Chapter 1, Data science - Bird's-eye view, explains that data science or machine learning is the process of giving the machines the ability to learn from a dataset without being told or programmed. For instance, it will be extremely hard to write a program that takes a hand-written digit as an input image and outputs a value from 0-9 according to the number that's written in this image. The same applies to the task of classifying incoming emails as spam or non-spam. To solve such tasks, data scientists uses learning methods and tools from the field of data science or machine learning to teach the computer how to automatically recognize digits by giving it some explanatory features that can distinguish each digit from another. The same for the spam/non-spam problem, instead of using regular expressions and writing hundred of rules to classify the incoming emails, we can teach the computer through specific learning algorithms how to distinguish between spam and non-spam emails.
Chapter 2, Data Modeling in Action - The Titanic Example, linear models are the basic learning algorithms in the field of data science. Understanding how a linear model works is crucial in your journey of learning data science because it's the basic building block for most of the sophisticated learning algorithms out there, including neural networks.
Chapter 3, Feature Engineering and Model Complexity – Titanic Example Revisited, covers model complexity and assessment. This is an important towards building a successful data science system. There are lots of tools that you can use to assess and choose your model. In this chapter, we are going to address some of tools that can help you to increase the value of your data by adding more descriptive features and extracting meaningful information from existing ones. We are also going to address other tools related to optimal number features and learn why it's a problem to have a large number of features and fewer training samples/observations.
Chapter 4, Get Up and Running with TensorFlow, gives an overview of one of the most widely used deep learning frameworks. TensorFlow has big community support that is growing day by day, which makes it a good option for building your complex deep learning applications
Chapter 5, Tensorflow in Action - Some Basic Examples, will explain the main computational concept behind TensorFlow, which is the computational graph model, and demonstrate how to get you on track by implementing linear regression and logistic regression.
Chapter 6, Deep Feed-forward Neural Networks - Implementing Digit Classification, explains that a feed-forward neural network (FNN) is a special type of neural network wherein links/connections between neurons do not form a cycle. As such, it is different from other architectures in a neural network that we will get to study later on in this book (recurrent-type neural networks). The FNN is a widely used architecture and it was the first and simplest type of neural network. In this chapter, we will go through the architecture of a typical ;FNN, and we will be using the TensorFlow library for this. After covering these concepts, we will give a practical example of digit classification. The question of this example is, Given a set of images that contain handwritten digits, how can you classify these images into 10 different classes (0-9)?
Chapter 7, Introduction to Convolutional Neural Networks, explains that in data science, a convolutional neural network (CNN) is specific kind of deep learning architecture that uses the convolution operation to extract relevant explanatory features for the input image. CNN layers are connected as an FNN while using this convolution operation to mimic how the human brain functions when trying to recognize objects. Individual cortical neurons respond to stimuli in a restricted region of space known as the receptive field. In particular, biomedical imaging problems could be challenge sometimes but in this chapter, we'll see how to use a CNN in order to discover patterns in this image.
Chapter 8, Object Detection – CIFAR-10 Example, covers the basics and the intuition/motivation behind CNNs, before demonstrating this on one of the most popular datasets available for object detection. We'll also see how the initial layers of the CNN get very basic features about our objects, but the final convolutional layers will get more semantic-level features that are built up from those basic features in the first layers.
Chapter 9, Object Detection – Transfer Learning with CNNs, explains that Transfer learning (TL) is a research problem in data science that is mainly concerned with persisting knowledge acquired during solving a specific task and using this acquired knowledge to solve another different but similar task. In this chapter, we will demonstrate one of the modern practices and common themes used in the field of data science with TL. The idea here is how to get the help from domains with very large datasets to domains that have smaller datasets. Finally, we will revisit our object detection example of CIFAR-10 and try to reduce both the training time and performance error via TL.
Chapter 10, Recurrent-Type Neural Networks - Language Modeling, explains that Recurrent neural networks (RNNs) are a class of deep learning architectures that are widely used for natural language processing. This set of architectures enables us to provide contextual information for current predictions and also have specific architecture that deals with long-term dependencies in any input sequence. In this chapter, we'll demonstrate how to make a sequence-to-sequence model, which will be useful in many applications in NLP. We will demonstrate these concepts by building a character-level language model and see how our model generates sentences similar to original input sequences.
Chapter 11, Representation Learning - Implementing Word Embeddings, explains that machine learning is a science that is mainly based on statistics and linear algebra. Applying matrix operations is very common among most machine learning or deep learning architectures because of backpropagation. This is the main reason deep learning, or machine learning in general, accepts only real-valued quantities as input. This fact contradicts many applications, such as machine translation, sentiment analysis, and so on; they have text as an input. So, in order to use deep learning for this application, we need to have it in the form that deep learning accepts! In this chapter, we are going to introduce the field of representation learning, which is a way to learn a real-valued representation from text while preserving the semantics of the actual text. For example, the representation of love should be very close to the representation of adore because they are used in very similar contexts.
Chapter 12, Neural Sentiment Analysis, addresses one of the hot and trendy applications in natural language processing, which is called sentiment analysis. Most people nowadays express their opinions about something through social media platforms, and making use of this vast amount of text to keep track of customer satisfaction about something is very crucial for companies or even governments.
In this chapter, we are going to use RNNs to build a sentiment analysis solution.
Chapter 13, Autoencoders – Feature Extraction and Denoising, explains that an autoencoder network is nowadays one of the widely used deep learning architectures. It's mainly used for unsupervised learning of efficient decoding tasks. It can also be used for dimensionality reduction by learning an encoding or a representation for a specific dataset. Using autoencoders in this chapter, we'll show how to denoise your dataset by constructing another dataset with the same dimensions but less noise. To use this concept in practice, we will extract the important features from the MNIST dataset and try to see how the performance will be significantly enhanced by this.
Chapter 14, Generative Adversarial Networks, covers Generative Adversarial Networks (GANs). They are deep neural net architectures that consist of two networks pitted against each other (hence the name adversarial). GANs were introduced in a paper (https://arxiv.org/abs/1406.2661) by Ian Goodfellow and other researchers, including Yoshua Bengio, at the University of Montreal in 2014. Referring to GANs, Facebook's AI research director, Yann LeCun, called adversarial training the most interesting idea in the last 10 years in machine learning. The potential of GANs is huge, because they can learn to mimic any distribution of data. That is, GANs can be taught to create worlds eerily similar to our own in any domain: images, music, speech, or prose. They are robot artists in a sense, and their output is impressive (https://www.nytimes.com/2017/08/14/arts/design/google-how-ai-creates-new-music-and-new-artists-project-magenta.html)—and poignant too.
Chapter 15, Face Generation and Handling Missing Labels, shows that the list of interesting applications that we can use GANs for is endless. In this chapter, we are going to demonstrate another promising application of GANs, which is face generation based on the CelebA database. We'll also demonstrate how to use GANs for semi-supervised learning setups where we've got a poorly labeled dataset with some missing labels.
Appendix, Implementing Fish Recognition, includes entire piece of code of fish recognition example.
