An Introduction to Machine Learning

Machine learning algorithms have shown great promise in providing solutions to complex problems. Some of the applications we use every day from searching the Internet to speech recognition are examples of tremendous strides made in realizing the promise of machine learning.

Machine learning algorithms derive their strength from their ability to learn from the available data. This chapter introduces you to the most prominent learning models. Subsequent chapters will dive deeper into the various algorithms or applications that depend on these models. The primary learning models that will be considered are Supervised, Unsupervised, Semi-supervised and Reinforcement learning.

Regression analysis is a statistical and machine learning technique that allows you to build a model based on a labelled dataset (such as past data for stocks) that can be used to make predictions. This chapter starts getting into the actual aspects of machine learning. By the end of this chapter, you will be able to perform some simple predictions for new data, based on learning from the labelled dataset (from prior observations).

In this chapter, we will further refine the techniques learnt in the previous chapter. With these refinements, you will be able to achieve quicker convergence and get improved predictions.

Classification refers to the problem of identifying the category to which an input belongs to among a possible set of categories. The possible set of categories are labelled, and models are generally learned from training data. Classification models can be created using simple thresholds, regression techniques, or other machine learning techniques like Neural Networks, Random Forests, or Markov models.

Random Forests are effective and intuitive models used in classification and regression problems. They are intuitive because they provide clear path to a result and are based on underlying Decision Tree structures. A Decision Tree is a machine learning model built using series of decisions based on variable values to take one path or the other. A Random Forest is a collection of Decision Trees that improve the prediction over a single Decision Tree.

The learnings in this chapter will help you determine if your choice of features and the number of datasets are sufficient or should you increase datasets and/or increase/decrease the number of features forming your hypothesis.

This chapter will cover the basics of artificial neural networks (ANNs) which are also called multilayer perceptrons. Neural networks are networks of interconnected artificial neurons. Their structure is heavily inspired by the brain’s neuron network. A neural network is generally used to create supervised machine learning models for classification, similar to a Logistic Regression model, and is useful in cases where Logistic Regression may not provide reasonable accuracy. Neural networks form the basis of many of the complex applications and algorithms of machine learning. You will subsequently see some of these applications in Chaps. 16 and 17 (Reinforcement Learning), 11 (Recurrent Neural Network Used in Language Processing), and 15 (Convolutional Neural Network). A good understanding of a neural network is necessary to understand these and other applications that have raised so much interest in machine learning. Neural networks are also used in unsupervised learning for compressed representation and/or dimensionality reduction.

Natural language processing (NLP) is understanding and interpreting human languages, spoken or written, using machine processing. NLP is useful in a variety of applications including speech recognition, language translations, summarization, question responses, speech generation, and search applications. NLP is an area of research which has proven to be difficult to master. Deep learning techniques have started to solve some of the issues involved in natural language processing.

Deep learning has seen significant success in recent time with applications like speech recognition, image processing, language translation, and list goes on. Deep neural networks in general refer to neural networks with many layers and large number of neurons, often layered in a way that is generally not domain specific. Availability of compute power and large amount of data has made these large structures very effective in learning hidden features along with data patterns.

The learnings in this chapter will help you determine if something is anomalous. Anomalous in this context effectively means out of range. It could be about a defect in a component or a fraudulent exchange. For example, if a person usually types with a certain speed, and suddenly the system sees a different speed, it is anomalous behavior, giving an indication of possible impersonation. Or, for a component, certain measured parameters being outside the range compared to the range exhibited by normal components could indicate a defective component. This information can be used for quality checks, either during manufacturing, shipment, acceptance stages, or also during preventive maintenance. The concept of anomaly detection depends on being able to observe certain parameters and then being able to form an opinion about some dependent behavior, which is difficult to directly observe at this time. For example, predict the expected remaining life of an aircraft engine based on noise (something that can be observed now). Similarly, use measured value of typing speed to predict if the person (not visible) at the other end is impersonating. Anomaly detection is used very extensively in monitoring computer equipment in large data centers. All the computers in a data center are monitored for many parameters. If a specific computer seems to exhibit a behavior that is very different from other computers, this computer could be defective and may need to be taken out of the network. For example, if it’s getting too few jobs compared to other computers, maybe, it’s working slowly. Or, if it’s getting too many jobs compared to others, maybe, it’s not working properly and terminating jobs immediately that come to it and, thus, be ready to take on additional jobs.

This chapter explains how to build recommender systems. Recommender systems are the systems that make recommendations to you, based on prior choices exercised by you and by others who have exhibited similar tastes in their choices. When you visit an e-commerce site and look for a specific dress, you start seeing several other dresses which are similar. Or, when you watch a video on YouTube, it starts recommending several other videos which are similar. Similarly, when you read a specific news item, the website starts showing you information related to that same category of items. When you search for a book on amazon.com, it also shows other books in the same genre or by the same author.

Convolution is a very important concept in the world of machine learning. In many of the previous chapters, you have read about various algorithms, and you have seen how they work on numbers. Convolution is a technique which automates extraction and synthesis of significant features needed to identify the target classes, useful for machine learning applications. Fundamentally, convolution is feature engineering guided by the ground truth and cost function. Thus, convolution is used for some of the coolest applications of machine learning, such as image recognition, handwriting reading, interpreting street signs, etc. As you can well imagine, one of the most famous applications of machine learning – ADAS (autonomous driver assistance system) – depends on convolution as a component of the whole system to identify objects and to interpret signs!!

Reinforcement learning (RL) is one of the fundamental areas of machine learning and is widely regarded as a necessary component of artificial general intelligence (AGI). While RL has been researched for a few decades, the advent of deep learning has resulted in the so-called deep reinforcement learning algorithms that utilize deep neural networks and large-scale computing power to significantly improve the capabilities of RL. They have resulted in computer programs that play video games and board games at super human level and even beat handcrafted computer programs, e.g., chess program AlphaZero beat the best handcrafted program Stockfish 8. Besides video and board games, RL is being applied to enable robots to learn complex tasks in autonomous driving and industrial automation. This chapter covers the fundamental concepts of RL and the taxonomy of RL algorithms. In the next chapter, you will learn some of the key algorithms in reinforcement learning. An understanding of this chapter will provide you with a good feel for RL.

In the previous chapter, you were introduced to major aspects of reinforcement learning. This chapter takes you through the next steps on dealing with those challenges in order to form the algorithms for reinforcement learning. Reinforcement learning is an area of very active research, and new variations of algorithms are proposed regularly. An understanding of this chapter will provide you with a good basis, so that you can appreciate not just the current generation algorithms but also understand new research findings in this area. RL involves interaction of multiple components and concepts, which you have already seen in prior chapters in this book.

In the previous chapters, you have seen various algorithms and how they apply to specific problem domains. This chapter will help you get into the finer details of designing a machine learning system. The concepts explained in this chapter are less about individual algorithms; they are about making choices for implementing your algorithms.