When training ML model need Orthogonalization in order to Determine what to tune, and observe the effect it has.
Each button does one thing.
Example: Car with Accelerator and angle of steering wheel.
Assumptions (controls for tuning):
- model works well with cost functions
- Try Adam Optimizer, bigger network
- work on training set
- Regularisation
- Try bigger training set
- works on test set of data
- Try bigger training set.
- works well in real life.
- Change training set
- Change cost function.
Avoid early stopping as effects network size, and training set size.
Related links: