Machine Learning Approaches
Machine learning is probably the most important field within modern artificial intelligence. As its name implies, it aims to provide machines the ability to learn. More specifically, it concerns itself with algorithms that can automatically learn models and patterns from data, without being explicitly told where to look.
In many applications, we are required to build models: representations of something else (e.g. a process or system) based on general rules, concepts or principles. Models can serve many purposes – for example, they can allow us to gain a better understanding of how some system works. Also, they might allow us to make predictions about the things they represent – and using these predictions, they might tell us how to take better actions.
For example, a model of the atmosphere allows us to make weather and climate forecasts. A robotic lawnmower will build a model of its environment so it can efficiently mow the lawn. Financial models allow traders to make better trades, etc.
Sometimes, it is easy to manually define rules and concepts in order to build a model that corresponds to a real-world system or phenomenon, or that explains the patterns in a certain dataset. In many cases though, this is a very difficult and labor intensive task. This is where machine learning comes in — in many of those cases, it offers a solution, by automating the task of building the model, with comparatively little instruction.
There are three main approaches of machine learning:
- Supervised learning. In this area, the goal is to learn a function or rule that maps example inputs to desired target outputs. This is done by a “supervisor” giving the algorithm example input-output pairs during training. For example, the algorithm may be required to recognize input faces, and output the corresponding persons’ names.
- Unsupervised learning. In this area, the goal is to find structure, patterns, groups, clusters, or new representations within the input data. There is no “supervisor” or any specific target. For example, an algorithm may be shown a number of various pictures, and asked to group them by how similar they are, without giving any examples of how to do this. Unsupervised learning is either a goal in itself (e.g. to find hidden patterns in the data), or a means to an end (e.g. to prepare a different representation of the data to be used as the input for supervised learning).
- Reinforcement learning. In this area, the problem concerns an agent that interacts with an environment. The agent can observe the environment, and must learn to take actions in it in order to accomplish a certain goal. For example, the agent must learn to drive a car to a specific destination, play a game, or operate a machine.
Getting started in machine learning
If you want to start learning about ML, head over to the section on supervised learning. It's a good place to start: it is easy to learn the basics, and with a base knowledge of supervised learning, you will be acquainted with many of the important concepts underlying modern AI, which you can then apply to more advanced topics like reinforcement learning.