Artistic expression

« Artificial intelligence Artificial intelligence (AI) tries to make computer systems (of various kinds) do what minds can do: interpreting a photograph as depicting a face; offering medical diagnoses; using and translating language; learning to do better next time.

AI has two main aims.

One is technological: to build useful tools, which can help humans in activities of various kinds, or perform the activities for them.

The other is psychological: to help us understand human (and animal) minds, or even intelligence in general.

Computational psychology uses AI concepts and AI methods in formulating and testing its theories.

Mental structures and processes are described in computational terms. Usually, the theories are clarified, and their predictions tested, by running them on a computer program.

Whether people perform the equivalent task in the same way is another question, which psychological experiments may help to answer.

AI has shown that the human mind is more complex than psychologists had previously assumed, and that introspectively 'simple' achievements - many shared with animals - are even more difficult to mimic artificially than are 'higher' functions such as logic and mathematics.

There are deep theoretical disputes within AI about how best to model intelligence.

Classical (symbolic) AI programs consist of formal rules for manipulating formal symbols; these are carried out sequentially, one after the other.

Connectionist systems, also called neural networks, perform many simple processes in parallel (simultaneously); most work in a way described not by lists of rules, but by differential equations.

Hybrid systems combine aspects of classical and connectionist AI.

More recent approaches seek to construct adaptive autonomous agents, whose behaviour is self-directed rather than imposed from outside and which adjust to environmental conditions.

Situated robotics builds robots that react directly to environmental cues, instead of following complex internal plans as classical robots do.

The programs, neural networks and robots of evolutionary AI are produced not by detailed human design, but by automatic evolution (variation and selection).

Artificial life studies the emergence of order and adaptive behaviour in general and is closely related to AI.

Philosophical problems central to AI include the following.

Can classical or connectionist AI explain conceptualization and thinking? Can meaning be explained by AI? What sorts of mental representations are there (if any)? Can computers, or non-linguistic animals, have beliefs and desires? Could AI explain consciousness? Might intelligence be better explained by less intellectualistic approaches, based on the model of skills and know-how rather than explicit representation? 1 Historical beginnings Artificial intelligence (AI) researchers make two assumptions.

The first is that intelligent processes can be described by algorithms (sometimes called 'effective procedures'), which are rules where each step is so clear and simple that it can be done automatically, without intelligence.

This is an empirical hypothesis, which some critics of AI accept (Searle, in Boden 1990: ch.

3) and others reject (Penrose 1989).

The second is that all algorithms can be implemented on some general-purpose computer.

This assumption is generally accepted. It is based on the Church-Turing thesis, which states that a universal Turing machine, to which general-purpose computers are approximations, can compute any algorithmically computable function (see Church's thesis; Turing machines).

The best-known types of AI - classical AI and connectionism - share these two assumptions.

But they differ in other ways (see §§2-4).

Classical AI involves serial, or one-by-one, processing of (sometimes complex) formal instructions, whereas connectionism involves parallel, or simultaneous, processing among many simple units. Classical computation uses programs made up of formal rules to generate, compare and alter explicit symbol structures (see Mind, computational theories of; Language of Thought).

Connectionist computation typically uses numerical (statistical) rules to determine the activation within networks of locally interacting units and, in systems that can learn, to alter the firing thresholds of individual units and the (excitatory or inhibitory) 'weights' on their interconnections.

The system's 'knowledge' is contained implicitly in the constellation of connection weights (see Connectionism).

Some philosophers use 'computation' to apply only to the classical type, first defined by Turing in 1936.

AI researchers themselves normally use the term to cover both kinds of information processing.

Despite their differences, both these types of AI started from the same source: a seminal article written in 1943 by McCulloch and Pitts (Boden 1990: ch.

1).

This discussion of 'A Logical Calculus of the Ideas Immanent in Nervous Activity' integrated three powerful ideas of the early twentieth century: propositional logic, the neuron theory of Charles Sherrington, and Turing computability.

The authors showed that simple combinations of (highly idealized) neurons could act as 'logic gates'.

For instance, a McCulloch-Pitts neuron with two inputs could fire if and only if both inputs were firing (an 'and'-gate), or if only one input were firing (an 'or'-gate), or if some specific input were not firing (a 'not'-gate).

Since every truth-function can be expressed with 'not' and 'or' alone, McCulloch and Pitts were able to show that every function of the propositional calculus is realizable by some neural net; that every net computes a function that is computable by a Turing machine; and that every computable function can be computed by some net.

Their work inspired early efforts in both classical and connectionist AI because they appealed to logic and Turing computability, but described the implementation of these notions as a network of abstractly defined 'neurons' passing messages to their neighbours.

The neural networks discussed in this entry were extremely simple.

For example, any link always had the same amount of influence, whereas most modern connectionist systems allow for continuous changes in the weight of each connection.

But the authors' theoretical ambitions were vast.

Perception, reasoning, learning, introspection, motivation, psychopathology and value judgments: all, said McCulloch and Pitts, could in principle be understood in their terms.

The whole of psychology would in future consist of the definition of various kinds of nets capable of doing the things minds do - that is, capable of computing the sorts of things which minds compute.

Neurophysiology and neuroanatomy would show how networks are implemented in the brain, but psychology would define their logical-computational properties.

Their views on the relation between psychology and physiology (or mind and body) anticipated later developments in the philosophy of mind (see Functionalism). McCulloch and Pitts' 1943 paper made AI possible in three ways.

It influenced von Neumann, in designing the digital computer, to use binary arithmetic and binary logic (see Neumann, J.

von).

It gave both psychologists and technologists the confidence to model propositional (symbolic) reasoning, as opposed to only arithmetical calculation, on logic-based computers.

And it inspired people to start studying the computational properties of. »