Machine Learning Foundations

(機器學習基石)

Lecture 2: Learning to Answer Yes/No

Hsuan-Tien Lin (林軒田)
[email protected]

Department of Computer Science

& Information Engineering
National Taiwan University
(國立台灣大學資訊工程系)

Hsuan-Tien Lin (NTU CSIE) Machine Learning Foundations 0/22

Learning to Answer Yes/No

Roadmap
1 When Can Machines Learn?

Lecture 1: The Learning Problem

A takes D and H to get g

Lecture 2: Learning to Answer Yes/No

Perceptron Hypothesis Set
Perceptron Learning Algorithm (PLA)
Guarantee of PLA
Non-Separable Data

2 Why Can Machines Learn?

3 How Can Machines Learn?
4 How Can Machines Learn Better?

Hsuan-Tien Lin (NTU CSIE) Machine Learning Foundations 1/22

Learning to Answer Yes/No Perceptron Hypothesis Set

Credit Approval Problem Revisited

Applicant Information
age 23 years
gender female
annual salary NTD 1,000,000
unknown target function
year in residence 1 year
f: X →Y
year in job 0.5 year
(ideal credit approval formula) current debt 200,000

training examples learning final hypothesis

D : (x1 , y1 ), · · · , (xN , yN ) algorithm g≈f
A
(historical records in bank) (‘learned’ formula to be used)

hypothesis set
H

(set of candidate formula)

what hypothesis set can we use?

Hsuan-Tien Lin (NTU CSIE) Machine Learning Foundations 2/22
Learning to Answer Yes/No Perceptron Hypothesis Set

A Simple Hypothesis Set: the ‘Perceptron’

age 23 years
annual salary NTD 1,000,000
year in job 0.5 year
current debt 200,000

• For x = (x1 , x2 , · · · , xd ) ‘features of customer’, compute a

weighted ‘score’ and
Xd
approve credit if wi xi > threshold
i=1
Xd
deny credit if wi xi < threshold
i=1


• Y: +1(good), −1(bad) , 0 ignored—linear
d
! ! h ∈ H are
formula
X
h(x) = sign wi xi − threshold
i=1

called ‘perceptron’ hypothesis historically

Hsuan-Tien Lin (NTU CSIE) Machine Learning Foundations 3/22
Learning to Answer Yes/No Perceptron Hypothesis Set

Vector Form of Perceptron Hypothesis

d
! !
X
h(x) = sign wi xi −threshold
i=1
 
d
!
X
= sign  wi xi + (−threshold) · (+1)
 
| {z } | {z }
i=1 w0 x0
d
!
X
= sign w i xi
i=0

T
= sign w x

• each ‘tall’ w represents a hypothesis h & is multiplied with

‘tall’ x —will use tall versions to simplify notation

what do perceptrons h ‘look like’?

Hsuan-Tien Lin (NTU CSIE) Machine Learning Foundations 4/22
Learning to Answer Yes/No Perceptron Hypothesis Set

Perceptrons in R2
h(x) = sign (w0 + w1 x1 + w2 x2 )

• customer features x: points on the plane (or points in Rd )

• labels y : ◦ (+1), × (-1)
• hypothesis h: lines (or hyperplanes in Rd )
—positive on one side of a line, negative on the other side
• different line classifies customers differently

perceptrons ⇔ linear (binary) classifiers

Hsuan-Tien Lin (NTU CSIE) Machine Learning Foundations 5/22

Learning to Answer Yes/No Perceptron Hypothesis Set

Fun Time
Consider using a perceptron to detect spam messages.
Assume that each email is represented by the frequency of keyword
occurrence, and output +1 indicates a spam. Which keywords below
shall have large positive weights in a good perceptron for the task?
1 coffee, tea, hamburger, steak
2 free, drug, fantastic, deal
3 machine, learning, statistics, textbook
4 national, Taiwan, university, coursera

Hsuan-Tien Lin (NTU CSIE) Machine Learning Foundations 6/22

Learning to Answer Yes/No Perceptron Hypothesis Set

Fun Time
Consider using a perceptron to detect spam messages.
Assume that each email is represented by the frequency of keyword
occurrence, and output +1 indicates a spam. Which keywords below
shall have large positive weights in a good perceptron for the task?
1 coffee, tea, hamburger, steak
2 free, drug, fantastic, deal
3 machine, learning, statistics, textbook
4 national, Taiwan, university, coursera

Reference Answer: 2
The occurrence of keywords with positive
weights increase the ‘spam score’, and hence
those keywords should often appear in spams.

Hsuan-Tien Lin (NTU CSIE) Machine Learning Foundations 6/22

Learning to Answer Yes/No Perceptron Learning Algorithm (PLA)

Select g from H
H = all possible perceptrons, g =?

• want: g ≈ f (hard when f unknown)

• almost necessary: g ≈ f on D, ideally
g(xn ) = f (xn ) = yn
• difficult: H is of infinite size
• idea: start from some g0 , and ‘correct’ its
mistakes on D

will represent g0 by its weight vector w0

Hsuan-Tien Lin (NTU CSIE) Machine Learning Foundations 7/22

Learning to Answer Yes/No Perceptron Learning Algorithm (PLA)

Perceptron Learning Algorithm

start from some w0 (say, 0), and ‘correct’ itsy= +1
mistakes on D w+y x

For t = 0, 1, . . . y= +1 w+y x

w
1 find a mistake of wt called xn(t) , yn(t)
x
  w
sign wTt xn(t) 6= yn(t)
y= −1 w
2 (try to) correct the mistake by
y=
w+−1
yx w x
wt+1 ← wt + yn(t) xn(t)
x
. . . until no more mistakes w+y x
return last w (called wPLA ) as g

That’s it!
—A fault confessed is half redressed. :-)

Hsuan-Tien Lin (NTU CSIE) Machine Learning Foundations 8/22

Learning to Answer Yes/No Perceptron Learning Algorithm (PLA)

Practical Implementation of PLA

start from some w0 (say, 0), and ‘correct’ its mistakes on D

Cyclic PLA
For t = 0, 1, . . .


1 find the next mistake of wt called xn(t) , yn(t)
 
sign wTt xn(t) 6= yn(t)

2 correct the mistake by

wt+1 ← wt + yn(t) xn(t)

. . . until a full cycle of not encountering mistakes

next can follow naïve cycle (1, · · · , N)

or precomputed random cycle
Hsuan-Tien Lin (NTU CSIE) Machine Learning Foundations 9/22
Learning to Answer Yes/No Perceptron Learning Algorithm (PLA)

Seeing is Believing

initially

worked like a charm with < 20 lines!!

(note: made xi  x0 = 1 for visual purpose)
Hsuan-Tien Lin (NTU CSIE) Machine Learning Foundations 10/22
Learning to Answer Yes/No Perceptron Learning Algorithm (PLA)

Seeing is Believing

update: 1

w(t+1)

x1

worked like a charm with < 20 lines!!

(note: made xi  x0 = 1 for visual purpose)
Hsuan-Tien Lin (NTU CSIE) Machine Learning Foundations 10/22
Learning to Answer Yes/No Perceptron Learning Algorithm (PLA)

Seeing is Believing

update: 2

w(t)

w(t+1)

x9

worked like a charm with < 20 lines!!

(note: made xi  x0 = 1 for visual purpose)
Hsuan-Tien Lin (NTU CSIE) Machine Learning Foundations 10/22
Learning to Answer Yes/No Perceptron Learning Algorithm (PLA)

Seeing is Believing

update: 3

w(t+1)

w(t)

x14

worked like a charm with < 20 lines!!

(note: made xi  x0 = 1 for visual purpose)
Hsuan-Tien Lin (NTU CSIE) Machine Learning Foundations 10/22
Learning to Answer Yes/No Perceptron Learning Algorithm (PLA)

Seeing is Believing

update: 4

x3

w(t)

w(t+1)

worked like a charm with < 20 lines!!

(note: made xi  x0 = 1 for visual purpose)
Hsuan-Tien Lin (NTU CSIE) Machine Learning Foundations 10/22
Learning to Answer Yes/No Perceptron Learning Algorithm (PLA)

Seeing is Believing

update: 5

w(t)
w(t+1)

x9

worked like a charm with < 20 lines!!

(note: made xi  x0 = 1 for visual purpose)
Hsuan-Tien Lin (NTU CSIE) Machine Learning Foundations 10/22
Learning to Answer Yes/No Perceptron Learning Algorithm (PLA)

Seeing is Believing

update: 6

w(t+1)

w(t)

x14

worked like a charm with < 20 lines!!

(note: made xi  x0 = 1 for visual purpose)
Hsuan-Tien Lin (NTU CSIE) Machine Learning Foundations 10/22
Learning to Answer Yes/No Perceptron Learning Algorithm (PLA)

Seeing is Believing

update: 7

w(t)

w(t+1)

x9

worked like a charm with < 20 lines!!

(note: made xi  x0 = 1 for visual purpose)
Hsuan-Tien Lin (NTU CSIE) Machine Learning Foundations 10/22
Learning to Answer Yes/No Perceptron Learning Algorithm (PLA)

Seeing is Believing

update: 8

w(t+1)

w(t)

x14

worked like a charm with < 20 lines!!

(note: made xi  x0 = 1 for visual purpose)
Hsuan-Tien Lin (NTU CSIE) Machine Learning Foundations 10/22
Learning to Answer Yes/No Perceptron Learning Algorithm (PLA)

Seeing is Believing

update: 9

w(t)

w(t+1)

x9

worked like a charm with < 20 lines!!

(note: made xi  x0 = 1 for visual purpose)
Hsuan-Tien Lin (NTU CSIE) Machine Learning Foundations 10/22
Learning to Answer Yes/No Perceptron Learning Algorithm (PLA)

Seeing is Believing

finally

wPLA

worked like a charm with < 20 lines!!

(note: made xi  x0 = 1 for visual purpose)
Hsuan-Tien Lin (NTU CSIE) Machine Learning Foundations 10/22
Learning to Answer Yes/No Perceptron Learning Algorithm (PLA)

Some Remaining Issues of PLA

‘correct’ mistakes on D until no mistakes

Algorithmic: halt (with no mistake)?

• naïve cyclic: ??
• random cyclic: ??
• other variant: ??

Learning: g ≈ f ?
• on D, if halt, yes (no mistake)
• outside D: ??
• if not halting: ??

[to be shown] if (...), after ‘enough’ corrections,

any PLA variant halts

Hsuan-Tien Lin (NTU CSIE) Machine Learning Foundations 11/22

Learning to Answer Yes/No Perceptron Learning Algorithm (PLA)

Fun Time
Let’s try to think about why PLA may work.
Let n = n(t), according to the rule of PLA below, which formula is true?
 
sign wTt xn 6= yn , wt+1 ← wt + yn xn

1 wTt+1 xn = yn
2 sign(wTt+1 xn ) = yn
3 yn wTt+1 xn ≥ yn wTt xn
4 yn wTt+1 xn < yn wTt xn

Hsuan-Tien Lin (NTU CSIE) Machine Learning Foundations 12/22

Learning to Answer Yes/No Perceptron Learning Algorithm (PLA)

Fun Time
Let’s try to think about why PLA may work.
Let n = n(t), according to the rule of PLA below, which formula is true?
 
sign wTt xn 6= yn , wt+1 ← wt + yn xn

1 wTt+1 xn = yn
2 sign(wTt+1 xn ) = yn
3 yn wTt+1 xn ≥ yn wTt xn
4 yn wTt+1 xn < yn wTt xn

Reference Answer: 3
Simply multiply the second part of the rule by
yn xn . The result shows that the rule
somewhat ‘tries to correct the mistake.’

Hsuan-Tien Lin (NTU CSIE) Machine Learning Foundations 12/22

Learning to Answer Yes/No Guarantee of PLA

Linear Separability
• if PLA halts (i.e. no more mistakes),
(necessary condition) D allows some w to make no mistake
• call such D linear separable

(linear separable) (not linear separable) (not linear separable)

assume linear separable D,

does PLA always halt?

Hsuan-Tien Lin (NTU CSIE) Machine Learning Foundations 13/22

Learning to Answer Yes/No Guarantee of PLA

PLA Fact: wt Gets More Aligned with wf

linear separable D ⇔ exists perfect wf such that yn = sign(wTf xn )

• wf perfect hence every xn correctly away from line:

yn(t) wTf xn(t) ≥min yn wTf xn > 0

n

• wTf wt ↑ by updating with any xn(t) , yn(t)

wTf wt+1 = wTf wt + yn(t) xn(t)

≥ wTf wt + min yn wTf xn

n
> wTf wt + 0.

wt appears more aligned with wf after update

(really?)
Hsuan-Tien Lin (NTU CSIE) Machine Learning Foundations 14/22
Learning to Answer Yes/No Guarantee of PLA

PLA Fact: wt Does Not Grow Too Fast

wt changed
only when mistake
⇔ sign wTt xn(t) 6= yn(t) ⇔ yn(t) wTt xn(t) ≤ 0

• mistake ‘limits’ kwt k2 growth, even when updating with ‘longest’ xn

kwt+1 k2 = kwt + yn(t) xn(t) k2

= kwt k2 + 2yn(t) wTt xn(t) + kyn(t) xn(t) k2
≤ kwt k2 + 0 + kyn(t) xn(t) k2
≤ kwt k2 + max kyn xn k2
n

start from w0 = 0, after T mistake corrections,

wTf wT √
≥ T · constant
kwf k kwT k

Hsuan-Tien Lin (NTU CSIE) Machine Learning Foundations 15/22

Learning to Answer Yes/No Guarantee of PLA

Fun Time
Let’s upper-bound T , the number of mistakes that PLA ‘corrects’.
wTf
Define R 2 = max kxn k2 ρ = min yn xn
n n kwf k
We want to show that T ≤ . Express the upper bound  by the two
terms above.
1 R/ρ
2 R 2 /ρ2
3 R/ρ2
4 ρ2 /R 2

Hsuan-Tien Lin (NTU CSIE) Machine Learning Foundations 16/22

Learning to Answer Yes/No Guarantee of PLA

Fun Time
Let’s upper-bound T , the number of mistakes that PLA ‘corrects’.
wTf
Define R 2 = max kxn k2 ρ = min yn xn
n n kwf k
We want to show that T ≤ . Express the upper bound  by the two
terms above.
1 R/ρ
2 R 2 /ρ2
3 R/ρ2
4 ρ2 /R 2

Reference Answer: 2
wt wT
The maximum value of kwf k kwtk
is 1. Since T
f
mistake corrections
√ increase the inner
product by T · constant, the maximum
number of corrected mistakes is 1/constant2 .
Hsuan-Tien Lin (NTU CSIE) Machine Learning Foundations 16/22
Learning to Answer Yes/No Non-Separable Data

More about PLA

Guarantee
as long as linear separable and correct by mistake
• inner product of wf and wt grows fast; length of wt grows slowly
• PLA ‘lines’ are more and more aligned with wf ⇒ halts

Pros
simple to implement, fast, works in any dimension d

Cons
• ‘assumes’ linear separable D to halt
—property unknown in advance (no need for PLA if we know wf )
• not fully sure how long halting takes (ρ depends on wf )
—though practically fast

what if D not linear separable?

Hsuan-Tien Lin (NTU CSIE) Machine Learning Foundations 17/22
Learning to Answer Yes/No Non-Separable Data

Learning with Noisy Data

unknown target function
f: X →Y
+ noise
(ideal credit approval formula)

training examples learning final hypothesis

D : (x1 , y1 ), · · · , (xN , yN ) algorithm g≈f
A
(historical records in bank) (‘learned’ formula to be used)

hypothesis set
H

(set of candidate formula)

how to at least get g ≈ f on noisy D?

Hsuan-Tien Lin (NTU CSIE) Machine Learning Foundations 18/22

Learning to Answer Yes/No Non-Separable Data

Line with Noise Tolerance

• assume ‘little’ noise: yn = f (xn ) usually
• if so, g ≈ f on D ⇔ yn = g(xn ) usually
• how about
N r
X z
wg ← argmin 6 sign(wT xn )
yn =
w
n=1

—NP-hard to solve, unfortunately

can we modify PLA to get

an ‘approximately good’ g?

Hsuan-Tien Lin (NTU CSIE) Machine Learning Foundations 19/22

Learning to Answer Yes/No Non-Separable Data

Pocket Algorithm
modify PLA algorithm (black lines) by keeping best weights in pocket

initialize pocket weights ŵ

For t = 0, 1, · · ·
1 find a (random) mistake of wt called (xn(t) , yn(t) )
2 (try to) correct the mistake by

wt+1 ← wt + yn(t) xn(t)

3 if wt+1 makes fewer mistakes than ŵ, replace ŵ by wt+1

...until enough iterations
return ŵ (called wPOCKET ) as g

a simple modification of PLA to find

(somewhat) ‘best’ weights

Hsuan-Tien Lin (NTU CSIE) Machine Learning Foundations 20/22

Learning to Answer Yes/No Non-Separable Data

Fun Time
Should we use pocket or PLA?
Since we do not know whether D is linear separable in advance, we
may decide to just go with pocket instead of PLA. If D is actually linear
separable, what’s the difference between the two?
1 pocket on D is slower than PLA
2 pocket on D is faster than PLA
3 pocket on D returns a better g in approximating f than PLA
4 pocket on D returns a worse g in approximating f than PLA

Hsuan-Tien Lin (NTU CSIE) Machine Learning Foundations 21/22

Learning to Answer Yes/No Non-Separable Data

Fun Time
Should we use pocket or PLA?
Since we do not know whether D is linear separable in advance, we
may decide to just go with pocket instead of PLA. If D is actually linear
separable, what’s the difference between the two?
1 pocket on D is slower than PLA
2 pocket on D is faster than PLA
3 pocket on D returns a better g in approximating f than PLA
4 pocket on D returns a worse g in approximating f than PLA

Reference Answer: 1
Because pocket need to check whether wt+1 is
better than ŵ in each iteration, it is slower than
PLA. On linear separable D, wPOCKET is the
same as wPLA , both making no mistakes.
Hsuan-Tien Lin (NTU CSIE) Machine Learning Foundations 21/22
Learning to Answer Yes/No Non-Separable Data

Summary
1 When Can Machines Learn?

Lecture 1: The Learning Problem

Lecture 2: Learning to Answer Yes/No
Perceptron Hypothesis Set
hyperplanes/linear classifiers in Rd
Perceptron Learning Algorithm (PLA)
correct mistakes and improve iteratively
Guarantee of PLA
no mistake eventually if linear separable
Non-Separable Data
hold somewhat ‘best’ weights in pocket
• next: the zoo of learning problems
2 Why Can Machines Learn?
3 How Can Machines Learn?
4 How Can Machines Learn Better?
Hsuan-Tien Lin (NTU CSIE) Machine Learning Foundations 22/22