Biomedical Engineering Reference
In-Depth Information
a) firing activity in circuit and readout
"one", speaker 3
"
o
ne", speaker 5
"five", speaker 1
"eight", speaker 4
135
90
45
0
1
2
3
4
5
6
7
8
9
10
0
0.2
0.4
0
0.2
0.4
0
0.1
0.2
0.3
0
0.1
0.2
time [s]
time [s]
time [s]
time [s]
b) 10 selected liquid states
class "one"
class "
other"
0.43
0.098
0.68
0.3
0.11
−0.6
−0.37
−0.01
−0.16
−0.21
135
135
90
90
45
45
1
1
1
2
3
4
5
6
7
8
9
10
state number
state number
c) weight vector
1
0
−1
1
45
90
135
Figure 18.5
Readout defined equivalence classes of liquid states.
a)
The firing activity of the
microcircuit for the speech recognition task from
Figure 18.4
is reexamined.
b)
The
liquid state
x
weight number
(
)
is plotted for 10 randomly chosen time points
t
(see arrowheads in
panel a). The target output of the readout neuron is 1 for the first 5 liquid states,
and 0 for the other 5 liquid states. Nevertheless the 5 liquid states in each of the 2
equivalence classes are highly diverse. But by multiplying these liquid state vectors
with the weight vector of the linear readout (see panel c), the weighted sums yields
the values shown above the liquid state vectors, which are separated by the threshold
0 of the readout (and by the firing threshold of the corresponding leaky integrate-and-
t
Search WWH ::
Custom Search