Artificial Intelligence

This book was written to repair a specific imbalance. Artificial intelligence is at present discussed more widely than any technical subject in living memory and understood, in proportion to the discussion, less than almost any. The reasons are not mysterious. The field moves quickly; its vocabulary is borrowed from psychology, statistics, and engineering in roughly equal parts and used in ways that none of the lenders would recognise; and the systems that have captured public attention are, by construction, easier to use than to explain. A reader who wishes to know what these systems are - not what they can be made to do in a demonstration, but what they are, in the sense that a physician knows what a kidney is or an engineer what a bridge is - will find a great deal of instruction and very little explanation. The purpose of this treatise is to supply the explanation. Its method follows from its purpose. Nothing in the pages that follow is asserted without being derived, and nothing is derived from anything that has not itself been introduced. When an equation appears it has been built, in view of the reader, from definitions stated a page or a chapter earlier; when an algorithm is named it is stated in full and exercised on a worked example; when a claim is made about what a system can or cannot do, the claim is tied to an evaluation whose design the reader has been shown how to criticise. This discipline makes the book longer than a survey would be and, the author hopes, more useful than one, since a survey tells the reader what is believed and a derivation tells the reader why. The twenty chapters are arranged in five parts, moving from the philosophical and historical foundations of the subject, through its mathematics and the classical methods of reasoning under uncertainty, to machine learning, deep neural computation, generative and foundation models, and the multimodal, embodied, and production systems in which they are deployed. Each chapter closes with exercises whose solutions are given in full, and with references in the style of the American Psychological Association, so that every claim can be traced to its source and every method reproduced.The manner of the argument is dialectical, and the reader should know in advance what that means here. Where the subject is contested - and there is scarcely a page on which it is not - the book states the strongest form of each position, sets them against one another, and reports the resolution that the evidence supports, or the absence of one. This is slower than announcing a conclusion. It is also, in the author's experience of teaching the subject to clinicians, engineers, and graduate students, the only method that leaves the reader able to think about the next development rather than merely informed about the last. Two further features require a word. The first is that computations are carried out, wherever the reader might reasonably wish to reproduce them, on a scientific calculator whose keystrokes are shown in full, and in free software whose code is given. The choice is deliberate: a result that can be obtained on a device that fits in a coat pocket is a result the reader owns. The second is that the illustrations, graphs, and diagrams throughout the book were produced by the author for it, from the equations in the text, and the legends say so; where a figure reproduces external material the source is given. A picture in a technical book should be as accountable as a sentenceThe book is addressed to graduate students in computing, mathematics, and the quantitative sciences; to researchers and engineers who use these systems and wish to understand them; to physicians and other professionals whose work is being transformed by them; and to the serious general reader who is prepared to follow an argument through its derivation.