CS156 Machine Learning -1 The Learning Problem 笔记

大纲

example of machine learning

Predicting how a viewer will rate a movie

Movie rating - a solution

spectrum 范围
The essece of machine learning:

• A pattern exists
• We cannot pin it down mathematially.
• We have data on it.

The learning approah

Components of learning

Metaphor(比喻): Credit approval
Applicant information:

kind	info
age	23 years
gender	male
annual salary	$30,000
years in residence	1 year
years in job	1 year
current debt	$15,000
...	...

可认为这是一个d-维向量
其元素依次是 salary, years in residence, years in job, current debt...

y在这里仅表示 extend credit (1) & not to extend credit (-1)

It is a function from domain X,
X is a set of all input x (the set of vectors of d-dimention), it's a d-dimention Euclidean space(欧氏空间)
y : a binary co-demain
endeavors 尽力

f is unknown, but g is known, and we credit it
the value of g is supposed to approximates f

为什么需要hypothesis set

• no downside for including a hypothesis set in the formalization, but there is an upside

  • no downside:

    • no loss of generality in putting

  • upside:

    • the hypothesis set will play a pivotal role in the theory of learning, make the theory go through?

Solution components

quadratic programming 二次编程???
glorious 最好的,极好的
pinpoint 精确查找

A simple model

A simple hypothesis set - the `pereptron'

x1 salary, x2 years in residence, x3 years in job, x4 current debt ... xd ...

根据实际情况,分别给不同的权重
视为credit score
threshold 临界值

we start with random weights that will give a random line
notation 符号,记法

引入 x0 = 1 可以化简表达式

A simple learning algorithm - PLA

通过这种方法,我们尽可能的使这些点被正确分类

只要是线性的,当迭代次数足够多后,总能全部分类正确

Types of learning

Basic premise of learning

premise 前提
underlying process 基本过程

Supervised Learning

vending 贩卖

Unsupervised Learning

虽然没办法知道具体类别,但是可以做出分类

Reinforcement Learning

a learning puzzle

想这种样例少,为给出确定函数的,实际上根据不同规则是可以有不同答案的