Here we define four key terms related to CSET’s work: artificial intelligence, machine learning, deep learning, and neural networks. Below, we suggest other glossaries that define related terms. For more information about some popular types of models — generative AI, large language models, and foundation models — see our blog post explaining how these categories overlap and how they differ.

Artificial Intelligence:

Artificial intelligence (AI) is a broad term with no single authoritative definition — and frequently, people mean different things when they use it. 

However, AI commonly means: 

1. The capability of a non-human system to perform functions typically thought of as requiring human intelligence.
2. A field of study dedicated to developing these systems.

The term AI is often imprecise. AI is sometimes used interchangeably with machine learning, but the two terms are not identical. Machine learning is one promising set of techniques used to develop AI, but others exist. We disambiguate machine learning and AI more fully below. 

AI is also a moving target. The “AI effect” is a paradox in which problems thought to require AI, once largely solved, are no longer seen as requiring “intelligence.” This dynamic further contributes to ambiguity around the definition of AI.

Machine Learning: 

Machine learning is a set of techniques by which a computer system learns how to perform a task through recognizing patterns in data and inferring decision rules, rather than through explicit instructions. Machine learning also refers to the subfield of computer science and statistics studying how to advance those techniques.

Machine learning has led to most of the recent advances in artificial intelligence. These advances have been incorporated into systems used by millions of people every day, such as Google’s search and translate tools, Amazon’s digital assistant Alexa and Netflix’s movie recommendation algorithm. It also includes specialized systems such as AlphaGo, text generators like BERT and GPT-2 and game-playing systems like OpenAI Five, AlphaStar and poker-playing Pluribus.

The essence of machine learning is a system recognizing patterns of relationships between inputs and outputs. For example, the U.S. Postal Service used machine learning to train a system to recognize handwritten zip codes on mail. The system was fed images of handwritten examples paired with the corresponding numbers as typed by humans, so it learned to identify what features were common in handwritten digits and how they varied. Once trained, the system could correctly identify previously unseen examples of handwritten digits.

Not all computing is machine learning and not all artificial intelligence systems that use computing use machine learning. Many computer programs, including most commonly used software, use rule-based systems where programmers set the actions the system should take. However, machine learning is useful for applications where it is difficult for human designers to specify the correct actions to take. For example, IBM’s Deep Blue used a rule-based, exhaustive-search approach to beat the world chess champion. Deep Blue is therefore an example of an AI system not based on machine learning. On the other hand, DeepMind used machine learning to create AlphaGo, an AI system capable of out-performing humans in Go. While it is theoretically possible to solve Go with rule-based algorithms, the search space is so large that winning against a human would have been impossible. Machine learning allowed AlphaGo to infer strategies not yet discovered by humans, leading it to beat the world champion.

Neural Networks:

Deep learning is a statistical technique that uses neural networks composed of multiple hidden layers of nodes and typically trained on large amounts of data to capture patterns and relationships in data. 

Neural networks (also known as artificial neural networks or neural nets) are one common type of machine learning algorithm. They were loosely inspired by aspects of biological brains. In the brain, signals cascade between neurons; similarly, a neural net is organized as layers of nodes that can send signals of varying strengths based on the inputs they receive. Analogous to human learning, training a neural network involves adjusting how and when the nodes in different layers activate. 

At its simplest, a neural network can be made up of just three layers: an input layer where data is observed, a hidden layer where data is processed and the output layer where a conclusion is communicated. When a neural network has multiple hidden layers it is called a “deep neural network.” While shallow neural networks have uses, most of the recent advances in AI have come from deep neural networks. Because of this, in contemporary usage the term neural network usually refers to a deep neural network. The term neural network is also sometimes used interchangeably with deep learning, which technically refers to the process of training the deep neural network.

Deep Learning:

Neural networks (also known as artificial neural networks or neural nets) are one common type of machine learning algorithm. They were loosely inspired by aspects of biological brains. In the brain, signals cascade between neurons; similarly, a neural net is organized as layers of nodes that can send signals of varying strengths based on the inputs they receive. Analogous to human learning, training a neural network involves adjusting how and when the nodes in different layers activate. 

Deep learning is a subfield of machine learning that has proven particularly promising over the last decade or so and is responsible for many of the well-known developments in artificial intelligence such as in computer vision and autonomous vehicles. 

Deep learning is a statistical technique for fitting the parameters of deep neural networks. This process is often referred to as “training” the deep neural net. While neural networks can comprise any number of layers, deep learning uses neural networks with multiple hidden layers to process data, allowing it to recognize more complex patterns. More colloquially, any application using this approach is referred to as “deep learning.” 

Deep learning has shown promise in a wide range of areas including image recognition, natural language processing, photo generation, game play, robotics, self-driving cars, drug discovery and music generation. For example, in image recognition, earlier layers (those toward the beginning of the process) may identify shapes, edges and other abstract features. Later layers may identify how those shapes come together to form ears, noses and other facial features.

Other Sources:

While there is no comprehensive glossary on machine learning and AI, we recommend looking at the following documents for further information:

Machine Learning for Policymakers: In depth explanation of supervised learning, unsupervised learning and reinforcement learning written for policymakers. 

Defense Innovation Board’s AI Principles (Appendix I): Definitions of terms related to artificial intelligence and national security.

Google’s Glossary for Developers: Short explanations of a large range of machine learning terms, including simple explanations of dozens of technical terms.