Artificial intelligence has been in the making for more than 50 years, yet we are still learning, evolving, and determining best practices to ensure global adoption. Part of that is because the industry is still discovering new use cases and gaining a holistic understanding of what types of machine learning are best for certain problems.
But, before I jump into the different types of machine learning, let’s first go over the basics of artificial intelligence (AI) and machine learning (ML).
Artificial Intelligence vs. Machine Learning
In a nutshell, artificial intelligence is making computers function like humans. AI is about transforming raw data into valuable information. Data is good, but information is better, as information is what we ultimately use to make decisions.
Whereas machine learning is an application of artificial intelligence that provides systems the ability to automatically learn and improve from experience without being explicitly programmed.
To be more specific, machine learning ultimately is a computer program that is said to learn from experience (E) with respect to some class of tasks (T) and performance measure (P), if its performance at tasks in T, as measured by P, improves with experience (E). For example:
- Task (T): classify whether a comment on a forum is appropriate or not
- Performance (P): classification accuracy – number of comments predicted correctly as either appropriate or not
- Experience (E): corpus of data representing whether a comment on a forum was appropriate or not
What this basically means is you give a set of experiences (training data) to a machine to perform a certain set of tasks and you measure based on a certain set of criteria (P).
Put another way, machine learning enables computers to problem-solve through training and experiences to perform a task.
Alright, now let’s get back to the title of this blog!
Different Types of Machine Learning
Machine learning can roughly be broken down into three types of learning: supervised, unsupervised and reinforcement learning. A researcher at Google once said he imagines this as a cake where the top layer is supervised, the middle layer is unsupervised and the bottom is reinforcement.
When we think about how children learn, they learn in one of these same ways:
- Supervised learning would be if you show a picture of a cat with the word cat underneath.
- Unsupervised is when you try and learn a pattern on your own – for example, Sesame Street’s ‘one of these things is not like the others.’
- Reinforcement learning would be if you punish for a wrong answer and reward for the right one. We do this all the time with kids and sometimes it’s autonomous like if they burn their fingers on a hot stove.
So, let’s break these categories down even further.
Within supervised learning, there are two types: classification and regression. They are a variance of the same thing, but classification is discrete and regression is continuous. Things like diagnostics, image classification, and fraud detection all fall under classification problems.
Regression could be things like market forecast, determining life expectancy, and network prediction. The difference is classification puts you into one of X number of buckets where regression determines a value between zero and a million, so it’s not a clear classification, though they are variants of each other.
Unsupervised learning is clustering and dimensionality reduction. Clustering could be customer segmentation, targeted marketing, etc. Dimensionality could be determining a customer that exists in many dimensions. For example, each customer has an age, level of education, where they live, occupation, etc., and each factor translates to a dimension. It’s incredibly hard for humans to see them all and make use of the information. But ML can help turn those dimensions into valuable insights.
Lastly, reinforcement learning is used a lot in game AI, skill acquisition, and robotics. Alpha Go is a classic case – teach an AI to play a game and do it well.
Regardless of which technique is used to train a model, the key is training data or also known as the ground truth. Large amounts of data need to be prepared to accurately train models. Your model can only be as good as your ground truth.
To train a machine learning model in any of the three ways we just discussed, you first need to ensure you have high-quality training data before then determining which technique is best for the problem you are trying to solve.