Lecture 3:Types of Learning

2018-08-25 本文已影响0人薛家掌柜的

Learning with Different Output Space Y

Multiclass Classification

例如：

written digits $\Rightarrow$ 0,1,...,9
pictures $\Rightarrow$ apple,orange,strawberry
email $\Rightarrow$ spam,primary,social,promotion,update

Regression

例如：

patient features $\Rightarrow$ how many days before recovery
company data $\Rightarrow$ stock price
climate data $\Rightarrow$ temperature

Structured Learning

sentence $\Rightarrow$ structure(class of each word)
protein data $\Rightarrow$ protein folding
speech data $\Rightarrow$ speech parse tree
Mini summary
binary classification: $y = \{-1,+1\}$
multiclass classification: $y = \{1,2,3,...,k\}$
regression: $y = R$
structured learning: $y$ = structures

Learning with Different Data Label $y_n$

Supervised Learning(every $x_n$ with corresponding $y_n$ )

Unsupervised Learning( without $y_n$ ,clustering,a challenge but useful problem)

articles $\Rightarrow$ topics
consumer profiles $\Rightarrow$ consumer groups
unsupervised learning :diverse,with possibly very different performance goals

Semi-supervised(with some $y_n$ )

• face images with a few labeled ⇒ face identifier (Facebook)
• medicine data with a few labeled ⇒ medicine effect predictor
semi-supervised learning: leverage unlabeled data to avoid ‘expensive’ labeling

Reinforcement Learning(a very 'different' but natural way of learning)

learning with 'partial/implicit information'(often sequentially)
比如说，你教小狗‘sit down’，但是它并不会真的‘sit’，但是它会学到‘sit is good’。
因此，可以说，reinforcement: implicit $y_n$ by goodness( $\tilde y$ )
Mini summary
supervised learning:all $y_n$
semi-supervised:some $y_n$
unsupervised:no $y_n$
reinforcement learning:implicit $y_n$ by goodness $\tilde y$

Learning with Different Protocol $f ⇒ (x_n,y_n)$

Batch Learning(Batch Supervised multiclass classification:learning from all known data)

batch of (email,spam?) $\Rightarrow$ spam filter
batch of(patient,cancer?) $\Rightarrow$ concer classifier
batch of patient data $\Rightarrow$ group of patients

Online Learning(hypothesis 'improves' through receiving data instances sequentially)

online spam filter,with sequentially:

observe an email $x_t$
predict spam status with current $g_t(x_t)$
receive 'desired label $y_n$ ' from user,then update $g_t$ with( $x_n,y_n$ )

Active learning(Learning by 'Asking')

当输入一个 $x_n$ 时，算法不知道对应的 $y_n$ ，然后向目标函数 $f$ query the $y_n$ of chosen $x_n$ ，这就是Active learning :improve hypothesis with fewer labels(hopefully)by asking questions strategically
Mini summary
batch:all known data
online:sequence(passive)data
active:strategically-observed data

Learning with Different Input Space X

Concrete Features(each dimension of $x \subseteq R^d$ represents 'sophisticated physical meaning')

(size,mass)for coin classification
customer info for credit approval
patient info for cancer diagnosis

Raw Features(often need human or machines to convert to concrete ones)

Digit Recognition Problem:features $\Rightarrow$ meaning of digit(simple physical meaning,thus more difficult for ML than concrete features)
image pixels
speech signal

Abstract Learning(again need features'conversion/extraction/construction')

rating prediction problem
student ID in tutoring system
advertisement ID in online ad system
Mini summary
concrete:sophisticated(and related) physical meaning
raw:simple physical meaning
abstract:no(or little)physical meaning

Lecture 3:Types of Learning

Learning with Different Output Space Y

More Binary Classification Problems

Multiclass Classification

Regression

Structured Learning

Learning with Different Data Label $y_n$

Supervised Learning(every $x_n$ with corresponding $y_n$ )

Unsupervised Learning( without $y_n$ ,clustering,a challenge but useful problem)

Semi-supervised(with some $y_n$ )

Reinforcement Learning(a very 'different' but natural way of learning)

Learning with Different Protocol $f ⇒ (x_n,y_n)$

Batch Learning(Batch Supervised multiclass classification:learning from all known data)

Online Learning(hypothesis 'improves' through receiving data instances sequentially)

Active learning(Learning by 'Asking')

Learning with Different Input Space X

Concrete Features(each dimension of $x \subseteq R^d$ represents 'sophisticated physical meaning')

Raw Features(often need human or machines to convert to concrete ones)

Abstract Learning(again need features'conversion/extraction/construction')

猜你喜欢

热点阅读

Lecture 3:Types of Learning

Learning with Different Output Space Y

More Binary Classification Problems

Multiclass Classification

Regression

Structured Learning

Learning with Different Data Label

Supervised Learning(every with corresponding )

Unsupervised Learning( without ,clustering,a challenge but useful problem)

Semi-supervised(with some )

Reinforcement Learning(a very 'different' but natural way of learning)

Learning with Different Protocol

Batch Learning(Batch Supervised multiclass classification:learning from all known data)

Online Learning(hypothesis 'improves' through receiving data instances sequentially)

Active learning(Learning by 'Asking')

Learning with Different Input Space X

Concrete Features(each dimension ofrepresents 'sophisticated physical meaning')

Raw Features(often need human or machines to convert to concrete ones)

Abstract Learning(again need features'conversion/extraction/construction')

猜你喜欢

热点阅读

Learning with Different Data Label $y_n$

Supervised Learning(every $x_n$ with corresponding $y_n$ )

Unsupervised Learning( without $y_n$ ,clustering,a challenge but useful problem)

Semi-supervised(with some $y_n$ )

Learning with Different Protocol $f ⇒ (x_n,y_n)$

Concrete Features(each dimension of $x \subseteq R^d$ represents 'sophisticated physical meaning')