Page 229 - Artificial Intelligence in the Age of Neural Networks and Brain Computing

220 CHAPTER 11 Deep Learning Approaches to Electrophysiological

1. INTRODUCTION

The processing of multivariate time-series for identification, classification, predic-
tion, or feature extraction problems is a relevant research and application topic
related to many different science domains. The development of information
technology and the prevalence of sensor networks for the emergence of the Internet
of Things (IoT) have strongly motivated a resurgence of interest in ML and its
potential impact on multivariate time-series analysis. DL techniques are a new trend
in ML, well founded in classical neural network theory. They use many hidden
layers of neurons to generate a lower-dimensional projection of the input space
that corresponds, for example, to the signals generated by the network of sensors
in monitoring applications. The successive hidden layers are able to build effective
high-level abstraction of the raw data. State-of-the-art DL processors present archi-
tectural advantages and benefit of novel training paradigms synergic with other
approaches, like compressive sensing and sparse representations. The high number
of neurons and links is reminiscent of brain networks and allows the storage of the
essential features of the underlying inputeoutput mapping. In biomedical signal
processing, many diagnostic systems produce multivariate time-series, and the
automatic extraction of features without human intervention is of high interest for
supporting clinical diagnoses and for highlighting latent aspects hidden to standard
visual interpretation. For example, in medical imaging, small irregularities in tissues
may be a precursor to tumors that can be detected in the successive levels of abstrac-
tions of DL network. The development of efficient DL systems can have a significant
impact on public health, also given the possibility of incorporating real-time
information in the existing computational models. In this chapter, DL methods are
briefly presented in the historical perspective of NN studies. Electroencephalo-
graphic (EEG) multivariate data are considered as many application domains
spanning from brain computer interface (BCI) to neuroscience take advantage of
this noninvasive and cheap technique as the basis of brain studies. Two different
DL architectures will be proposed that successfully solve difficult neurology
applications.

2. THE NEURAL NETWORK APPROACH

Through ML, computers develop the ability to autonomously learn and interact with
their environment. By exploiting the available data, they learn optimal behaviors
without the need of a specific programming step. NN are machines explicitly
designed to possess this ability. NNs are collection of elementary processing nodes
suitably arranged in various topological architectures. The elementary node of the
network is referred to as neuron and includes a linear part taking a weighted linear
combination of its inputs and a nonlinear part where a selected nonlinear function
transforms the input in the final output of the node. The inputs of the neuron