机器学习 vs. 深度学习

1. bias/vairance

• Trend # 1：Scale driving Deep Learning process.

  
  
  

2. feature learning 的方式

• A common computer vision pipeline before 2012：
  
  • a. find interest points.
  • b. crop patches around them.
  • c. represent each patch with a sparse local descriptor.
  • d. combine the descriptors into a representation of the image.
• neural networks are trainable pipelines：
  
  • 神经网络的组织架构本身（trainable pipelines）就能帮助我们驾驭：
    
    • training data（$\left(x_n,y_n\right)_{n=1,\ldots,N}$）
    • parallel hardware