This Is How Machines Learn! (Part 1)
An Introduction of the underlying Ideas in Machine Learning
Our goal with this series is to enable everyone to understand AI phenomena in their daily lives, as well as to actively shape the growing influence of AI on our society. Therefore, we do not consider any technical details or provide an introduction on how to use certain machine learning frameworks. Instead, we focus on explaining the underlying ideas of machine learning which empower everyone to understand and shape the digital world that surrounds us.
How can cars “learn” to drive autonomously? How do computers recognize cancer cells? And does my online shop know what I want to buy? Recent progress in the field of AI is largely driven by machine learning — a subfield of AI that deals with algorithms that improve through experience over time, i.e. “learn”.
Machine learning algorithms differ from typical algorithms that power web servers or describe how to prepare a meal, where we can follow step by step how input data is converted into output data. It also differs from more traditional approaches to artificial intelligence in which knowledge is explicitly represented, e.g. by modeling the underlying facts and rules to draw logical conclusions. In machine learning, rules, behavior or patterns are instead derived based on (typically large amounts of) data — in other words, they are “learned”. The knowledge that was learned is stored in a so-called model. While in classical AI, our task is to model knowledge in such a way that it can be explicitly represented and processed in a computer system, in machine learning, data needs to be prepared and a (general) machine learning procedure needs to be configured in such a way that a model can be derived from the data.
Machine learning is mainly used when knowledge cannot be explicitly represented in a way that computers can process. If we want to recognize a cat in a picture, we look for four paws, fur, whiskers, and so on. Writing down explicit rules for all of this, however, is far from easy. What is a general description for recognizing four paws, fur, whiskers, and so forth, for a computer to process? Also, we must consider that not all four paws might be visible at the same time on a given picture. Although recognizing a certain animal in a picture is easy for humans, it is difficult, if not impossible, to represent the knowledge required for this task explicitly. By using machine learning, the computer itself can identify rules that help recognize cats. These rules are not necessarily those that we would have used as humans.
But how exactly can machine learning be used to find such rules and patterns and thus recognize a cat? How can these rules be transferred to new, unknown data? In other words: How does a machine learn?
This series addresses these questions. Therefore, we will distinguish three ways of how machines can learn:
- through supervised learning using labeled data (part 2)
- through unsupervised learning using unlabeled data (part 3)
- and through reinforcement learning using rewards and punishments (part 4)
We will use a little robot to visualize the different types of machine learning. This robot is merely an analogy for a computer executing a machine learning algorithm. But it will help us grasp the underlying idea of what machine learning is all about. Finally, we’ll also explore how machine learning affects our society (part 5).
Go to the next part.
Written by Stefan Seegerer, Tilman Michaeli and Ralf Romeike.
The robot is adapted from https://openclipart.org/detail/191072/blue-robot and licensed under CC0.