Definitely not for the faint-hearted, this book does not waste any time explaining the basic concepts in machine learning; rather, it assumes that the reader is already moderately knowledgeable in the area. To make things worse, the (understandable) abundance of acronyms and external references throughout the text makes it even harder for anyone unfamiliar with the terminology and the state of the art to follow. However, it does provide a brief history about how deep learning has matured since Hinton’s pivotal publications in 2006 [1,2].

What’s nice about this monograph is that it introduces each “deep learning” method/architecture through an example application. These example applications are chosen from the published works that popularized the method discussed, in a particular context. For example, the deep autoencoders (a special type of deep neural networks) are introduced in the context of extracting binary speech codes from the raw speech spectrogram data. Similarly, there are case studies of applications of deep learning in the areas of audio, natural language modeling and processing, information retrieval, object recognition and computer vision, and multi-modal and multi-task learning.

I found of particular interest the chapter on “Selected Applications in Object Recognition and Computer Vision,” where recent successful applications of convolutional neural networks on natural images are analyzed. Over the years, the traditional approach used in computer vision was to use features extracted using scale invariant feature transform (SIFT), histograms of oriented gradients (HOG), and so on, followed by some kind of high-level feature building that then became the input to some trainable classifier, for example support vector machines. A key observation in this traditional approach is that the features are not learned but are rather handcrafted by the researchers. This has worked quite well, especially in recent years with the introduction of SIFT, speeded-up robust features (SURF), HOG, and so on; in fact, one can argue that the quality of the features often determines the success of the entire algorithm. This chapter describes how convolutional neural networks can be used to create a hierarchy of trainable feature extractors (at each layer), which “learn” the features directly from the images; instead, each layer extracts the features from the output of the previous layer. The chapter continues with a detailed description of the architecture of the deep model of the neural network that won the ImageNet 2012 object recognition competition.

This is definitely not a textbook for class, but rather a reference book with example applications tailored for researchers and enthusiasts in machine learning interested in learning the details (and finding references) on how to implement and apply deep learning.

On a side note, after reading the monograph, I am left wondering whether there is anything new here other than the context of the applications. Convolutional neural networks were proposed in the late 1980s [3] for digit recognition, but failed to launch in other areas due to the computationally intensive calculations. It is evident that the recent advancements in hardware, and in particular programmable graphics processing units (GPUs) and high-performance computing, have resulted in the resurfacing of these methods to make these calculations more tractable, which begs the question: What other methods that were proposed in the past and were ahead of their time should we revisit next?