Nonlinear Filters

8Deep Learning

8.1 Introduction

The Dartmouth conference in 1956 is regarded as the birth of artificial intelligence [136, 137]. A carefully selected group of scientists gathered together to work on the following topics:

• Automatic computers
• How can a computer be programmed to use a language
• Neural nets
• Theory of the size of a calculation
• Self‐improvement
• Abstraction
• Randomness and creativity

Since then, artificial intelligence has been increasingly deployed in different branches of science and engineering. Nowadays, learning algorithms have become part and parcel of daily life. Learning algorithms are categorized as [138]:

• Supervised learning refers to learning with a teacher and relies on labeled datasets. Teacher's knowledge is represented as input‐output examples.
• Unsupervised learning refers to learning without a teacher and relies on unlabeled datasets consisting of only input signals (stimuli) without the desired or target outputs (responses). The learning algorithm is tuned to the statistical regularities of the input data based on a task‐independent measure of the quality of the learned representation.
• Semi‐supervised learning relies on datasets that consist of labeled as well as unlabeled samples.
• Reinforcement learning refers to learning through continued interaction with the environment. The learning agent tries to optimize a scalar index of performance, which is the expected collected reward received from the environment.

Most of the machine‐learning algorithms ...