拼命三娘冲（20191120)_Why Neural Network?

Ø Problem 1: Non-linear classification

• If treat this problem with sigmoid function
  
• If we include enough polynomial terms, maybe we can get a hypothesis that separate the positive and negative examples.

Ø Problem 2: predict whether the house would be sold next year

Given 100 features of the house: .

• If only considering the quadratic terms
  
  # quadratic terms
  

• If only considering the subset of quadratic terms
  
  May not have the irregular shape we want, only constrained to ellipse in the above example.

• If only considering the cubic terms
  
  # cubic terms
  
  Ø Problem 3: computer vision for car detection

• Understanding computer vision problem for car detection as a classification problem using two pixels
  
  The nature of this problem is a classification problem.

• If take 50*50 images (whichis quite small),->2,500 pixels
  
  Using matrx of pixel intensity values that tell us the brightness of each pixel in the image(每个像素的亮度值矩阵)
  

Summary

Logistic regression works well when you have only 2 features, like in Problem 1, because you can include al the polynomial terms. But for many interesting machine learning problems, which will have a lot more features than just 2, simple logistic regression together with adding in the quadratic or the cubic features is not a good way to learn complex non-linear hypotheses when the number of features is large because you will just end up with too many features [1]. Neural Network is a good way to learn complex non-linear hypotheses, even when the input feature space is very large.

[1] too many features may cause: (1) overfitting; (2) computational expensive

参考视频: 8 - 1 - Non-linear Hypotheses (10 min).mkv