李宏毅
- 2006, Restricted Boltzmann Machine. Complex. Used to initialize multi-layer perceptron (1980), to be called Deep Learning
- Network -> function set
- Deep -> many hidden layers
- 2015, Residual Net, 152 layers.
- Define the network structure <=> define a good function
- Non-deep model -> deep model <=> feature engineering -> network design. 换句话说，如果问题能很容易地抽象feature，不像video或者audio的问题中，那么可能deep model不一定是个好的选择。