THE HKPU, DEPARTMENT OF EIE
Lecture Notes of Machine Learning
Lecture 1 55y
January 12, 2015
1 SUPERVISED LEARNING
˜k˜‰´§x(i ) L«\C‰\A§y (i ) L«C‰
8I" (x(i ), y (i ) ¡Ø§|{(x(i ), y (i ));i = 1,2,...,m} ¡8§X L
«\C§Y L«C§5‘§•˜¢Œ"
The goal of supervised learning can be described as follows: given a training set, to
learn a function h : X → Y so that h(x) can predict the output value as close as the corre-
sponding value of y. The function h is called hypothesis.
˜u8I/“§Supervised Learning '⁄a.
Regression Problem££8flK⁄: C·ºY.
Classification Problem£'aflK⁄: C·l.
1.1 5£8
•k˜{5£8flK§•b\Cx ·3
m2§•k–eL“
hθ(x) = θ0 + θ1x1 + θ2x2
1
θi ·ºŒ (k¡)§øºŒ–\X 5NY. •\
Vg§-x0 = 1, ªL“–U⁄
hθ(x) = n
θi xi = θT x
i=0
y3§|km 8§•F"Ø•ºŒθ§ƒhθ(x)
8IƒUC"OºŒθ§•‰´Xecost function:
m
i=1
J(θ) = 1
2
(hθ(xi )− y i )2
(1)
ø…Œ…ŒØ§e5§•?Xƒ)ºŒθ"
1.2 LMS ALGORITHM
ƒ‘ºŒθ ƒ…Œ§•–^Fe{?
1ƒ)§ºŒθ #–L«⁄Xe
θj := θj − α
∂
∂θj
J(θ)
¥§α ¡Learning rate. •k…ŒŒ§{B§k˜k
Ø„(x, y).
∂
∂θj
J(θ) = ∂
∂θj
= 2· 1
2
(hθ(x)− y)2
1
2
(hθ(x)− y)
· (hθ(x)− y)· ∂
n
∂θj
θi xi − y)
(
= (hθ(x)− y)· ∂
∂θj
= (hθ(x)− y)x j
i=0
?§•–kØ#OK
θj := θj + α(y i − hθ(xi ))xi
j
(2)
ø·k¶LMS#K§Widrow-Hoff˘SOK§ºŒθ #ßu
" lØ„§•ºŒθ X#ƒ…Œ–´æ"
fl¢§Øu„kı„§k{–غŒθ ?1#§
·batch model§,·stochastic model"
2
batch model#K·zg#H{⁄k§Xe⁄«
Repeat until convergence
θj := θj − α
(y i − hθ(xi ))xi
j
m
i=1
{
for every j
}
stochastic modelz‹?1g#§Xe⁄«
Loop
{
{
for i = 1 to m,
θj := θj − α(y i − hθ(xi ))xi
}
j
for every j
}
l¡«“–w§batch model zg#H{⁄k§ø·
~§AO·3Œ~ı§stochastic model ·=#§z
#g§ˇ~„e§stochastic model ‹e/’batch model fl§
Lstochastic model"·kUˆ{´æ§·3NC5
£6˜§øV·¡stochastic model ˇ§“´æk‰¯
5§,,·´æNC§Øıø«Cqfi†~C
§stochastic model ˙p§⁄–stochastic model‹‘k˜
{"
ˇ~«{·(«“§8'⁄Øıbatch§,
|^stochastic model?1#§zbatch§?1g#§ø|
^stochastic modelp§3‰§~3NC6˜"
REFERENCES
[1] Andrew Ng, "Machine Learning", Stanford University.
3