Information Technology Reference
In-Depth Information
neuron of different CNN layer to
i
th neuron in current CNN layer,
x
*
j
(
t
) is the output
value of
j
th neuron in different CNN layer at time
t
.
g
is a sigmoid function.
When a MCNN is composed by two CNN layers, learning rules of the connections
between the CNN layers
W
*
ij
are described by Eq. (16) and Eq. (17).
(
⋅
)
cnn
ij
1
⋅
cnn
2
cnn
i
1
cnn
j
2
Δ
W
=
β
⋅
x
(
t
)
x
(t)
.
(16)
cnn
ij
2
⋅
cnn
1
cnn
i
2
cnn
j
1
(17)
Δ
W
=
β
⋅
x
(
t
)
x
(
t
)
.
Where β
is a parameter of learning rate, usually
β
=
1
/
m
,
m
is the number of stored
patterns.
Conventionally, we calculated the temporal change of the internal state of a CNN
layer
△
x
(
t
), and when
△
x
(
t
) is less than a threshold
θ
, the chaotic retrieval of the layer
is stopped by changing values of parameters
k
r
,
k
f
into zero, as a result, the CNN
layer becomes to a Hopfield model. The recalled pattern of one CNN layer provides
an input pattern to the other CNN layer, and the network realizes mutual association
and one-to-many retrieval for plural time series patterns [7] [8] [9].
2.2 An Amygdala-Hippocampus Model
Balkenius & Moren's computational amygdala model [3] [13], which is shown in the
right part of limbic system model in Fig. 1, is combined with hippocampus-neocortex
described above to evaluate and promote the performance of memory processing.
Recently, we proposed an amygdala-hippocampus model which showed faster storage
ability and higher precision of recollection of plural time series patterns dynamical
association [10]. The main idea of the adoption of emotional model comes from the
consideration that unstable state of hippocampus may result emotional response, i.e.,
arousal of amygdala, or the high value of amygdale output may enhance memory
processing happened in hippocampus.
The dynamics of the amygdale model is described as follows:
(18)
A
=
V
S
.
i
i
i
(19)
O
=
W
S
.
i
i
i
(20)
∑
−
∑
E
=
A
O
.
i
i
i
i
∑
(21)
Δ
V
=
α
(
S
max(
0
R
-
A
)
.
i
AMY
i
j
j
(
∑
(22)
Δ
W
=
β
S
(
O
−
R
)
)
.
i
AMY
i
j
j
