1 of 19

Machine Learning - �Basic Principles & Practice�12. Epilogue

Cong Li 李聪

机器学习 – 基础原理与实践

12. 尾声

2 of 19

What We’ve Learned �我们学了什么

Machine Learning – Basic Principles & Practice: 12. Epilogue

机器学习 – 基础原理与实践:12. 尾声

Popular science course 科普课程

A cross-section view of machine learning�机器学习的一个刨析视角

3 of 19

Inductive Learning 归纳学习

  • Nearest Neighbor 最近邻
  • Linear Classifiers 线性分类器
  • Learning w/ Margin 带间距的学习
  • Ensemble Learning 合奏学习
  • Complex Networks 复杂网络

Machine Learning – Basic Principles & Practice: 12. Epilogue

机器学习 – 基础原理与实践:12. 尾声

4 of 19

Real World Problems 真实问题

  • Handwritten ZIP Code Recognition�手写邮政编码识别
  • Word Sense Disambiguation�词义消歧

Machine Learning – Basic Principles & Practice: 12. Epilogue

机器学习 – 基础原理与实践:12. 尾声

We see performance improvement w/ multiple enhancements

随着各种改进,我们看到性能的提升

5 of 19

Principle 原理 (1)

Machine Learning – Basic Principles & Practice: 12. Epilogue

机器学习 – 基础原理与实践:12. 尾声

No assumption, no learning!�无假设,不学习!

Nearest neighbor 最近邻

Close neighbors come from the same class

相似的数据来自于同一类

6 of 19

Principle 原理 (1)

Machine Learning – Basic Principles & Practice: 12. Epilogue

机器学习 – 基础原理与实践:12. 尾声

No assumption, no learning!�无假设,不学习!

Perceptron or logistic regression 感知器或算术回归

Data are separable by a hyperplane

数据能被超平面分割

7 of 19

Principle 原理 (1)

Machine Learning – Basic Principles & Practice: 12. Epilogue

机器学习 – 基础原理与实践:12. 尾声

No assumption, no learning!�无假设,不学习!

Convolutional neural network�卷积神经网络

Local & high-level patterns are indicative�局部模式和高层次模式具有指示意义

8 of 19

Principle 原理 (2)

Machine Learning – Basic Principles & Practice: 12. Epilogue

机器学习 – 基础原理与实践:12. 尾声

Occam’s razor�奥卡姆剃刀

Make as fewer assumptions as possible

尽量少做假设

Simple models w/ good training performance can be powerful �在训练中表现出色的简单模型会很强大

9 of 19

Principle 原理 (2)

Machine Learning – Basic Principles & Practice: 12. Epilogue

机器学习 – 基础原理与实践:12. 尾声

Occam’s razor�奥卡姆剃刀

Not sure about assumption validity: use margins or improve training confidence

对于假设的合理程度无法确信时:

使用间距或改善训练的置信度

10 of 19

But There’s Another 但还有一个

Machine Learning – Basic Principles & Practice: 12. Epilogue

机器学习 – 基础原理与实践:12. 尾声

No free lunch: Everything comes at a cost!�没有免费的午餐:凡事皆有代价!

Nearest neighbor 最近邻

You need to define the distance

你必须确定距离

11 of 19

But There’s Another 但还有一个

Machine Learning – Basic Principles & Practice: 12. Epilogue

机器学习 – 基础原理与实践:12. 尾声

No free lunch: Everything comes at a cost!�没有免费的午餐:凡事皆有代价!

Eager learning 积极学习

What classifier to choose? A linear one?

选什么分类器?线性分类器吗?

12 of 19

But There’s Another 但还有一个

Machine Learning – Basic Principles & Practice: 12. Epilogue

机器学习 – 基础原理与实践:12. 尾声

No free lunch: Everything comes at a cost!�没有免费的午餐:凡事皆有代价!

Learning w/ margins 带间距的学习

How to determine the margin parameter?�怎么确定间距参数?

13 of 19

But There’s Another 但还有一个

Machine Learning – Basic Principles & Practice: 12. Epilogue

机器学习 – 基础原理与实践:12. 尾声

No free lunch: Everything comes at a cost!�没有免费的午餐:凡事皆有代价!

Ensemble learning 合奏学习

How to construct different classifiers?�怎么构建不同的分类器?

14 of 19

But There’s Another 但还有一个

Machine Learning – Basic Principles & Practice: 12. Epilogue

机器学习 – 基础原理与实践:12. 尾声

No free lunch: Everything comes at a cost!�没有免费的午餐:凡事皆有代价!

Kernel methods 核计算方法

What is the kernel? What is its parameter?�哪个核计算?参数选什么?

15 of 19

But There’s Another 但还有一个

Machine Learning – Basic Principles & Practice: 12. Epilogue

机器学习 – 基础原理与实践:12. 尾声

No free lunch: Everything comes at a cost!�没有免费的午餐:凡事皆有代价!

Complex networks 复杂网络

What is the network structure? How to determine the dropout rate?�网络结构是什么?信息遗失率取多少?

16 of 19

But There’s Another 但还有一个

Machine Learning – Basic Principles & Practice: 12. Epilogue

机器学习 – 基础原理与实践:12. 尾声

No free lunch: Everything comes at a cost!�没有免费的午餐:凡事皆有代价!

Making the right choice rewards you, e.g, CNN for USPS dataset: 3.29% error rate

正确的选择让你得到好的回报,例如卷积神经网络用于USPS数据集:3.29%的错误率

Nearest neighbor w/ tangent distance: 2.6%!!!

使用正切距离的最近邻:2.6%!!!

17 of 19

Popular Science Course 科普课程

  • Demonstrate Power of Methods�展示各种方法的威力
  • Not Tell How Choices Are Made �没告诉你这些选择是怎么作出的
    • This is a much more complex but important topic 这是一个复杂得多但也重要得多的课题
    • There is a line between scientific exploration & cheating 在科学探索和投机作弊间存在着一道界限

Machine Learning – Basic Principles & Practice: 12. Epilogue

机器学习 – 基础原理与实践:12. 尾声

18 of 19

Machine Learning 机器学习

  • A Hot Area Today 当下的热点
  • If You’re Interested 如果你感兴趣
    • You may learn much more �你有很多很多可学
    • You need to follow scientific methodology & engineering practice �你需要遵循科学方法论和工程规范

Machine Learning – Basic Principles & Practice: 12. Epilogue

机器学习 – 基础原理与实践:12. 尾声

19 of 19

The End