Introduction
In our last post we covered “AI and data science, what is it?” explaining key concepts relating to AI and data science. Now we’ll cover the different types of machine learning technology and use cases for each. We hope this information will give you a clear understanding of the types of problems you can solve with machine learning. Let’s dive into the insightful, unexpected, and profitable capabilities ML has to offer.
There are several different categories of ML techniques each with unique benefits and suitable for different types of applications. The types of ML that we will discuss today are Supervised Learning, Unsupervised Learning and Reinforcement Learning.
Supervised Learning
Supervised learning is the most common type of machine learning. In supervised learning an algorithm is provided examples of data and corresponding labels for each piece of data. For a particular question you’re trying to answer about the data the labels represent “yes” or “no”.
For example, if you’re trying to predict whether an email is spam (yes) or not spam (no), you’d provide an ML algorithm emails that are labeled as being either spam (yes) or not spam (no). The algorithm is trained on those historical data examples that have been labeled and then uses those to classify new pieces of data provided. In this case a new piece of data would be the contents of an email. Without knowing whether that new email is spam, the ML algorithm can use the historical examples it was trained on to output a probability about whether this new email is either spam or not spam. Some other examples of supervised learning could include:
- From data about houses (size, number of bedrooms, location, etc) and predicting what price a house should be.
- From data about online advertisements, user demographic information, and user behavior patterns (such as what users have clicked on an advertisement), predicting which future users will also click on a similar advertisement.
- From pictures of goods produced on a factory assembly line with scratches, dents, or defects, and without scratches, dents, or defects, automatically identifying goods that come off the assembly line with scratches, dents, or defects.
- From data about movie recommendations, suggesting movies to other users that have similar preferences.
- From photos of X-rays of patients that do and do not have cancer, automatically detecting if new X-rays images contain cancer.
Using supervised learning, a computer can make predictions about almost any situation where you can clearly label your data in a yes or no format (yes this is an example of a cat picture, no this isn’t an example of a cat picture). It’s not always the case, but generally speaking the more data you have (especially when using deep learning or neural networks which we’ll dive into in a future post), the more accurate output predictions you can get. Luckily, the amount of data that businesses have access to has been growing exponentially over the last several years. We’re at an inflection point where there is more data than ever at our fingertips.
Unsupervised Learning
Unsupervised learning is a type of machine learning where unlike in supervised learning you do not have clearly labeled ‘yes’ or ‘no’ examples for a given prediction. Unsupervised learning is used to draw conclusions and find meaningful patterns or groupings in data rather than train a computer to reach a conclusion that you already know the outcome of.
Some examples of unsupervised learning include:
- Grouping customers into segments for advertising and determining for a given product what is the demographic (age, location, education level) that prefers that product over another product
- Using image data to determine patterns regarding what driving on the correct side of the road should look like to automate self-driving cars
- Figuring out how to compress and aggregate data to reduce complexity without sacrificing the usefulness of the data
Reinforcement Learning
Reinforcement learning functions on the principles of punishment and reward. An algorithm is created with agents that perform actions in an environment. The agent takes decisions based on the rewards and punishments. For example, in baseball the agent would be a batter. When the batter strikes out the batter gets a negative point. When the batter gets a hit the batter gets a positive point.
The agents try every action they can and are given a reward based on the guidelines of their problem. This is repeated over and over again as the program learns the best course of action for the agents.
In the baseball example the program would learn what is the best way to get a hit and avoid striking out.
Some good use cases to apply reinforcement learning can include:
- Resource management (learning the best allocation of computer resources to waiting jobs)
- Robotic control (learning to use robotic limbs for a variety of applications)
- Personalized recommendations (determining recommendations for complex personalized features)
Conclusion
Machine learning comes in many varieties and these different approaches can be used to elevate any company into a cutting edge data driven organization. By thinking about an AI strategy there are many ways your business can thrive. Now is the time to empower your company to use data science and AI to unlock insights in your business. In our next post we will explore some of the business specific pitfalls to avoid when implementing machine learning technology.
We at Lumio are giving businesses data superpowers. If you have questions about AI, data science, machine learning, and how it can apply to your business get in touch with us at hello@lumio.ai, or by visiting our website at lumio.ai.
