February 14, 2022 [reading notes] - life science based on deep learning Chapter 2 Introduction to deep learning (Part 1)

• 
• The traditional method is to design a function by hand , Start by analyzing the problem , Recognize features by writing computer code , And reliably generate results ; Machine learning takes a completely different approach , There is no need to design functions manually , Instead, it allows the computer to learn its own functions from the data .
• Linear model （ perceptron ）
  • y=Mx+b
  • M It's a matrix / The weight ,b It's a vector
  • There are many limitations , Not suitable for most real datasets . When the data dimension becomes very high , The problem will get worse .
• Multilayer perceptron （MLP）
  • It can be simply understood as superposition of multiple linear transformations one by one .
  • Transfer a linear function to another nonlinear function （ Activation function φ(x)）
    • y=M2φ(M1x+b1)+b2
  • You can also stack any number of transformations together ：
    • h1=φ1(M1x+b1) 
    • h2=φ2(M2h1+b2) 
    • ...
    • hn-1=φn-1(Mn-1hn-2+bn-1) 
    • y=φn(Mnhn-1+bn)
  • At present, there are three popular activation functions ：
    • Correction of linear element function
    • Hyperbolic tangent function
    • sigmoid function
  • MLP The other two properties of ： Width and depth
    • Definition
      • Width ： Refers to the size of the hidden layer , Choose each hhhh The length of
      • depth ： Refers to the number of layers in the model
    • The choice of width and depth requires trying multiple combinations , There are some principles to guide ：
      • Containing a hidden layer MPL Is a universal approximator
      • The deep model requires fewer parameters than the shallow model
      • Deep models are often more difficult to train than previous models