Electronic apparatus and method for generating trained model
A method for generating a trained model is provided. The method for generating a trained model includes: receiving a learning data; generating an asymmetric multi-task feature network including a parameter matrix of the trained model which permits an asymmetric knowledge transfer between tasks and a feedback matrix for a feedback connection from the tasks to features; computing a parameter matrix of the asymmetric multi-task feature network using the input learning data to minimize a predetermined objective function; and generating an asymmetric multi-task feature trained model using the computed parameter matrix as the parameter of the generated asymmetric multi-task feature network.
1. A method for generating an asymmetric multi-task feature trained model, comprising:
receiving a learning data;
generating an asymmetric multi-task feature network including a first parameter matrix of the trained model which permits an asymmetric knowledge transfer between tasks and a feedback matrix for a feedback connection from the tasks of the first parameter matrix to features of the first parameter matrix;
computing a second parameter matrix of the asymmetric multi-task feature network using the received learning data to minimize a predetermined objective function; and
generating an asymmetric multi-task feature trained model using the second parameter matrix as parameters of the generated asymmetric multi-task feature network,
wherein the feedback matrix is an additional weight matrix for the feedback connection along with an original feedforward connection.
2. The method as claimed in claim 1 , wherein the predetermined objective function includes a loss function for the trained model which permits the asymmetric knowledge transfer, an auto-encoder term using the feedback matrix and inducing a nonlinear combination of task parameters, and a weight decay regularization term.
3. The method as claimed in claim 1 , wherein the predetermined objective function is:
A
,
min
.
{
W
(
l
)
}
l
=
1
L
∑
t
=
1
T
(
1
+
α
a
t
o
1
)
{
ℒ
(
w
t
(
L
)
;
Z
t
,
y
t
)
+
ℒ
^
t
}
+
γ
σ
(
f
(
ZW
(
L
)
)
A
)
-
Z
2
2
+
λ
∑
l
=
1
L
W
(
l
)
F
2
,
,
wherein Z represents a matrix of the asymmetric multi-task feature network configured of a multi layer, W (L) represents a weight matrix for the last layer of the asymmetric multi-task feature network, A represents the feedback matrix, α t o ϵ 1×K represents a t-th row of an A vector, {circumflex over (L)} i represents an average validation loss of single task learning for a task t, t := (w t (L) , W (L−1) , . . . W (1):X t , y t ), σ represents a nonlinear function, and α, γ, λ each represent model parameters for adjusting weights of each term.
4. The method as claimed in claim 1 , wherein the asymmetric multi-task feature network includes a plurality of hidden layers.
5. An electronic apparatus, comprising:
a memory storing a learning data; and
a processor configured to:
generate an asymmetric multi-task feature network which has a first parameter matrix of a trained model permitting an asymmetric knowledge transfer between tasks and a feedback matrix for a feedback connection from the tasks of the first parameter matrix to features of the first parameter matrix,
compute a second parameter matrix of the asymmetric multi-task feature network using the stored learning data to minimize a predetermined objective function, and
generate an asymmetric multi-task feature trained model using the second parameter matrix as parameters of the generated asymmetric multi-task feature network,
wherein the feedback matrix is an additional weight matrix for the feedback connection along with an original feedforward connection.
6. The electronic apparatus as claimed in claim 5 ,
wherein the predetermined objective function includes a loss function for the trained model which permits the asymmetric knowledge transfer, an auto-encoder term using the feedback matrix and inducing a nonlinear combination of task parameters, and a weight decay regularization term.
7. The electronic apparatus as claimed in claim 5 , wherein the predetermined objective function is:
A
,
min
.
{
W
(
l
)
}
l
=
1
L
∑
t
=
1
T
(
1
+
α
a
t
o
1
)
{
ℒ
(
w
t
(
L
)
;
Z
t
,
y
t
)
+
ℒ
^
t
}
+
γ
σ
(
f
(
ZW
(
L
)
)
A
)
-
Z
2
2
+
λ
∑
l
=
1
L
W
(
l
)
F
2
,
,
wherein Z represents a matrix of the asymmetric multi-task feature network configured of a multi layer, W (L) represents a weight matrix for the last layer of the asymmetric multi-task feature network, A represents the feedback matrix, α i o ϵ 1×K represents a t-th row of an A vector, t represents an average validation loss of single task learning for a task t, t := (w t (L) , W (L−1) , . . . W (L) :X t , y t ), σ represents a nonlinear function, and α, γ, λ each represent model parameters for adjusting weights of each term.
8. The electronic apparatus as claimed in claim 5 , wherein the asymmetric multi-task feature network includes a plurality of hidden layers.
9. A non-transitory computer readable recording medium including a program for executing a method for generating an asymmetric multi-task feature trained model in an electronic apparatus, wherein the method for generating the asymmetric multi-task feature trained model comprises:
receiving a learning data;
generating an asymmetric multi-task feature network including a first parameter matrix of the trained model which permits asymmetric knowledge transfer between tasks and a feedback matrix for a feedback connection from the tasks of the first parameter matrix to features of the first parameter matrix;
computing a second parameter matrix of the asymmetric multi-task feature network using the received learning data to minimize a predetermined objective function; and
generating an asymmetric multi-task feature trained model using the second parameter matrix as parameters of the generated asymmetric multi-task feature network,
wherein the feedback matrix is an additional weight matrix for the feedback connection along with an original feedforward connection.