Playground

google順手架的類神經網路遊樂園區：playground.tensorflow.org

可以一眼看出以前機器學習和現在深度學習的差別
都是類神經網路但是...

Old-school

Few-layers, focus on trying out different features, use smooth activation such as sigmoid

特徵統統都給我輸進去！

有些特徵可以省略

Deep-learning

學特徵吧！

再試些不同的網路結構

Multiple layers, use raw-inputs, focus on trying different architectures, ReLu activation

Watch out for Overfitting!

用適當的regularization調味很重要

在這裡L1 regularization不如L2，而且值也太小