TDWI | Training & Research | Business Intelligence, Analytics, Big Data, Data Warehousing

How Do Machines Learn?

Machine learning is a hot topic today and businesses around the world are using it to gain a competitive advantage. It is not a magical technology, rather the application of time-tested statistical practices to common business processes.

• By Troy Hiltbrand
• September 16, 2016

Every year, Gartner, a leading industry analyst firm, releases a report called the "Hype Cycle on Emerging Technologies." The purpose of the report is to identify up-and-coming technologies and visualize how close they are to becoming mainstream. The Hype Cycle is a graphic representation of the path that a technology goes through from obscurity through hype to stabilization and mainstream use.

In 2016, at the top of this hype cycle was the term machine learning. This means that machine learning is a hot topic across all industries. Gartner expects machine learning to go through a relatively short transition from hype to mainstream within the next two to five years.

What does this mean for businesses looking for the next technology that will differentiate them from their competition? It means that it is time to learn what machine learning is and what it is not, to identify if there is actual business value associated with it, and to start experimenting in applying it to business problems.

Defining Machine Learning

When many people hear "machine learning," they think of artificially intelligent robots who are acting completely autonomously, but that's not what we mean by the term. Machine learning is a branch of artificial intelligence where systems are developed not by codifying rules but by feeding the system data and letting the system evolve based on that data. This is where the learning occurs.

Just like humans, these systems learn from past experience. The biggest difference between human learning and machine learning is the speed at which machine learning can happen. Thomas H. Davenport, analytics thought leader, stated in The Wall Street Journal that "humans can typically create one or two good models a week; machine learning can create thousands of models a week." With machine learning, business value can be realized quickly through the evolving development of new models.

Machine learning is not a mystical and magical process; it's founded in traditional statistics, which has been studied for hundreds of years. Of course, statistics haven't been researched as a basis for machine learning until recently, and these studies have shifted only very recently from the laboratory to the business environment in an effort to monetize them.

Throughout this progress from science to research to business reality, the algorithms and approaches have evolved but the basics have remained the same.

Supervised Versus Unsupervised Learning

Today, there are two main categories of machine learning: supervised and unsupervised learning. The difference between them is what is known about the data that the machine learns from. Depending on the input data, different techniques and algorithms will be applied to the data to generate results.

This can be explained with a simple example. Let's say that a business has a truckload of rubber bouncy balls that it wants to divide by color. What is already known about those bouncy balls would determine which type of learning would be used. In this example, the goal for the machine learning is to learn to determine ball color .

Supervised learning: The first approach requires that the target attribute be known for a subset of the data. This would mean that the business has a sampling of the balls where the color is known. Each ball would be defined as a set of attributes. This could include ball size, weight, bounciness, and, for this subset, color.

This data about this subset of balls would be fed into an algorithm that would identify an association between the other attributes and the known color of the ball. Once this initial training is complete, the resulting model would be able to process the remaining balls -- with a similar set of attributes as the first set but unknown colors -- and guess the color of each ball based on the model.

In effect, the machine would have learned how to sort the balls by color by looking at the other attributes and creating a correlative model.

Unsupervised learning: The second approach would work if there is no subset of balls where color is known. Under these circumstances, the entire data set of attributes about the balls would be fed into an algorithm.

The algorithm would then find clusters of similarities in the attributes and divide the balls accordingly. In the end, the balls would be separated into groups. The algorithm would not necessarily know what color each ball is, only that there are differences and similarities in the groups it has generated.

From this process, the model would be able to take any future ball and sort it into one of these categories based on its attributes.

Machine Learning Models Lead to Emerging Intelligent Systems

The models generated by supervised and unsupervised learning are the heart of machine learning. It doesn't mean that the machine rises to the level of human intelligence, but as these models are applied to different parts of the business flow, business is able to work faster and often more accurately. When models are chained together, very complex and very intelligent systems start to emerge. The culmination of systems that bring together multiple machine learning models are often referred to as smart machines.

Examples of industry use are plentiful. From autonomous vehicles being researched and prototyped by all of the major automobile manufacturers to virtual personal assistants, such as x.ai, who will coordinate meetings between multiple attendees, machine learning is becoming more mainstream every day.

Today, machine learning is still classified as an emerging technology, but we are quickly seeing it applied to business problems across the spectrum to deliver tangible business value.