AI Fundamentals

Despite decades of research, there is no single, universally agreed-upon definition of artificial intelligence, and different researchers have historically approached the field from different angles. In their widely used textbook Artificial Intelligence: A Modern Approach, Stuart Russell and Peter Norvig organize historical definitions of AI along two dimensions: whether a system is judged by its thought processes or by its behavior, and whether the standard of success is fidelity to human performance or to an idealized standard of rationality. Crossing these two dimensions yields four categories of definitions: systems that think like humans, systems that think rationally, systems that act like humans, and systems that act rationally. The "thinking humanly" approach, sometimes called cognitive modeling, is illustrated by John Haugeland's 1985 description of AI as "the exciting new effort to make computers think . . . machines with minds, in the full and literal sense." The "thinking rationally" approach traces back to Aristotle's work on syllogisms and formal logic, and is captured in Patrick Winston's 1992 definition of AI as "the study of the computations that make it possible to perceive, reason, and act." These two "thinking" categories judge a system by the accuracy of its internal reasoning process rather than by what it ultimately does.

The remaining two categories judge AI systems by their behavior rather than their internal reasoning. The "acting humanly" approach is closely associated with the Turing test and is reflected in Ray Kurzweil's 1990 definition of AI as "the art of creating machines that perform functions that require intelligence when performed by people." The "acting rationally" approach, by contrast, defines AI as the design of agents that act to achieve the best expected outcome given what they know, a view echoed in George Luger and William Stubblefield's 1993 description of AI as "the branch of computer science that is concerned with the automation of intelligent behavior." Russell and Norvig adopt the rational-agent approach as the organizing theme of their textbook, because it can be stated as a precise engineering goal and applies even to environments, such as mathematics or logistics, where there is no obvious human standard of comparison. Under this view, an intelligent agent need not think or act the way a person would; it only needs to select actions that are expected to best achieve its goals given its available information. This distinction between imitating humans and behaving rationally becomes important later in the course when we compare AI systems to human cognition and to formal notions of optimality.

The field's name was coined in a 1955 funding proposal titled "A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence," written by John McCarthy of Dartmouth College, Marvin Minsky of Harvard University, Nathaniel Rochester of IBM, and Claude Shannon of Bell Telephone Laboratories. The proposal, dated August 31, 1955, made the bold conjecture that "every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it," and it proposed a two-month workshop to be held at Dartmouth College in the summer of 1956. That 1956 workshop is now widely regarded as the founding event of AI as an organized field of research, and initial optimism was high: in 1957 Herbert Simon predicted that within ten years a computer would become world chess champion. Progress proved harder than expected, and by the early 1970s reduced funding followed setbacks such as the United Kingdom's 1973 Lighthill report, which criticized AI's inability to handle the "combinatorial explosion" of real-world problems and led the British government to sharply cut AI research funding, a slowdown later nicknamed an AI winter. The 1980s saw a resurgence built on knowledge-based expert systems, followed by a shift from the late 1980s onward toward more rigorous statistical and probabilistic methods and a renewed interest in neural networks after the rediscovery of the back-propagation training algorithm. Most recently, deep learning has driven major advances since 2012, when a convolutional neural network called AlexNet, built by Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton, won the ImageNet Large Scale Visual Recognition Challenge with a top-5 error rate far below its closest competitor.

In 1950, the British mathematician Alan Turing published a paper titled "Computing Machinery and Intelligence" in the philosophy journal Mind, opening with the sentence, "I propose to consider the question, 'Can machines think?'" Rather than trying to define the words "machine" and "think," Turing proposed replacing the question with a more precise one based on what he called the "imitation game." In Turing's original game, an interrogator communicates by written questions with two hidden parties, one of whom is later replaced by a machine, and the interrogator must decide which respondent is the machine and which is the human. Communication is restricted to text, for example through a teletype, so that the interrogator cannot judge based on voice or appearance and must rely solely on the content of the answers. Turing suggested that if a machine could regularly succeed at making an average interrogator misidentify it as the human, that success should count as evidence of machine intelligence, and he speculated that by the year 2000 machines might be good enough to keep an average interrogator's chance of correct identification below about 70 percent after five minutes of questioning.

As the Stanford Encyclopedia of Philosophy's entry on the Turing test explains, the game is fundamentally a test of a machine's ability to imitate human conversational behavior rather than a direct test of thought, consciousness, or understanding. Passing the test does not, by itself, prove that a machine reasons the way people do internally; it only shows that the machine's outward behavior in that narrow conversational setting is indistinguishable from a human's. This gap between behaving intelligently in one setting and possessing broad, flexible intelligence motivates an important distinction covered throughout this course: narrow AI versus general AI. Narrow AI, sometimes called weak AI, refers to systems built to perform one task or a limited family of related tasks, however impressively, without the ability to transfer that competence to unrelated problems. Artificial general intelligence, or AGI, refers to a hypothetical system with the kind of broad, flexible intellectual competence that lets humans learn and reason across many different domains.

The Stanford Encyclopedia of Philosophy's entry on artificial intelligence illustrates narrow AI with the example of IBM's Deep Blue, which defeated world chess champion Garry Kasparov, and IBM's Watson, which won at the quiz show Jeopardy!, noting that neither system could transfer its expertise to a new but related game without being rebuilt. Every AI system studied in this course, from chess-playing programs to today's large-scale conversational systems, is a form of narrow AI, since no system yet demonstrates the general, cross-domain competence associated with AGI. To organize the rest of this course, it helps to preview the major subfields that make up modern AI, starting with search and planning algorithms that find sequences of actions toward a goal, and knowledge representation and reasoning methods for encoding and manipulating facts about the world. A third core subfield, machine learning, was precisely defined by researcher Tom Mitchell: a computer program is said to learn from experience E with respect to a class of tasks T and a performance measure P if its performance at tasks in T, as measured by P, improves with experience E. Beyond these core areas, the field also includes natural language processing, which enables machines to understand and generate human language, computer vision, which extracts meaning from images and video, and robotics, which connects perception and reasoning to physical action in the world. Each of these subfields will reappear in later modules of this course, and together they form the toolkit that modern AI practitioners draw on to build systems that act rationally within their environments.