Biomedical Engineering Reference
In-Depth Information
FIgURE 5.9:
State representation of ICHMM vs. CHMM.
single neural channel, the number of parameters is greatly reduced and can support the amount of
training data. Specifically, the individual HMM chains in the ICHMM contain around 70 param-
eters for a training set of 10 000 samples as opposed to almost 18 000 parameters necessary for a
comparable FCHMM (due to the dependent states).
Each of the individual HMM chains is trained with the Baum-Welch algorithm as before,
but now directly from the input data. We compose the likelihood ratio for the decision by
D
∏
∏
( )
i
( )
i
P O
(
|
λ
)
T
M
l O
(
)
=
i
=
1
>
ζ
(5.60)
D
( )
i
( )
i
P O
(
|
λ
)
T
r
i
=
1
To give a qualitative interpretation of these weak classifiers, we present in Figure
5.10
the
probabilistic ratios from 14 single-channel HMM chains (shown between the top and bottom
FIgURE 5.10:
Example of probabilistic overlap.
