What is machine learning?

What changes is a set of numbers used in a calculation. What machine learning builds is called a “model”: a calculation that takes something in (a photo, an email) and gives an answer. The numbers it uses are called “parameters.” For example, in a model that predicts rainfall from temperature and humidity, the numbers say how much weight to give each clue (an example from Google’s introduction to machine learning).

At the start, the numbers are set to random values near zero. Then it repeats three steps: ① give an answer with the current numbers, ② compare it with the right answer and measure how wrong it was (the “loss”), and ③ move the numbers a small amount in the direction that makes the mistake smaller. When the mistakes stop shrinking, learning is done. This method is called “gradient descent.” Google’s course says the goal of training is to minimize the loss, reducing it to its lowest possible value.

People do not set the numbers one by one. In large language models (LLMs), there can be hundreds of billions or even trillions of them (Google’s course; see the “Large language model” page).

Google’s introduction to machine learning lists the main kinds as supervised learning, unsupervised learning and reinforcement learning, plus generative AI (models that create new content from what the user puts in).

In supervised learning, the model sees many examples with the right answers (“labels”) and learns which clues lead to those answers. Some models predict a category, like “spam or not” or “is there a cat in this photo”; others predict a number, like the amount of rain or the price of a house.

In unsupervised learning, the model looks for meaningful patterns in data that has no answers attached. A common technique is “clustering,” grouping similar things together. For example, clustering weather data by temperature might reveal groupings like the seasons. Naming the groups is up to people who understand the data.

In reinforcement learning, the system tries actions, gets rewards or penalties, and finds the way of acting that earns the most reward. It is used to train robots to do things like walking, and programs like AlphaGo to play the game of Go. It was also used to teach reasoning models how to think (see the “Reasoning model” page).

It is one way of doing machine learning, using a kind of calculation called a “neural network.” A neural network is built from layers, and the layers are made of parts called “neurons.” The layers between what goes in (the clues) and what comes out (the answer) are called “hidden layers,” and Google’s glossary calls a neural network with more than one hidden layer a “deep model.”

For example, in 2012, Alex Krizhevsky and colleagues at the University of Toronto trained a neural network with eight layers and 60 million parameters to sort 1.2 million photos into 1,000 categories. In a contest to name what is in photos (ILSVRC-2012), a version of it got a top-5 error rate of 15.3%: the share of photos where none of the AI’s five guesses was right. The second-best entry got 26.2% (lower is better).

Large language models are deep learning too. Their most widely used design, the “Transformer,” is an enormous neural network that stacks layer upon layer (Google’s course).

It may be “overfitting.” Google’s course describes overfitting as matching (memorizing) the data used for learning (the training data) so closely that the model fails to make correct predictions on new data. It compares an overfit model to an invention that performs well in the lab but is worthless in the real world.

To spot it, you keep some data aside that is not used for learning, and during learning you compare the size of the mistakes on both. At first, both fall in a similar way. If, from some point on, mistakes on the training data keep falling (or hold steady) while mistakes on the held-out data start rising, that is a sign of overfitting.

It is caused by one or both of two problems: the training data doesn’t represent real data well, or the model is too complex. Google’s course advises testing a model on different examples from the ones used for training, and deleting any test examples that duplicate training examples.

Because machine learning learns from the examples it is shown, skewed examples can lead to skewed answers. Google’s course gives this example. A model is trained to predict future sales of a new product from phone surveys of people who bought the product. People who bought a competing product instead were not surveyed, so they are missing from the training data.

Skew has been found with face photos too. In 2018, researchers Joy Buolamwini and Timnit Gebru showed that two photo collections used to test face analysis systems (IJB-A and Adience) were overwhelmingly made up of lighter-skinned people (79.6% and 86.2%). They then built a collection balanced by gender and skin type and tested three companies’ systems that guess gender from a face. Error rates were as high as 34.7% for darker-skinned women, and at most 0.8% for lighter-skinned men.

In 2019, NIST, a US government research lab, also tested 189 face recognition programs from 99 developers. When checking whether two photos show the same person, the rate of wrongly matching two different people was often 10 to 100 times higher for Asian and African American faces than for white faces, depending on the program. But some programs made in Asian countries showed no such large gap between Asian and white faces. The report’s main author, noting that the study did not look into what causes the gaps, called this “an encouraging sign that more diverse training data may produce more equitable outcomes.”

Email, for one. Google wrote in 2019 that, with its machine learning models and other protections, Gmail blocks more than 99.9% of spam, phishing and malware from reaching inboxes, and that with its machine learning tool TensorFlow it was blocking about 100 million additional spam messages a day (both are the company’s own claims). Machine learning helps, it says, by finding patterns in large amounts of data that the people who write the rules might miss.

Speech recognition, which turns voices into text, and systems that recognize what is in photos have also been built with machine learning. In a 2006 essay, Tom Mitchell of Carnegie Mellon University wrote that the commercial speech recognition systems then available all used machine learning in some way, because training a system gives more accurate results than programming it by hand. He also wrote that over 85% of handwritten mail in the US was sorted automatically, using handwriting software trained with machine learning (all as of 2006).

In the same essay, Mitchell named two kinds of jobs where machine learning was already the best method available: jobs too complex for people to design the steps by hand, and jobs where the software must adapt to its users or surroundings. He also wrote that some software will remain where machine learning may never be useful, such as programs for multiplying matrices.

In July 1959, Arthur Samuel published a paper in IBM’s research journal titled “Some Studies in Machine Learning Using the Game of Checkers.” It opens by saying the studies were about programming a computer to behave in a way that, if people or animals did it, would be described as learning. Its abstract says the computer could learn to play a better game of checkers than the person who wrote the program, and that this took about 8 or 10 hours of the machine playing.

There are several ways to define it. On the page for his 1997 textbook “Machine Learning,” Tom Mitchell of Carnegie Mellon University describes machine learning as the study of computer algorithms that improve automatically through experience. Google’s glossary calls it a program or system that trains a model from input data.

In his 2006 essay, Mitchell put it more precisely: a machine learns with respect to a task T, a performance measure P and a type of experience E, if its performance P at task T reliably improves after experience E. Applied to sorting spam, T is “sorting emails,” P is “the share sorted correctly,” and E is “seeing labeled emails” (an example made up for this page).

In the same essay, Mitchell wrote that while computer science had focused mainly on how to program computers by hand, machine learning asks how to get computers to program themselves, from experience plus some initial structure.