How machines learn
- Article
- Machine Learning
Machine learning is the branch of artificial intelligence that builds programs which improve from data instead of from hand-written rules. A conventional program follows instructions a person wrote. A machine learning model is given examples and a way to measure its mistakes, and it adjusts itself until the mistakes get small. IBM’s explainer on the topic, which this post summarizes, lays out the process in three parts.
Three parts of a learning system
The first part is a decision process. The model takes an input, such as the pixels of an image or the words of an email, turns it into numbers, and produces a guess. The guess might be a label (spam or not spam), a number (tomorrow’s temperature), or a grouping (which customers behave alike).
The second part is an error function, often called a loss function. It compares the guess with the correct answer and turns the difference into a single number. A model that guesses well gets a low score, a model that guesses badly gets a high one.
The third part is model optimization. The model contains many adjustable parameters. An optimization algorithm nudges those parameters in the direction that lowers the error, then checks again. Repeated over thousands or millions of examples, this is what “learning” means in practice: the parameters settle on values that make the guesses accurate.
Three ways to learn
Supervised learning uses labeled data. Every example comes with the right answer, and the model learns to reproduce it. Most everyday applications work this way, from spam filters to systems that read handwritten postal codes.
Unsupervised learning uses data without labels. The model looks for structure on its own, such as clusters of similar records or a smaller set of variables that explain most of the variation. This is useful when nobody has labeled the data, which is the usual case.
Reinforcement learning has no fixed answer key at all. An agent takes actions in an environment and receives rewards or penalties, and it learns a strategy that collects the most reward over time. Game-playing programs and robot controllers are the common examples. Semi-supervised methods sit in between, mixing a small labeled set with a large unlabeled one.
Why it matters
Machine learning already sits behind speech recognition, product recommendations, fraud detection, and medical image screening. The explainer also lists the parts that are harder than they look: getting data into a clean and consistent form, making sure a model works on data it has not seen, the cost of the computing needed to train large models, and monitoring a deployed model so it does not quietly degrade as the world changes.
The original Facebook post credited the 2021 version of the IBM page by Tucci and Burns. IBM has since rewritten the page, and the link below goes to the current version.
Source: What is machine learning? IBM