1. Don’t use a neural net if there’s another way
See Sklearn machine learning roadmap for ideas
2. Don’t leak data across datasets,
...and also, Validation is not test
3. Plan for data preparation
Use the table in the data readiness article to communicate the state of readiness.
4. Capture a baseline model
More complex models may not actually beat a simpler ones, start small.
5. Hyperparameter rules of thumb
- Batchsize=32
- Start with single hidden layer, total nodes = halfway between input and output layers
6. Preprocessing rules of thumb
- Balance the data
- Use class weights
- Know your split and make sure accuracy is better than random guess
- Normalize the data - 0 to 1
7. General pipeline: Create model > Build/Compile > Fit > Evaluate [repeat]
Pay close attention to input and output shapes, these are where the most common errors lie.
8. Experiment, experiment, experiment
Change one thing at a time (e,g., add a layer, increase learning rate, not both) and record experimental results.
9. Visualize visualize visualize
Be creative and curious. Visualize input datasets, to understand idiosynchracies and also predictions to understand performance. For example, you may find, via visualization that mis-classified observations are really mislabeled observations.
10. Transfer learning when few samples are available
But remember - transfer learning improves predictive performance often by adding complexity, and can significantly boost prediction time.
11. Use online learning for continuous improvement
Train with new data as it comes online, CI/CD can help.
12. Keep the saw sharp on state of the art
- http://paperswithcode.com
- Conferences, like NeurIPS