Biomedical Engineering Reference
In-Depth Information
TaBlE 3.4:
Best performing network
BEST NETwoRK
TRaININg
Run
85
Epoch
999
Minimum MSE
0.018601749
Final MSE
0.018616276
3.2.2.1 Echo-State Networks.
The improvement in performance of the RMLP is gained at the
expense of a significant leap in the computational costs of training. The problem is accentuated by
the fact that these models may have to be retrained occasionally to adjust to the changing statistics
and environment. This can severely restrict practical, real-world application of BMIs. We show here
a different architecture that has performance similar to that of an RMLP, but the training has linear
complexity as opposed to the (BPTT) RMLP training algorithm [
39
].
Echo state networks (ESNs) were first proposed by Jaeger [
42
] and are appealing because of
their computational abilities and simplified learning mechanisms. ESNs exhibit some similarities
to the “liquid state machines” proposed by Maas et al. [
43
], which possess universal approximation
80
60
40
20
0
-20
-40
0
50
100
150
200
250
300
350
400
450
500
Time (100ms)
FIgURE 3.14:
Testing performance for a RMLP for a reaching task. Here, the red curves are the de-
sired
x
,
y
, and
z
coordinates of the hand trajectory, whereas the blue curves are the model outputs.
