The monograph brings in deep learning methods as a tool to model a large fraction of artificial intelligence information. These methods (for example) should support understanding of scenes and describe them in natural language. The text can be divided into two main parts. The first one contains an informal introduction of deep architectures, their parameter space and types of algorithms which are successfully used to some different AI problems. This part includes Section 2 where theoretical results are reviewed, and it is shown that insufficient depth can be detrimental for learning. Section 3 discusses the question of Local vs. Non-Local generalization.
The second part contains the description and the analysis of a particular family of algorithms, the Deep Belief Network, which are based on the Restricted Boltzmann Machine. The last section introduces potentially interesting open questions.
The text is written in a readable style and contains rich references. It is recommended to researchers, students and teachers.