The project was done in collaboration with Michele Cafagna For more information please read the report.
- Randomly initialised weights
- Try a minimum number of initial configurations (e.g. 10 test frameworks)
- The test examples must be provided randomly
- Momentum
- Check the learning curve (draw it with each test)
- Use average gradients (eta/l), especially if using mini batch NN-part2 slide 25
- Optimisation
- Stopping criteria, in any case stop after too large a number of epochs
- Regularisation
- Use heuristics (slide)