Page 235 - Artificial Intelligence in the Age of Neural Networks and Brain Computing

226 CHAPTER 11 Deep Learning Approaches to Electrophysiological

layer computes the dot product between the input image X and a set of K j learnable
filters. Each filter K j sized k 1 k 2 moves across the input space performing the
convolution with local subblocks of inputs, providing Y j , the feature maps
P
(Y j ¼ X K j þ B j , where B is the bias term). A Rectified Linear Unit (ReLU)
activation function is commonly applied to each feature map, in order to improve
computational efficiency by inducing sparsity and to reduce the effect of the vanish-
ing gradient problem [8,9]. The nonlinear convolutional layer is followed by the
pooling layer which performs a maximum or average subsampling of the feature
maps previously obtained in the previous step: a filter sized k 1 k 2 moves across
the input feature map taking the maximum (max pooling) or the average (average
pooling) of the neighbor values selected by the filter. Finally, the learned feature
maps are the input of a standard NN that performs classification tasks.

4. ELECTROPHYSIOLOGICAL TIME-SERIES

4.1 MULTICHANNEL NEUROPHYSIOLOGICAL MEASUREMENTS OF
THE ACTIVITY OF THE BRAIN
The brain generates bioelectromagnetic fields related to the activity and the synaptic
action of interacting neurons. This activity can be detected through sensors yielding
neurophysiological measurements like electroencephalography (EEG) and magne-
toencephalography (MEG).

4.2 ELECTROENCEPHALOGRAPHY (EEG)
The EEG collects the measurements of the brain’s electrical activity. It consists of
recording and displaying, over the time, the voltage difference between two scalp
sites: the location of interest and the “reference” location. Since its discovery in
1924 by Berger, EEG has become a routine examination in neurology [10] and
the basic neurophysiological measurement of many brain computer interface
(BCI) applications [11].
EEG electrodes are located at the surface of the scalp (Fig. 11.5) and collect the
electrical activity generated by networks of neurons. Extracellular current flow is
generated because of the excitatory and inhibitory postsynaptic potentials produced
by cell bodies and dendrites of pyramidal neurons. The EEG waveforms are mainly
produced by layers of pyramidal neurons whose synchronous activity produces a
bioelectromagnetic field that propagates from the sources to the recording scalp
electrodes. Fields propagate through tissues that have different conduction proper-
ties and overlap with the fields generated by other neuronal populations. As a result,
the potentials recorded at a specific electrode site will reflect the combination of the
contributions of different cortical sources (volume conduction effect). The issue of
volume conduction is not trivial in EEG analysis as it can result in apparently high
functional connectivity between channels thus leading to a wrong neurophysiolog-
ical interpretation of the results. It has long been discussed in the literature and