Biomedical Engineering Reference
In-Depth Information
Therefore,
dw
i
(
t
)
dt
=
G
∞
0
dt
∞
−
∞
(11.26)
dt
K
t
)
t
)
∑
j
w
j
r
i
j
(
t
)
r
in
i
t
)
−
θ
(
P
(
t
−
(
t
+
r
in
(
A
+
−
A
−
)
.
For simplicity, let us assume that the scale of STDP is slow compared to the spiking
rate of the neurons, we can therefore replace
r
i
j
(
t
)
r
in
i
t
)
t
−
(
t
+
with its average
r
i
j
(
t
)
r
in
i
t
)
>
t
,or
Q
ij
(
−
t
−
t
)
t
−
t
)
<
is stationary
for all subsequent calculations. We can now calculate the rate of change in synaptic
strengths for correlated inputs. From Section 11.1,
t
−
(
t
+
. We assume that
Q
ij
(
−
C
ij
r
in
2
e
−|
t
|/
τ
r
in
2
(
)=
+
+
δ
(
)
,
Q
ij
t
c
t
r
in
(11.27)
ij
and
1
e
−
t
/
τ
K
(
t
)=
.
(11.28)
m
m
Upon collecting terms,
dw
i
r
in
w
i
A
+
τ
+
+
τ
Hr
in
2
r
in
2
)
)
dt
=
G
(
j
C
ij
w
j
−
(
A
−
−
A
+
)
j
w
j
+
m
τ
+
τ
(11.29)
m
−
θ
r
in
(
A
+
−
A
−
)
,
where
∞
0
1
A
+
τ
+
[
e
−
t
/
τ
dt
e
t
/
τ
+
e
−
(
t
+
t
)
/
τ
dt
=
{
H
m
c
−
t
0
m
−
t
∞
A
−
τ
−
dt
e
t
/
τ
+
e
t
+
t
)
/
τ
dt
e
−
t
/
τ
−
e
−
(
t
+
t
)
/
τ
+
]
−
}
c
c
−
∞
0
c
A
+
τ
A
+
τ
τ
+
A
+
τ
τ
+
c
c
=
)
−
)
+
(
τ
−
τ
+
)(
τ
+
τ
(
τ
−
τ
+
)(
τ
+
+
τ
(
τ
+
τ
+
)(
τ
+
+
τ
)
.
(11.30)
c
c
m
c
m
c
m
c
A
−
τ
−
(
τ
c
+
τ
−
)(
τ
c
+
τ
m
)
c
A
+
(
τ
−
τ
A
−
(
τ
+
τ
=
+
τ
−
)
[
A
+
−
A
−
+
c
)
−
c
)
(
τ
+
τ
)(
τ
+
τ
+
)(
τ
c
m
c
c
2
A
+
τ
τ
+
(
τ
+
τ
−
)
m
c
+
]
.
c
(
τ
+
+
τ
)
m
The first term in the brackets in Equation (11.29) corresponds to the effect of corre-
lations in the inputs, the second term corresponds to the effect of the average input,
and the last term takes into account the correlation introduced by spiking. It is inter-
esting that STDP can switch from a Hebbian to an anti-Hebbian rule depending on
the time scale of the input correlation. In order for the rule to be Hebbian,
H
has to
be greater than zero. Otherwise, it is anti-Hebbian. Solving
H
>
0 for
c
yields
−
x
2
x
−
4
(
A
−
−
A
+
)
y
<
,
(11.31)
c
2
(
A
−
−
A
+
)
where
2
A
+
τ
τ
+
m
x
=
A
+
τ
−
−
A
−
τ
+
+
m
,
(11.32)
τ
+
+
τ
Search WWH ::
Custom Search