try below and retrain):
- More Data
- Regularization
- New Model Arch
- No:
- Done