THE AI COMPASS (eBook)

What is Machine Learning?

One of the most promising subfields of artificial intelligence is machine learning. It involves the process where systems can “learn” through statistics, trial and error as well as data to enable them optimize processes and also innovate much faster. Machine learning empowers computers to develop human-like capabilities that make it possible for them to resolve various challenges facing the world like climate change, cancer, HIV/AIDS and several others. So, in what ways is machine learning empowering computer systems with human-like capabilities?

The process of machine learning is automated and all through the learning process, it is usually fine-tuned based on the machines’ experiences. The machines are fed with high-quality data and machine learning models are developed with different algorithms which we shall look at shortly. The type of algorithm used is based on the available data as well as the kind of activity that is being automated. One question that comes to mind at this point is, how exactly does machine learning differ from traditional programming?

The answer is simple – we feed the input data (and a well-developed and tested program) into a machine to enable it generate an output. But this is not the case with machine learning as input and output data are both fed into the machine in the course of learning and the machine will work out a program for itself. Take a look at the illustration below.

Figure 9: Machine learning process

Generally, computer programs often depend on code to inform them on the things they should do or the information they should store. This is also regarded as “explicit knowledge” which encompasses things that can easily be recorded or written such as videos, manuals or textbooks. Presently, computers are acquiring tacit knowledge – knowledge acquired from context and personal experience – courtesy of machine learning. It is hard to transfer this kind of knowledge from one individual to another through verbal communication or text.

An excellent example of tacit knowledge is facial recognition. Have you observed that when we recognize people’s faces, it is not always easy to explain how or why we even recognize them accurately? What happens is that when we see a person, we depend on our personal knowledge database to tacitly make the conclusions and recognize an individual based on their face. Have you ever tried explaining how to ride a bike to a friend or family member before? You will agree with me that it is usually an easier task to just show them exactly how to ride a bike than trying to explain how it is done.

This is also what machine learning is all about. It is no longer compulsory for computers to depend on billions of lines of codes before they can execute calculations. With machine learning, they now have the power of tacit knowledge and this enables them easily make such connections, identify patterns and also leverage the things they already learned before in making predictions. The use of tacit knowledge by machine learning has undoubtedly made it extremely useful for virtually all industries – government, fintech, weather, healthcare, etc. We will be looking at how AI and machine learning are being used in different industries in section II.

Deep Learning

One subfield of machine learning that is also increasingly gaining traction is deep learning. Deep learning is getting more useful because of its unique ability to accurately extract data. To extract higher-level features from raw data, deep learning leverages Artificial Neural Networks (ANN). More on deep learning later in this chapter.

Common Types of Machine Learning

Figure 10: Types of machine learning

For machine learning to establish parameters, actions and end values, it also requires algorithms just like all systems with AI. The purpose of these algorithms is to serve as a guide for machine-learning-enabled programs while they go through several options and evaluate various factors. Computers actually use hundreds of algorithms based on different factors such as diversity and data size. I will not be going through all the available machine learning algorithms because that is beyond the scope of this book. But I will briefly discuss the most common types.

Supervised Learning

These types of algorithms help create mathematical models of data containing input and output information. Another word for supervised learning algorithms is training data and the reason for this name is that the programs understand the beginning and end results of the data. What it simply needs to do is to determine the most efficient way to achieve the result. To enable machine learning programs predict outputs based on a new set of inputs, they are constantly provided with these sets of supervised learning algorithms. Two examples of supervised learning algorithms that are more popular than the others are classification and regression algorithms.

Another name for regression analysis is linear regression and this algorithm is used in discovering and predicting relationships between outcome variables and at least an independent variable. It also serves as training data to enhance the ability of systems to predict and forecast.

The second most popular type of this set of algorithms are classification algorithms and their purpose is for training systems on object identification and placement in the right sub-category. An excellent example is the use of machine learning by email filters to automate incoming email flows for spam, promotion and primary inboxes.

Systems are usually exposed to a wide range of labeled data and this could be images of handwritten figures annotated to signify the number they correspond to. When a supervised-learning system is provided with sufficient examples, it would learn to identify the clusters of shapes and pixels linked with each number and finally identify handwritten examples and even distinguish between numbers 9 and 4, or 8 and 6 reliably. It is important to know that the training of these systems often demands vast amounts of labeled data. Some systems may need exposure to several millions of examples before they can finally master an activity or task.

Unsupervised Learning Algorithms

Unlike the first category of machine learning, unsupervised learning requires algorithms to identify patterns in data as it attempts to discover similarities that separate the data into categories. For instance, Airbnb clustering together homes that are available to rent based on different neighborhoods is a good example of unsupervised learning. Another example might also be Google News categorizing stories on topics that are similar every day. This set of learning algorithms are not designed to separate specific types of data. Instead, they are designed to search for data that they can group based on similarities or for irregularities that stand out.

Semi-Supervised Learning

The rise of semi-supervised learning may eventually lower the importance of vast sets of labeled data required for training machine learning systems. The name already explains what it means –training that is supervised and unsupervised. It is a method that trains systems by depending on a large amount of unlabeled data as well as a small amount of labeled data. A machine learning model will be partially trained with the labeled data and the partially trained model will, in turn, be used to label the unlabeled data.

The entire process is often regarded as pseudo-labeling. The resulting mix of pseudo-labeled data and labeled data will be used for training the model. Recently, semi-supervised learning’s viability has been enhanced by Generative Adversarial Networks (GANs). They are machine learning systems that are capable of generating entirely new data with labeled data which will assist in training a machine learning model. Once we get to the point where semi-supervised learning possesses the same level of effectiveness as supervised learning, then access to large, labeled datasets will no longer be as important to us for successfully training machine learning systems as access to vast amounts of computing power.

Reinforcement Learning

An easy way to understand what this type of machine learning means is to consider the process of learning an old-school computer game for the very first time, especially when the person is not conversant with ways of controlling the game or the rules. Although they initially tend to be a complete novice, with time, their performance will continue to improve as they understand the relationship between the different buttons they press, the actions that take place on the screen and the points they scored.

Google DeepMind’s Deep Q-network is perhaps the best example of reinforcement learning. Interestingly, the system has already defeated humans in different vintage video games. During the training process, the system is supplied with pixels from every single game and then it determines different information regarding the state of the game and this includes the distance between objects on the screen. Next, the system will examine how the state of the game and actions being performed in the game affects the score it gets. In the course of playing the game repeatedly, the system will successfully establish a model that has learned from many cycles and will maximize the score.

Factors Behind the Success of Machine Learning

First, you should know that machine learning is not a new training technique; instead, interest in the field increased dramatically in recent years. Perhaps some of the factors that are responsible for...