Skip to content

A Beginner’s Guide to Machine Learning

Share on twitter
Share on facebook
Share on linkedin

Machine learning is one of those buzzwords (along with deep learning and artificial intelligence) that’s become ubiquitous in science and tech news. In fact, if you look at Google Trends data, you’ll see that “machine learning” searches have been trending upwards since 2014.

Machine learning is seemingly everywhere, but a lot of people using the term would be hard pressed to properly explain what it is, what it does, and what it’s best used for.

This article aims to both define machine learning and provide clear, real-world examples of how the technology works in order to demonstrate why it’s so valuable. Then we’ll explore the different machine learning methods, consider how deep learning relates to these, and identify how machine learning is being employed to solve business problems. Finally, we’ll pull out our crystal ball to offer some brief thoughts on the future of machine learning.

What is Machine Learning?

In simple terms, machine learning (often abbreviated ML) is a subset of artificial intelligence that deals with extracting patterns from complex data. This can be used by computers to identify patterns and associations in massive amounts of data, and then make predictions and forecasts based on its findings.

Some forms of machine learning—like reinforcement learning (see below)—can even adjust their decision-making abilities over time as they are fed further data. In practice, this means that machine learning algorithms can become more “intelligent” on their own, whether with human guidance or without it.

How does Machine Learning Work?

So how do machine learning algorithms predict the future? They use computational techniques to gather information directly from data without relying on mathematical models or predetermined equations.

The best way to explain this is through reference to one of machine learning’s most high-profile applications: gaming. Machine learning has proven to be extremely effective at mastering board games like chess and Go.

In the past, AI programs that played these games would effectively try to “out compute” their opponents by evaluating millions of moves per second while applying strategies that reflected humans’ understanding of the game. When IBM’s Deep Blue beat reigning human world champion, Garry Kasparov, in a six-game match in 1996, it was essentially using a sophisticated version of this brute force technique.

However, machine learning offered a different approach. ML-powered programs would begin by knowing only the basic rules of the game and the desired outcome (victory). The program’s strategy for play would develop organically by analyzing data from past games and endlessly testing and refining this knowledge through new games (often against older versions of itself).

That’s exactly how AlphaGo, the program that beat the world’s best human Go player in 2016, was trained. This victory was monumental, as Go is a much more complex game than chess and the brute force of previous AI programs was never enough to beat the world’s top players. AlphaGo’s successor, AlphaZero, used the same machine learning techniques to reach a superhuman level of skill in chess by playing against itself over and over again for 24 hours, and a similar program (Stockfish) is now the world’s strongest CPU chess engine.

The same kinds of algorithms have also been used for video games. Check out the following video, which shows how machine learning was used to teach a program to play Super Mario World by simply giving it a very simple objective (advance rightward).

What Makes Machine Learning so Valuable?

So, machine learning is really good at playing Go and Super Mario World. Why does that matter for business? Or healthcare? Or the sciences?

Data is the lifeblood of all of these fields, and machine learning is really, really good at processing data to identifying obscure patterns. That’s why using machine learning to make objective-directed decisions that are reliable and repeatable is so useful for organizations grappling with big data. Machine learning has applications across industries, including manufacturing, retail, healthcare and life sciences, travel and hospitality, financial services, and energy, feedstock, and utilities.

What are the Different Types of ML Learning Styles?

The different algorithms and applications of machine learning are often categorized as representing one of two methods: supervised learning or unsupervised learning. If we pull back slightly to examine these learning styles, we’ll see that there are actually several more categories, each with distinct benefits and use cases.

Supervised learning

Supervised learning algorithms are trained using datasets that have labeled inputs and outputs based on historical data. We can think of these inputs and outputs as predictors and outcomes. These labels are designed to train (or “supervise”) the algorithms into classifying data or predicting outcomes, and then to measure accuracy and learn from results over time.

If you take a classic example like predicting customer churn, your dataset would have information about customer behavior—how often they use the product, how long they’ve had their account for, etc.—as the inputs or predictors, as well as whether or not they churned—the outcomes, or the expected output of your machine learning model once it’s trained.

Supervised learning is usually applied towards two types of problems: classification and regression.

  • Classification problems use an algorithm to accurately assign data into specific categories. Spam filters are an example of this. “Spam” can be a subjective category—the distinction between spam and non-spam messages is fuzzy—and the spam filter algorithm is always improving itself based on feedback from users (meaning email that humans mark as spam).
  • Regression, meanwhile, is useful for understanding the relationship between dependent and independent variables. Regression models can predict numerical values based on different data points, such as sales revenue projections for a given business. Some popular regression algorithms are linear regression, logistic regression, and polynomial regression.

Unsupervised learning

Unsupervised learning is used with data that has not been classified or labeled. In this case, the machine learning algorithm is used not just to predict outcomes, but also to find structure and patterns within the data, all without historical data guiding it towards desired outputs. The two most commonly used unsupervised learning techniques are clustering (in which data points with similar characteristics are grouped together and assigned to “clusters”) and association rules (essentially “if-then” statements that detail the probability of relationships between data items). These are the algorithms you’ll find under the hood for market research and audience segmentation applications.

Semi-supervised learning

As the name suggests, semi-supervised learning is a mixture of supervised and unsupervised methods in that it often involves a small number of labeled examples and a larger number of unlabeled examples. The model will then draw from this mixed dataset to identify patterns and make predictions. Semi-supervised learning has proven to be especially useful for medical images. A radiologist can analyze and label a small set of scans for tumors or diseases to help the model learn to recognize these features in larger datasets.

Reinforcement learning

Reinforcement learning is used to train machine learning models to make a sequence of goal-oriented decisions in an interactive environment. The gaming use cases above are perfect examples of this. You don’t need to feed AlphaZero thousands of past games of chess with each move labeled as “good” or “bad.” You simply teach it the rules of the games and the objective, and then let it try out random actions. When actions move the program closer to the objective (like establishing a strong pawn position), it receives positive reinforcement. When actions do the opposite (like moving the king around prematurely), it receives negative reinforcement. Through this model, the program can eventually master the game.

In the real world, reinforcement learning is used extensively in robotics to train machines for complex and hard-to-engineer behaviors. It’s also used with roadway infrastructure, such as traffic lights, in order to optimize traffic flow.

Self-supervised learning

Self-supervised learning is often listed as a subset of unsupervised learning, although that’s not entirely right. Rather, self-supervised learning is something closer to a hybrid, as it involves applying supervised learning techniques to surrogate tasks and data in order to build models that can be applied to the actual data. One common application of this is image manipulation, particularly filling in gaps in photos. If you train a machine learning model to understand the component parts of the image it can see, it can then use that information to predict what components should be present in the missing sections.

Trying to get your machine learning project off the ground?

Our Human’s Guide to ML Projects walks you through everything you need to do to build a solid project foundation, right from day one.

What about Deep Learning?

Deep learning is a subfield of machine learning focused on a class of algorithms called artificial neural networks that are meant to mimic the structure and function of the human brain. The “deep” in deep learning refers to the use of multiple layers in the network architecture. Each layer of these artificial neurons evaluates one decision-making criterion (such as identifying the edges in an image). Subsequent layers are trained on the previous layers’ outputs (so edges → shapes → object segmentation), such that the algorithm is able to identify higher-level features (like faces or body types), and its predictive and classification capacity increases. As the examples suggest, deep learning is often used in image and video processing, as well as in natural language processing.

Deep learning can be utilized in any of the ML learning styles described above, whether supervised, unsupervised, semi-supervised, reinforcement, or self-supervised. 

What Problems can Machine Learning Solve?

Businesses are grappling with more enterprise, mobile, and sensor data than ever before, but that data is just that—data—until machine learning processes it to uncover insights and predict the future. The ability to see patterns and connections within huge datasets is why machine learning is driving innovation in nearly every sector of business. Here are some of the most promising use cases across industries.

Fraud detection

Before machine learning, financial fraud detection was prohibitively expensive and often ineffective. Sorting through transaction data looking for irregularities took an enormous amount of time and resource allocation. Algorithms, though, never grow tired and can work around the clock to identify fraudulent transactions, flag compromised accounts immediately, and increase security for businesses and financial institutions. And the more data that machine learning algorithms process, the better they become at spotting patterns and understanding purchasing behavior to spot outliers that might indicate fraud.

Targeted marketing 

Websites and mobile devices collect a tremendous amount of data about users, and machine learning can be used to transform that data into highly targeted and personalized marketing. Machine learning algorithms are especially good at product and content recommendations (like what you see with Netflix and Spotify), predictive lead scoring (by identifying engagement patterns that suggest high conversion rates), customer lifetime-value forecasting (based on purchase history, logins, and other variables), and churn-rate prediction (through the identification of subtle patterns in behavior).

Traffic prediction

Machine learning algorithms—like the ones powering Google Maps—can use a mix of historical traffic data along with real-time GPS data transmitted from private cars and public buses to predict traffic congestion levels, plot out ideal navigation routes, and provide more accurate estimated arrival times.

Chatbots

The widespread deployment of chatbots across the internet reflects the success of machine learning-powered natural language processing. Today’s chatbots are far more sophisticated and conversational than their early predecessors, which often struggled to answer customer queries. The technology’s benefits are substantial, as chatbots offer automated and scalable end-to-end customer service. Companies like Drift are even allowing customers to effectively build chatbots in under an hour, without the need for development teams.

Facial recognition

The earliest facial recognition computer programs were developed in the 1960s. It wasn’t until 2006, though, with the launch of the Face Recognition Grand Challenge, that the technology came into its own thanks to increasingly powerful algorithms. Today, machine learning and neural networks are driving the deployment of facial recognition in all manner of applications, from identity verification at ATMs to access control at secure facilities. And in the future, this technology may facilitate everything from personalized digital advertising to completely staff-free stores.

What’s the Future of Machine Learning?

Interest in machine learning is clearly increasing as more businesses and large organizations employ the technology to solve particular problems or fuel innovation. This ongoing investment reflects a recognition that machine learning is delivering ROI, especially through some of the proven and replicable use cases described above. After all, if the technology is good enough for Netflix, Facebook, Amazon, Google Maps, and on and on, then chances are high it can help your business make the most of its data.

As new machine learning models are trained and released, we’re going to see an expanding number of applications that will play out across sectors. You’re already seeing this with facial recognition, as what seemed like a novel feature on your iPhone several years ago is now being incorporated into all manner of programs and apps, especially those based around public security.

The secret for most businesses looking to start out with machine learning is to look beyond the flashy futurism and identify the specific business problems that the technology can help you with.

 

If you want to dive deeper into machine learning, including how to get your first project off the ground, check out RapidMiner’s Human’s Guide to Machine Learning Projects.

Getting a machine learning project off the ground is hard

With various stakeholders, differing background knowledge among team members, and administrative hurdles, many projects die before they have a chance to fly. The solution to this problem is to build a solid project foundation from the very first stages to set yourself up for success.

The process outlined in this guide will help make that easier.

Additional Reading

Chris Doty

Chris Doty

Chris has a PhD in linguistics, and has previously worked on ML projects for Amazon's Alexa. As RapidMiner's Content Marketing Manager, he works to evangelize for the power of AI and ML to upskill and empower people in a changing world. When he isn't working, he enjoys learning languages and drawing.