Hands-On Deep Learning Architectures with Python

Create deep neural networks to solve computational problems using TensorFlow and Keras

Yuxi (Hayden) Liu, Saransh Mehta

• Publisher: Packt Publishing
• Year: 2019
• ISBN:
  9781788990509
• Edition: 1
• Topic: Informatica
• Subtopic: Intelligenza artificiale (IA) e semantica

Concepts, tools, and techniques to explore deep learning architectures and methodologies

Key Features

• Explore advanced deep learning architectures using various datasets and frameworks
• Implement deep architectures for neural network models such as CNN, RNN, GAN, and many more
• Discover design patterns and different challenges for various deep learning architectures

Book Description

Deep learning architectures are composed of multilevel nonlinear operations that represent high-level abstractions; this allows you to learn useful feature representations from the data. This book will help you learn and implement deep learning architectures to resolve various deep learning research problems.

Hands-On Deep Learning Architectures with Python explains the essential learning algorithms used for deep and shallow architectures. Packed with practical implementations and ideas to help you build efficient artificial intelligence systems (AI), this book will help you learn how neural networks play a major role in building deep architectures. You will understand various deep learning architectures (such as AlexNet, VGG Net, GoogleNet) with easy-to-follow code and diagrams. In addition to this, the book will also guide you in building and training various deep architectures such as the Boltzmann mechanism, autoencoders, convolutional neural networks (CNNs), recurrent neural networks (RNNs), natural language processing (NLP), GAN, and more—all with practical implementations.

By the end of this book, you will be able to construct deep models using popular frameworks and datasets with the required design patterns for each architecture. You will be ready to explore the potential of deep architectures in today's world.

What you will learn

• Implement CNNs, RNNs, and other commonly used architectures with Python
• Explore architectures such as VGGNet, AlexNet, and GoogLeNet
• Build deep learning architectures for AI applications such as face and image recognition, fraud detection, and many more
• Understand the architectures and applications of Boltzmann machines and autoencoders with concrete examples
• Master artificial intelligence and neural network concepts and apply them to your architecture
• Understand deep learning architectures for mobile and embedded systems

Who this book is for

If you're a data scientist, machine learning developer/engineer, or deep learning practitioner, or are curious about AI and want to upgrade your knowledge of various deep learning architectures, this book will appeal to you. You are expected to have some knowledge of statistics and machine learning algorithms to get the best out of this book

Section 1: The Elements of Deep Learning

In this section, you will get an overview of deep learning with Python, and will also learn about the architectures of the deep feedforward network, the Boltzmann machine, and autoencoders. We will also practice examples based on DFN and applications of the Boltzmann machine and autoencoders, with the concrete examples based on the DL frameworks/libraries with Python, along with their benchmarks.

This section consists of the following chapters:

• Chapter 1, Getting Started with Deep Learning
• Chapter 2, Deep Feedforward Networks
• Chapter 3, Restricted Boltzmann Machines and Autoencoders

Getting Started with Deep Learning

Artificial intelligence might work, and if it does, it will be the biggest development in technology ever.

– Sam Altman

Welcome to the Hands-On Deep Learning Architectures with Python! If you are completely unfamiliar with deep learning, you can begin your journey right here with this book. And for readers who have an idea about it, we have covered almost every aspect of deep learning. So you are definitely going to learn a lot more about deep learning from this book.

The book is laid out in a cumulative manner; that is, it begins from the basics and builds it over and over to get to advanced levels. In this chapter, we discuss how humans started creating intelligence in machines and how artificial intelligence gradually evolved to machine learning and eventually deep learning. We then see some nice applications of deep learning. Moving back to the fundamentals, we will learn how artificial neurons work and, in the end, set up our environment for coding our way through deep learning models. After completing this chapter, you will have learned about the following things.

• What artificial intelligence is, and how machine learning, deep learning relates to it
• The types of machine learning tasks
• Information about some interesting deep learning applications
• What an artificial neural network is, and how it works
• Setting up TensorFlow and Keras with Python

Let's begin with a short discussion on artificial intelligence and the relationships between artificial intelligence, machine learning, and deep learning.

Artificial intelligence

Ever since the beginning of the computer era, humans have been trying to mimic the brain into the machine. Researchers have been developing methods that would make machines not only compute but also decide like we humans do. This quest of ours gave birth to artificial intelligence around the 1960s. By definition, artificial intelligence means developing systems that are capable of accomplishing tasks without a human explicitly programming every decision. In 1956, the first program for playing checkers was written by Arthur Samuel. Since then, researchers tried to mimic human intelligence by defining sets of handwritten rules that didn't involve any learning. Artificial intelligence programs, which played games such as chess, were nothing but sets of manually defined moves and strategies. In 1959, Arthur Samuel coined the term machine learning. Machine learning started using various concepts of probability and bayesian statistics to perform pattern recognition, feature extraction, classification, and so on. In the 1980s, inspired by the neural structure of the human brain, artificial neural networks (ANN) were introduced. ANN in the 2000s evolved into today's so-called deep learning! The following is a timeline for the evolution of artificial intelligence through machine learning and deep learning:

Machine learning

Artificial intelligence prior to machine learning was just about writing rules that the machine used to process the provided data. Machine learning made a transition. Now, just by providing the data and expected output to the machine learning algorithm, the computer returns an optimized set of rules for the task. Machine learning uses historic data to train a system and test it on unknown but similar data, beginning the journey of machines learning how to make decisions without being hard coded. In the early 90s, machine learning emerged as the new face of artificial intelligence. Larger datasets were developed and made public to allow more people to build and train machine learning models. Very soon a huge community of machine learning scientists/engineers developed. Although machine learning algorithms draw inference from statistics, what makes it powerful is the error minimization approach. It tries to minimize the error between expected output provided by the dataset and predicted algorithm output to discover the optimized rules. This is the learning part of machine learning. We won't be covering machine learning algorithms in this book but they are essentially divided into three categories: supervised, unsupervised, and reinforcement. Since deep learning is also a subset of machine learning, these categories apply to deep learning as well.

Supervised learning

In supervised learning, the dataset consists of both the input data point and the expected output, commonly known as the label. The job of the algorithm is to learn a mapping function from inputs to expected outputs. The function could be a linear function such as y = mx + c or non-linear like y = ax3 + bx2 + cx + d, where y is the target output and x is the input. All the supervised learning tasks can be categorized into regression and classification.

Regression

Regression deals with learning continuous mapping functions that can predict values provided by various input features. The function can be linear or non-linear. If the function is linear, it is referred to as linear regression, and if it is non-linear, it is commonly called polynomial regression. Predicting values when there are multiple input features (variables), we call multi-variate regression. A very typical example of regression is the house prediction problem. Provided with the various parameters of a house, such as build area, locality, number of rooms, and so on, the accurate selling price of the house can be predicted using historic data.

Classification

When the target output values are categorized instead of raw values, as in regression, it is a classification task. For example, we could classify different species of flowers based on input features, petal length, petal width, sepal length, and sepal width. Output categories are versicolor, setosa, and virginica. Algorithms like logistic regression, decision tree, naive bayes, and so on are classification ...