Information Technology Reference
In-Depth Information
Function.
Hessian(
Φ
,
G
,
a
β
,
b
β
)
Input
: mixing feature matrix
Φ
, mixing matrix
G
, mixing weight prior
parameters
a
β
,
b
β
Output
:(
KD
V
)
×
(
KD
V
) Hessian matrix
H
get
D
V
,K
from shape of
V
1
H
←
empty (
KD
V
)
×
(
KD
V
)matrix
2
for
k
=1
to
K
do
3
g
k
← k
th column of
G
4
for
j
=1
to
k −
1
do
5
g
j
← j
th column of
G
6
H
kj
←−
Φ
T
(
Φ
⊗
(
g
k
⊗
g
j
))
7
kj
th
D
V
× D
V
block of
H
←
H
kj
8
jk
th
D
V
× D
V
block of
H
←
H
kj
9
a
β
k
,b
β
k
←
pick from
a
β
,
b
β
10
−
g
k
))) +
a
β
k
H
kk
←
Φ
T
(
Φ
⊗
(
g
k
⊗
(1
b
β
k
I
11
k
th
D
V
× D
V
block along diagonal of
H
←
H
kk
12
return H
13
Function.
TrainMixPriors(
V
,
Λ
−
V
)
Input
: mixing weight matrix
V
, mixing weight covariance matrix
Λ
−
1
V
Output
: mixing weight vector prior parameters
a
β
,
b
β
get
D
V
,K
from shape of
V
1
for
k
=1
to
K
do
2
v
← k
th column of
V
3
(
Λ
−
1
V
D
V
block along diagonal of
Λ
−
1
V
)
kk
←
k
th
D
V
×
4
a
β
k
← a
β
+
D
2
5
b
β
k
← b
β
+
2
Tr
(
Λ
−
V
)
kk
+
v
k
v
k
6
a
β
,
b
β
←{a
β
1
,...,a
β
K
}, {b
β
1
,...,b
β
K
}
7
return a
β
,
b
β
8
The posterior parameters of the prior on the mixing weights are evaluated
according to (7.56), (7.57), and (7.70) in order to get
q
β
(
β
k
) for all
k
. Function
TrainMixPriors
takes the parameters of
q
V
(
V
) and returns the parameters for
all
q
β
(
β
k
). The posterior parameters are computed by iterating over all
k
,and
in Lines 5 and 6 by performing a straightforward evaluation of (7.56) and (7.57),
where in the latter, (7.70) replaces
E
V
(
v
k
v
k
).
8.1.4
The Variational Bound
The variational bound
(
q
) is evaluated in Function
VarBound
according to (7.96).
The function takes the model structure, the data, and the trained classifier and
mixing model parameters, and returns the value for
L
L
(
q
). The classifier-specific
Search WWH ::
Custom Search