11. Improving your model Performance

supervised learning

 

1. fit training set (=avoidable bias)

- train bigger model(number of units, layers, etc...)

- train longer

- train with beteer optimization algorithm(momentum, RMSprop, Adam etc...)

- change NN architecture(RNN to CNN etx...)

- hyperparameter search

 

2. generalize to the dev/test set(=down variance)

- More data

- Reularization(L2, Dropout, data augumentation)

- change NN architecture

- hyperarameter serach