What are the different types of machine learning?
Valérie Bécaert Valérie Bécaert
February 13 5 min

What are the different types of machine learning?

Many people think machine learning is a single type of technology. It’s sometimes spoken about like a single algorithm that can be copied and pasted into a software’s code. But, as is always the case with AI and technology, it’s a little more complicated than that. In fact, There are several different types of machine learning algorithms, designed to complete different tasks in varying ways. Much of the time experts find it difficult enough to settle on what is and isn’t machine learning – never mind how to categorize different variants of the technology.

Types of machine learning subdivide into four main categories. In this article, we look through these and consider what makes them unique, and how they can be used in practice.

Supervised machine learning

The vast majority of machine learning is what’s known as “supervised machine learning.” This involves individual users feeding a vast swath of pre-solved information into an algorithm, which then analyzes it to discover patterns and learn rules. This type of machine learning is best suited to information where there is a clear X and Y variable, and the system is learning how to get one from the other.

One prominent example of this in practice comes from social media giant Facebook, who use supervised machine learning to detect inappropriate content on the platform. In this instance, the content of the social post is the X variable, and the Y variable is whether it can be flagged as inappropriate or not. To train the machine learning model, a team fed a range of content into the algorithm, which they’d manually flagged as appropriate or inappropriate. From there, the machine learning model could be left to its own devices to try to infer how and why the flagged content was inappropriate.

Unsupervised machine learning

Not all machine learning tasks have clear X and Y values. Much of the time, machine learning is employed not to discover a prescribed outcome, but to uncover patterns and structures that govern the relationship between disparate but related data points. Consider, for example, if you wanted to teach the algorithm to differentiate various types of vegetables – without having to identify which is which. Using a type of unsupervised machine learning known as “clustering,” the system could analyze pictures of the vegetables and infer similarities or differences between them. Doing this, they could learn more than the difference between “carrots” and “potatoes”; they could learn to categorize particular types of carrots, perhaps by weight, color, or shape.

The benefit of this is that it allows you to uncover patterns within data that you might never have even known were there. But there’s also a more practical benefit to unsupervised machine learning. As the name says, you don’t have to supervise it.

Semi-supervised machine learning

Semi-supervised machine learning takes the best of both approaches to create something of a hybrid. There are two types of situations where this tactic can be helpful:

  • If supervised machine learning would be your default option, but you lack the time or resources to dedicate to supervising the process fully. Therefore, you give the algorithm some direction —then leave it to its own devices.
  • On the other hand, if you’d use unsupervised machine learning, but you’d like to give the algorithm some direction, perhaps to save time or to direct the system towards a more structured, didactic outcome.

In many cases, machine learning that’s described as either supervised or unsupervised is actually a combination of both. This is why concrete classification can be difficult.

Reinforcement learning

Reinforcement learning is designed to sequentially find the quickest route to a given destination. One of the most common examples of reinforcement learning involves a game with a clear end-point or goal, and a number of different ways to achieve it. Let’s use a maze game as an example. The objective is to get one from one place to another in as few steps as possible. In this case, the reinforcement algorithm would investigate each possible route in a sequence and select the quickest.

In more practical situations, this has several benefits:

  • Predictive maintenance of machinery in manufacturing.
  • Optimizing energy consumption in factories and datacenters.
  • Identifying optimal treatments in the health sector.

The general theme of all these examples is one of optimization; feeding an existing system or relationship into the algorithm and finding the best possible outcome or destination.

Finding the right types of machine learning

Ideally, every type of machine learning algorithm would simply fit into one of these four categories, and it’d be easy to identify which you needed for a given task or problem. In reality, the broad categories feature a lot of crossover, and any particular system is almost bound to feature some combination of the different types of machine learning. Often, working out which type of machine learning you need is as complicated as the algorithm that powers it.

If you want to find out more about the different types of machine learning, read our recent blog ‘AI vs. machine learning’.