文章目录

• P1 机器学习介绍
• • 找出Function
  • • Framework
    • • Function Set -- Model
      • Training Data
      • Test && Training
      • Framework Step
      • • Step1:
        • - define a set of function:找出一个function set
        • Step2:
        • - goodness of function:衡量一个function的好坏
        • Step3:
        • - pick the best function:有一个好的方法，一个自动的演算法挑出好的function
  • Learining Map
  • • Regression--回归
    • Classification--分类
    • • Binary Classification--二元分类
      • Multi-class Classification
    • Model分类
    • Supervised Learning
    • Semi-supervised Learning
    • Transfer Learning
    • Unsupervised Learning
    • Structured Learning - Beyond Classification
    • Reinforcement Leaning
• P2 为什么要学习机器学习
• Summary

P1 机器学习介绍

• AI：Artificial Intelligence

• 机器学习：让机器具有学习的能力

• 人工智慧：达成的目标，机器学习是手段

• 深度学习：是机器学习的其中一个方法

• Machine Learning = Looking for a Function From Data

  • 语音辨识（Speech Recognition
  • 影像辨识（Image Recognition
  • Playing Go
  • Dialogue System

• 我理解的机器学习：从数据中找到一个函数，对于未学习过的事物进行判定

找出Function

Framework

Function Set – Model

• 要准备一个function set – Model

Training Data

• function input:
• function output:

从Function中pick出一个best Function

• best Function–f stat

• 对于不在Training Data中的cat通过f start输出为cat

Test && Training

Framework Step

Step1:

- define a set of function:找出一个function set

Step2:

- goodness of function:衡量一个function的好坏

Step3:

- pick the best function:有一个好的方法，一个自动的演算法挑出好的function

Learining Map

Regression–回归

• The output of the target function f is ‘scalar’
• For example:predict PM2.5
  • 预测PM2.5利用今天上午PM2.5,昨天上午PM2.5,预测明天上午PM2.5
  • 通过Training Data

Classification–分类

• 与Regression的不同输出的结果类型不同

Binary Classification–二元分类

• Output:Yes or No

Multi-class Classification

• Output:Class 1,Class 2,…Class N

• Example
  • Binary Classification : **Spam filtering:**判断邮件是否为垃圾邮件，只需要给Training Data

• Example
  • Multi-Class Classification:Document Classification

Model分类

• Linear Model
• Non-linear Model
  • Deep Learning：complex function

• Playing Go:看成19*19个选项的选择题，通过Training Data进行选择

Supervised Learning

• Training Data:Input / output pair of target function
• Function output = label
• 机器学习需要大量的label，就是要对结果进行人工标记
• Hard to collect a large amount of labelled data

Semi-supervised Learning

• 可以减少data用量

• Unlabeled data：也可能对学习有帮助

Transfer Learning

Unsupervised Learning

• 无师自通
• 机器如何自己学会

Structured Learning - Beyond Classification

Reinforcement Leaning

• Supervised v.s. Reinforcement

  • Supervised：

    • Learning from teacher
    • 告诉了机器正确的答案是什么

  • Reinfocement：

    • Learning from critics
    • 没有告诉机器正确的答案是是什么
    • 需要比较强的intelligence

• Reinforcement的步骤：对手是另外一个机器

• 没有办法做Supervised Learning 才去做Reinforcement Learning

P2 为什么要学习机器学习

• AI训练师要为机器选择合适的Model和loss function适合解决不同的问题、

Summary

• discriminate different scenario including Supervised Learning、Semi-supervised Learning、Transfer Learning、Unsupervised Learning、Reinforcement Learning
• discriminate different Task including Regression 、Classification and Structured Learning
• use different method including Linear Model 、 Deep Learning 、 SVM、decision tree、K-NN
• discriminate Regression and Classification
  • Regression:Output是scalar
  • Classification: including Binary Classification and Multi-class Classification
    • Binary Classification:Yes or No
    • Multi-class Classification:Class 1,Class 2 ,Class N and so on like a selection