1 Foundations of Artificial Intelligence

What artificial intelligence means

Artificial intelligence (AI) is a broad field concerned with building computer systems that perform tasks associated with abilities such as perception, language use, reasoning, learning, planning, and decision-making. There is no single definition used in every context. A practical description is a machine-based system that produces outputs—such as predictions, recommendations, or decisions—in pursuit of human-defined objectives.

AI describes a system's capabilities and methods; it does not prove that the system thinks or understands as a person does, or that it has human-like consciousness. Many AI systems are designed for limited tasks. Broad, human-level capability across many domains is a different and much more ambitious idea.

Common AI terms

Common terms describe approaches and components, but they are not requirements for every AI system. Some systems rely chiefly on explicit rules, others use models trained from data, and many combine both.

• Algorithm: A procedure or set of steps for solving a problem or carrying out a computation.

• Data: Information used by a computer system. In machine learning, examples in data may help a system learn patterns.

• Model: A computational representation used to produce an output, such as a prediction or classification.

• Training: The process of adjusting a model using data so that it performs a task more effectively.

• Inference: Using a trained model to produce an output from new input.

• Narrow AI: AI designed for a particular task or limited set of tasks. Recognizing speech is one example, and most current AI applications are of this kind.

• Artificial general intelligence (AGI): A proposed form of AI that could perform a wide range of intellectual tasks at a broadly human level. It is distinct from today's task-specific systems.

How the field developed

Ideas underlying AI predate the field's name. In 1950, mathematician Alan Turing raised the question of whether machines could exhibit intelligent behavior. In a 1955 proposal, John McCarthy and colleagues proposed a summer research project to investigate how machines might simulate aspects of learning and intelligence. The 1956 Dartmouth workshop that followed is commonly regarded as the beginning of AI as an organized research field.

Early AI research explored representing knowledge and solving problems with symbols and explicitly specified rules. Later, researchers increasingly developed methods that let computers learn patterns from examples. Neural-network research is one strand of this work, and improvements in data and computing helped make machine-learning approaches more capable across a range of tasks.

This history is not a simple replacement of one method by another. Rule-based and learning-based approaches remain useful for different problems.

Rules and learning from examples

All computer programs follow instructions, so AI is not separate from software in the sense of using different physical machinery. An important distinction is often how the system's behavior is specified.

In conventional, rule-based software, a developer writes explicit instructions for the cases the program should handle. For example, a program could apply a stated rule to grant a discount when a purchase exceeds a specified amount. Its behavior follows those programmed conditions.

In a machine-learning system, developers choose a model and a learning procedure, then provide examples. During training, the process adjusts the model to capture patterns in those examples. For instance, a model trained on labeled email examples may estimate whether a new message is spam. It does not need a programmer to write a separate rule for every possible message, but its output is an estimate based on what it learned, not a guarantee.

Strengths, limits, and combinations

The contrast between conventional software and machine-learning systems is useful, but not absolute. Conventional programs may contain many complex rules, and AI systems still depend on human choices about their objectives, data, design, and evaluation.

Rule-based software can be easier to inspect and predict in well-defined cases. Learning-based systems can handle patterns that are difficult to specify by hand, but their behavior may be less transparent and may vary with input. Many products combine learned models with ordinary software logic.

AI includes methods ranging from explicit rules to models learned from data. Its key distinction from conventional rule-based programming is often that some behavior is learned from examples rather than written out as a rule for every case.