Page 242 - Artificial Intelligence in the Age of Neural Networks and Brain Computing

5. Deep Learning Models for EEG Signal Processing 233

(A) (B)
Deep Learning
Deep Learning
Feature Feature Data Other
Data-Driven
Engineering
Engineering Driven Features
Features
Features
Step Step Features Extractor

Classifica on
Classifica on
Step
Step
FIGURE 11.11
(A) Serial and (B) parallel DL schemes.

features form a unique input vector for the classification step. The serial step is
preferred for reducing the information redundancy of the features. In a recent
paper [30], a serial scheme based on SAEs has been proposed that includes a
time-frequency transformation of the input recordings, an intermediate step of
data-driven feature combination, and a final classification stage. The SAE model
has been proposed for discriminating EEGs of subjects affected by early stage
CJD from other forms of rapidly progressive dementia (RPD).
Each AE is implemented by an MLP-NN that includes an encoding stage fol-
lowed by a decoder. Higher-order features have been obtained by stacking two levels
of nonlinear (sigmoidal) nodes. The outputs of the deepest hidden layer is the final
feature vector used as input of the classification step. The processing chain is
depicted in Fig. 11.12. It is worth noting that while the “engineered” features are
extracted channel by channel, the higher-order features generated by the SAE are
mixing information pertaining to all of the channels. This procedure can be limited
to selected areas, that is, frontal or parietal channels, in order to gain information on
brain areas mostly relevant for the classification.
The output of each hidden layer of the SAE is given by:
h ¼ sðF xÞ;

where s is the node nonlinearity, for example, the standard sigmoidal function:
z 1

s ¼ 1 þ e .
F is the learned matrix of the encoding layer of the MLP-NN, and z ¼ F x. The loss/
cost function to be minimized through learning is given by:

L ¼
e x x
þ l F 2

2