Artificial Intelligence is that branch of computer science that tries to write programs that do things that, in humans, require intelligence. The most common expression of the AI point of view was that the brain was a symbolic processing entity. It formed new knowledge by manipulating symbols representing ideas in such a way as to form new relationships among the symbols. This is very similar to the way AI researchers viewed mathematics. In mathematics you manipulate symbols representing numerical or algebraic concepts to derive new mathematical concepts. Intelligence was seen therefore as a computational system. A computational system is made up of symbols (numbers, words, ideas, etc) and a set of operations (addition, sentence formation, inference, etc) that are used to manipulate the symbols in prescribed ways. The brain for these researchers was simply the computational engine, the hardware so to speak, that performed the operations.. For them, the important part of intelligence was the operations carried out by the brain. They reasoned that anything, a computer for example, that could carry out these operations would be intelligent. AI researchers spent most of their time trying to discover just what the symbols and operations of intelligent thought might look like. They were basically engaged in what became know as the knowledge representation problem, i.e. how do we represent knowledge and its operations to account for human abilities to manipulate knowledge? Because of their belief in the importance of symbols and their manipulation, this way of doing artificial intelligence research became know as Symbolic AI.
The earliest problems investigated by AI researchers were game playing, natural language understanding, robotics, and machine vision. All of these problems seemed fairly simple on the surface and there was quite a bit of early progress in investigating these phenomena. Investigators soon found out however, that these phenomena were far more complex than they had imagined. Investigators discovered that intelligence consisted not only of ways of manipulating knowledge but also of the shared knowledge that all humans used to make sense of their activities. This shared knowledge was made up of all the things we know about the world and our knowledge of how human interactions take place. AI investigators called this shared knowledge "common sense". All of the early knowledge representation schemes, logic, frames and scripts for example, were found to be inadequate in representing common sense. Intelligence involved knowledge that we did not know how to put inside a machine. In 1984 Doug Lenat started a 10-year project called CYC (encyclopedia) where he and other researchers attempted to spoon-feed common sense to a computer. The project was not a success and people like Herbert Dreyfus argue that this failure essentially puts an end to symbolic AI.
However, an important consequence of the realization that we could not put common sense inside our machines was the realization that we could still do useful work if we artificially restricted the domain in which the machine operated. By restricting the domain we can remove much of the ambiguity inherent in all human activities. This realization gave birth to expert systems. Expert systems are computer programs that perform some tasks in a restricted domain as well as humans. Expert systems have been built, for example, to diagnose certain types of diseases. Expert systems are often characterized as having deep but narrow minds, that is, they know a lot but only about one thing. These systems perform their appointed tasks well but are totally useless outside of their domain. Some AI researchers point to expert systems as the success story of symbolic AI.
More recently, a group of AI researchers have argued that symbolic AI was a misguided effort. They pointed out that intelligence was not just the facts you know but also the things you know how to do. Your knowledge of the world comes from your manipulation of the things in your world. Since it takes a body to manipulate objects these researchers argued that intelligence was "embodied" and that our machines would become intelligent only if they could also learn from their manipulation of the world. One consequence of this view was that researchers began to seriously investigate the biology of intelligence. These researchers built simulations of human information processing based on how they thought the brain actually worked. The brain works by encoding information as patterns of neural activations and connections. So, these researchers build artificial neurons and connected them together into neural networks. Learning in a neural net consists of modifying the pattern of connections among the neurons in the net. This group of AI researchers is known as the Connectionists. Neural nets suffer from at least two problems. The first one is that the researcher is never quite sure of what a net actually learns when presented with a learning task. Researchers often find that tuning a net for one task adversely affects tasks that the net has already learned. The second problem is that attempts to make large neural nets results in combinatorial explosion. That is, the number of possible connections between neurons becomes so large that our computers can’t handle the necessary calculations in a reasonable amount of time.