Biomedical Engineering Reference
In-Depth Information
where
w
winner
is the winning filter's weight vector,
x
(
n
) is the present input,
e
winner
(
n
) is the error pro-
duced from the winning filter, η is the learning rate, and γ is the small positive constant. The data for
training is obtained from segments selected by the corresponding HMM (to avoid discontinuities,
one second of data is used to fill the FIR memory).
In terms of the final trajectory reconstruction, we see in Figure 5.8 that qualitatively the
bimodel system performs well in terms of reaching targets; this is especially evident for the first, sec-
ond, and the seventh peaks in the trajectory. Overall, prediction performance of the bimodel system
is approximately similar to the recurrent neural network (RNN), and superior to the single moving
average (MA) model (Wiener filter). The mean of the signal to error ratio (SER) averaged over all
dimensions for the single MA model, the RNN, and the bimodal system are -20.3 ± 1.6 (SD), -12.4
± 16.5, and -15.0 ± 18.8 dB, respectively, whereas the maximum SERs achieved by each model are
10.4, 30.1, and 24.8dB, respectively.
FIgURE 5.:
3D predicted (blue) and actual arm trajectory (dashed red) for (a) single MA model, (b)
RNN, and (c) bimodal system over the test data.
