Spee h a d La guage Pro essi g
Le ture
Ba esia et ork a d Ba esia i fere e
I fo atio a d Co u i atio s E gi ee i g Cou se
Takahi o Shi ozaki
Le ture Pla Shi ozaki’s part
. / e ote
Spee h e og itio ased o GMM, HMM, a d N‐g a . / e ote
Ma i u likelihood esti atio a d EM algo ith
. / e ote
Ba esia et o k a d Ba esia i fe e e
. / @TAIST
Va iatio al i fe e e a d sa pli g . / @TAIST
Neu al et o k ased a ousti a d la guage odels . / @TAIST
Weighted fi ite state t a sdu e WFST a d spee h de odi g
Toda ’s Topi
•
A s e s fo the p e ious e e ises
•
Ba esia et o k
E er ise .
•
Sho the de i atio p o ess of o tai i g
K k k k K k km
L
1 11
log
,
μ
n
m
k k
a i izi g
E er ise .
•
De i e the ML solutio { , } of the Gaussia
dist i utio . The de i atio p o ess ust e
des i ed
| ,
0E er ise .
•
Gi e a t ai i g data
D
ith
n
t ai i g sa ples
D={x
1, x
2,…,x
n},
o tai ML esti atio fo GMM
ith
M
i tu es
You a assu e the a ia e is fo si pli it
N i M m m i m N i M m m i m X w X w M M M 1 1 2 , , 1 1 2 , , 2 exp 2 1 log max arg 2 exp 2 1 max arg ˆ 2 1 2 1 GMME er ise .
•
Assu e ou ha e a i itial odel pa a ete
Θ
0.
P o e that
if you take
0 0,
then the lower bound
0 0is equal to the log
likelihood
0
|
0
|
,
0
,
0
log
P
X
J
P
H
X
P
H
|
X
,
0,
0
log
P
X
|
0
J
H
q
H
X
P
H
q
q
J
H
log
,
|
E er ise .
•
Co side the ‐ i GMM of the p e ious page.
Let
Θ . O tai s the follo i gs
0
Graphs
•
U di e ted g aph
• A g aph defi ed odes a d u di e ted a s
•
Di e ted g aph
• A g aph defi ed odes a d a di e ted a s
•
Di e ted A li G aph: DAG
• Di e ted g aph that does ot o tai a di e ted le
E a ples:
U di e ted g aph Di e ted g aph
Pare t, Child, A estor, Des e da t
B A
Node A is a pa e t of ode B
Node B is a hild of ode A
B A
C
D
Node B, C, a d D a e des e da t of ode A Node A, B, a d C a e
Bipartite
Whe odes of a g aph a e sepa ated to t o g oups a d
the e is o a i side the g oups, it is alled a ipa tite
Dire ted Graph a d Node Orderi g
A di e ted g aph is a DAG
The e is a o de i g of odes he e all a s fa e the sa e di e tio =The e is a u e i g of odes he e all a s go f o a lo e
u e ed to highe u e ed odes E
1
Outli e of the Proof
•
State e t A:
• The e is a o de i g of odes he e all the a s fa e the sa e di e tio
•
State e t B:
• A g aph does ot o tai a di e ted le
A
B
A
B
…Eas
E er ise .
•
Is the di e ted g aph a DAG?
Ba esia Net ork BN
•
BN is a g aphi al odel that ep ese ts a set of
a do a ia les a d thei o ditio al i depe de e
DAG
A
C B
De o positio of Joi t Pro a ilit a d BN
•
B the p odu t ule, a it a joi t p o a ilit is
de o posed to a p odu t of o ditio al p o a ilities
•
A DAG is ade
• Rep ese ti g the a ia les as odes
• Co e ti g the odes di e ted a s a o di g the o ditio al p o a ilities
A P B A
P C A B
P D A B C
P D
C B A
P( , , , ) | | , | , ,
A
C B
Co ditio al I depe de e a d Ar s
•
Co ditio al i depe de e is ep ese ted
a se e
of a s
A P B A
P C A B
P D A B C
P D
C B A
P( , , , ) | | , | , ,
) | ( ) , , | ( ) ( ) | ( B D P C B A D P B P A B P
A P B P C A B
P D B
P D
C B A
P( , , , ) | , |
Joi t Pro a ilit Defi ed BN
•
P odu t of o ditio al p o a ilities asso iated ith
DAG al a s satisf the su ‐to‐o e o st ai t
Si e a Ba esia et o k is a DAG, ith a p ope o de i g of the a ia les, the p odu t has the follo i g fo
P oof:
|
, { 1, 2, , 1}1
i iN
i
i
i C C X X X
X
P
That is, Xi does ot appea i the o ditio al pa t of X ,…, Xi‐ . B thi ki g the su atio of the follo i g o de , e ha e:
1 1
1 2
1 |
|
| 1 1
1 X X X
N N
N N
X X X
E er ise .
•
Rep ese t the follo i g joi t p o a ilit a BN
|
|
|
( | ) ), , , ,
(A B C D E P A P B A P C B P D B P E D
BN Represe tatio of a Categori al Distri utio
X
X … K
BN Represe tatio of a Gaussia Distri utio
2 22 2
2
1
exp
2
1
,
|
X
X
N
BN Represe tatio of a GMM
Gauss1
μ ,σ
Gauss μ ,σ
Gauss μ ,σ
(
|
)
,
|
i
X
P
i
P
X
Gauss
i
P
X
P
i
i i
i
i
X
Mi tu e eight P o a ilit of i de
Gaussia dist i utio
Ba esia Net ork Represe tatio of a HMM
HMM
S
X
Ti e
• The et o k has u olled st u tu e
• The le gth depe ds o the i put se ue e
Ba esia et o k
X X XT
S S ST
T a sitio P o a ilit e.g. P(St=a|St-1=b)
E issio dist i utio o ditio ed the state
e.g. P(Xt|St=a)
E a ple of Alig
e t
Featu e se ue e: , , , , , , State se ue e: a, ,a,a,a, ,
S
X X X X
S S S X = , X = , X = , X = , X = , X = , X =
S =a, S = , S =a, S =a, S =a, S = , S =
Represe tatio of a Repeated Stru ture
X
X
X
X
TV
V
X
tReprese tatio of Para eters
s all i les ep ese t pa a ete s
i
X
Ni
i i
X
Gauss
i
P
X
P
1
,
|
1,2,3N
N
S
1,
2,
3
E er ise .
•
Fill the la ks so that the follo i g HMM a d the
BN e o e e ui ale t
a
S
X X X XT
S S ST
b b
a a
, , ,
Xt st st
N | ,
X a a
N | ,
P a P . . .
. . .
.
.
I itial state
p o a ilit P SS t|St‐ St =a St =
t‐ =a
St‐ = . I itial
state
HMM
BN
X b b
Fa tor Graph
• A ipa tite g aph he e o e side of a ia les ep ese t a do a ia les a d the othe s ep ese t fu tio s
• The a s ep ese t depe de ies of the fu tio s to the a ia les
• A fa to g aph defi es a joi t p o a ilit
X1, X2, X3, X4
f1
X1, X2, X3
f2 X2 f3 X3, X4
P
X X X X
f f f
E a ple:
Va ia le odes Fa to odes
siables of
subsets s
s
N
f
X
X
X
X
P
var 2
Fa tor Graph Represe tatio of Ba esia Net ork
Ea h o ditio al p o a ilit a e ega ded as a fa to
A
C B
D E a ple
A P B A
P C A
P D A C
P | | | ,
Ba esia et o k
A B C D
P A P B|A P C|A P D|A,C
Pro a ilisti I fere e
•
Ma gi al a d o ditio al p o a ilities a e o tai ed
f o a joi t p o a ilit appl i g the su a d
p odu t ules
A
B
C
P
,
,
A C
C B A P B P , , ,
C B C B A P A P , , ,
C C B A P B AP , , ,
Distri utio Propert a d Co putatio al Cost
•
P odu t is dist i uti e o e additio
•
The sa e p ope t holds fo su a d a , a d
p odu t a d a
Nx N
x
x
f
a
x
af
1 1
af
x
a
f
x
x
x
max
max
a
f
x
a
f
x
x
x
max
max
Nu e of p odu ts:N
Nu e of su atio :N‐
Nu e of p odu ts:
Co putatio al Cost of Margi alizatio
1000 1 1000 1 1000 1 , , ,B C D
D C B A P A P
Suppose A, B, C, a d D take
possi le alues
# su
atio =
=
If the joi t p o a ilit is de o posed to:
A,B,C,D
P(A|C)P(B)P(C)P(D)P
1000 1 1000 1 1000 1 1000 1 1000 1 1000 1 | , , , D B C B C DD P B P C P C A P D C B A P A P
# su
atio =
Whe the Fa tor Graph is Li ear
X X X X
f f f
Suppose e a t P X
4 5 2 1 4 5 1 21 2 4 5
1 , , 1 , , , , , , , , , , 5 4 4 4 3 3 2 1 1 3 2 2 5 4 4 4 3 3 3 2 2 2 1 1 5 4 4 4 3 3 3 2 2 2 1 1 3 X X X X X X X X
X X X X
Message Passi g Vie of the I fere e
X X X
f f f
X X f 3 3 X f 2 1 X f
f2X3
4 4 X f
4 5 2 1 1 , , 1 ,, 3 1 1 2 3 3 4 4 4 5
2 2 3 X X X X X X f X X f X X f X X f X P 2 2 f X
X4 f3
2
1 X
f
X2f2
Pro a ilisti Models a d Their Para eters
x
x
w
p
x
w
AM
p
w
LM
p
,
|
,
|
x w
LM
AM
x
|
p
Gaussia dist i utio ,
ML Trai i g a d Predi tio
x
x
x x xN
x
|
*
p
T ai i g set D
Test sa ple
n n
x p D
p | argmax |
max arg
*
Ba esia Approa h
•
T eat pa a ete s as a do a ia les
x
x x xN
Λ
p
x
p D
D p
p D
p x
p D
p D x p
D x
p | | | |
, ,
|
P edi tio of a e sa ple is fo ulated as a e aluatio of o ditio al p o a ilit gi e a t ai i g set
,D, X
P P D |
P X |
pDefi itio s of Ter s
•
A p io i dist i utio of pa a ete s
•
P o a ilisti odel
•
A poste io i dist i utio of pa a ete s
•
P edi ti e dist i utio
D p
p D
p D
p | |
p
x
D
p
x
p
D
p
|
|
|
x
|
E aluatio of A Posteriori Distri utio
•
E ept fo e si ple odels, ho to e aluate the
a poste io i dist i utio is a ig issue si e it
e ui es i teg atio s o e a a ia les
D
p
p
D
p
D
Approa hes
•
A al ti al e aluatio
•
Ideal, ut o l appli a le fo e si ple odels
•
Fo p a ti al odels, losed fo solutio is usuall
ot o tai ed. Nu e i al i teg atio is also ot
feasi le he the e a e a a ia les
•
Va iatio al Ba es
•
Ca e applied to la ge odels if p ope a al ti al
app o i atio is i t odu ed
•
Sa pli g
Co jugate Prior
•
Fo so e o i atio s of p io a d p o a ilisti odel,
poste io takes the sa e fu tio al fo as the p io
Pro a ilisti odel Co jugate prior
Bi o ial dist i utio Beta dist i utio
Multi o ial dist i utio Di i hlet dist i utio
E er ise .
• Assu es a p o a ilisti odel P x|μ , a t ai i g sa ple x , a d
a p io dist i utio of a pa a ete P μ a e gi e as follo s.
x x
μ
2 exp 21 2
P
2 exp 2 1| 2
x x P
1 1 1 | | x P P x P xP
x x
P
x
P x
dP | 1
| | 1
Esti ate poste io dist i utio
Esti ate p edi ti e dist i utio
c dx
cx
2
exp
Note:
Gaussia dist i utio ith ea μ a d a ia e
Gaussia dist i utio ith ea a d
Le
a .
If a g aph does ot o tai a di e ted le, the the e e ist
at least o e ode that has o i o i g a
Notatio for Co ditio al I depe de e
•
Let A,B, a d C e disjoi t sets of a do a ia les.
Whe the follo i g e uatio holds, e sa that A is
i depe de t of B gi e C, a d de ote it as A
╨
B|C
A
B
C
P
A
C
P
|
,
|
A
B
C
P
A
C
P
B
C
P
,
|
|
|
A B C
P A B C
P B C
P , | | , |
A
╨
B|C
A
╨
B|C
Graph Stru ture a d Co ditio al I depe de e
B i estigati g the g aph st u tu e, e a ead
elatio ships et ee a do a ia les
A
C B
D
A
╨
B|C
A
╨
D|C,B
?
Tail‐To‐Tail
C A B C A B
C C C P C B P C A P C B A P B AP , , , | |
I ge e al, P A,B is ot e p essed as
P A P B . The efo , A╨B|Φ does ot hold.
Φ is a e pt set
P A,B|C is e p essed as P A|C P B|C . The efo e A╨B|C holds.
A C
P B C
Head‐To‐Tail
C A B C A B Head Tail
C C A C P C B P A P C B A P B AP , , , | |
A C
P B C
P C P C B P A P A C P C P C B A P C B A P | | | | , , | ,
I ge e al, P A,B is ot e p essed as
P A P B . The efo , A╨B|Φ does ot hold.
Head‐To‐Head
C
A B
C
A B
Head Head
A B PA B C P A P B P C A B P A P B P
C C
, ,
| ,,
C P
A C P B P A P C
P
C B A P C
B A
P , | , , |
I ge e al, P A,B is e p essed as P A P B . The efo , A╨B|Φ holds.
Blo ki g a Path
•
Fo a Ba esia et o k, let A a d B e a ode, a d C e a
set of odes that does ot i lude A a d B. We sa a path
f o A to B is lo ked he eithe of the follo i gs holds
• O the path f o A to B, the e is a ode i C a d the o e tio of the a s is tail‐to‐tail o head‐to‐tail
• At o e of the odes o the path f o A to B, the o e tio of the a s is head‐to‐head. I additio , the ode a d its all
des e da ts a e ot i luded i C
D A
B
C
C F
E
Blo k
d‐separatio
Fo a Ba esia et o k, let A, B, a d C e e lusi e
sets of odes
•
We sa A is d‐sepa ated f o B C if all the paths
sta ti g f o a ode i A a d e di g at a ode i B is
lo ked
•
Whe A is d‐sepa ated f o B C, A
╨
B|C holds fo
Ma i izatio of Joi t Pro a ilit
•
O tai ed epla i g Σ i the su ‐p odu t
algo ith ith a
Ma ‐p odu t Algo ith
N
X X
X
N
P
X
X
X
X
X
X
N
,
,
max
arg
,
,
,
1 2, , , 2
1
Su ‐Produ t Algorith for Tree
•
Message passi g
• Leaf odes
• Va ia le ode:
• Fa to ode:
• Va ia le ode to fa to ode:
• Fa to ode to a ia le ode:
•
Ma gi al p o a ilit
1f x x
x
f
x
x
f
i x f fx
x
ix
M m m f x x x M xf x f x x x m x
EM for HMM a d Effi ie
,0
HMM
K | X ,0
logHMM
K, X |
Q
K
a
•• TT x KX
•
Di e tl e u e ati g all the paths is i possi le
Q‐fu tio a e effi ie tl e aluated if poste io s a d a e o tai ed, he e s’ a d s a e HMM state ID
k s | X,0
P t
k 1 s ,'k s| X,0
P t t
T
k k
k
K 1, 2,, X x1,x2,,xT