Notes

Starts with Prof. James B. Wiesner, director of Massachusetts Institute of Technology 's Electronics Research Laboratory, in a television talk show, 1961.
[scenes of man playing chequers against a computer]
Prof. W. suggests that computers might be able to be programmed to do things that in human beings requires intelligence.

This idea originally proposed by Alan Turing in 40's. - a decade later MIT (Marvin Minsky and John McCarthy) sets up department to explore Artificial Intelligence (AI). Young mathematician Jim Sleight programs computer to do calculus - based on about a hundred rules/suggestions. Computer gets 'A' grade in MIT exam.

Having gone from 'number crunchers to machines which could handle logic and play games, hopes were high. It was thought that machines could become philosophers!

General view that 'thinking' located in 'abstract realm of mind' processed in physical brain.
[images of children in classroom imagining that they are on the moon - to demonstrate human ability to hold ideas in the mind and manipulate them as though they are the real thing]
"Ideas can be real or imaginary" - Likewise, computer can be used to conjure up real and imaginary worlds [image of flowers shooting seeds into space].

Rationale adopted that if brain can also be a mind then so can computer - seen as irrelevant that the brains hardware (wetware?) - neurones - was completely different from the computer's vacuum tubes.
[image of neurones].
Thoughts and ideas which are manipulated seen as main issue.
Prof. Edward Feigenbaum suggests that AI pioneers were not interested in how brain is structured.
Mind seen as analogous to software - brain as hardware, a symbolic processing unit.

[image of man trying to fly with a flapping-wings contraption] "It doesn't matter at all how the brain makes intelligence - we don't have to model the brain any more than we have to make something that flaps its wings in order to fly"

Early (1960's) experiment to have machine interact with world - machine with gripper for hand, TV camera for eye - set task to stack blocks (child's play) - more difficult than imagined - problems getting machine to recognise blocks - different angle, shadows, surface patterns and textures, only the start.
Set task to build tower of blocks, machine started with top-level block because it didn't know block would fall - which is obvious to a human two year old.
In 1971, in Edinburgh, a computer-machine takes ten minutes to recognise a cup.
In America, a huge computer takes many hours to manoeuvre a machine-cart through an area strewn with objects - a space a toddler would have no trouble with.

Many AI practitioners decided to avoid area of vision and physical interaction with world , to concentrate on 'disembodied mind'.

'Turing Test' devised by Alan Turing in early 50's - a test to demonstrate machine intelligence (not yet passed!)
Operator (you) sit at a terminal which provides a connection to another entity somewhere else (in another room) - the entity may be another human sitting at a terminal or it may be a computer. - by question and answer you must determine whether the other entity is human or not.
When the other entity is a computer and can use language intelligently enough for the questioner to be unable to be sure whether the correspondent is a person or not, then the computer will have passed the Turing Test and can be deemed intelligent!

Joseph Weisenbaum's 'Eliza' was one of the first programs to appear intelligent [demonstration] - reacted to key words - could be easily fooled.
Turns out that human ability to understand spoken and written word is vastly more complex than understanding required to do calculus
[example of 'Electronic Brain', $500,000 worth in 1963, attempting to translate Russian into English - problems - eg: with ambiguity.

Humans everywhere share vast pool of common knowledge - what we are, what we feel.. We are so good with language that before computers came along we did not understand how complex is the task of understanding language.
Marvin Minsky: "What we began to see is the things that people think are hard are actually rather easy .. " and vice-versa - the easy taken-for-granted things are really hard.
Calculus can be done with a few hundred pieces of code; but recognising faces, walking about, putting on clothes - the things we expect a child to do - cannot yet be done by robots ... (though NB that the first, 'recognising faces', is being done now, with security cameras in eg: Marks and Spencers).

1972 , 'What Computers Can't Do', by Hubert Dreyfus coincides with drying-up of funding for AI research. But Minsky and others don't give up. Terry Winograd develops program (SHRDLU) to use language intelligently in 'microworld' of simulated blocks..

Edward Feigenbaum and colleagues applied Winograd's concept to develop 'Expert' Systems for areas where knowledge was deep but narrow - specialities with little or no ambiguity in the language.- based on a few hundred or thousands of pieces of knowledge.

MYCIN - analyses chemical spectrograms; DENDRAL is expert on bacterial infections (see eg:: 'Understanding Computers and Cognition' by Terry Winograd & Fernando Flores.)

However human knowledge is broad and shallow.
Examples of attempts to get computers to understand simple children's stories (filled with 'unsaid' things which are too obvious to need saying).
Strategy devised to create 'scripts' or 'frames' which provide context for words , and include intuitive, background, and common-sense knowledge - ongoing attempt to create encyclopaedic database for computer 'intelligence' - CYC

But is this enough - lot of what we know is skill-based and body-experienced. - eg:: child's play with sand, water, blocks, - building up necessary 50,000 experienced behaviour cases. But then what about clear intelligence of disabled without ability to use body but still able to develop shared common-sense knowledge?

Return to looking at structure and models of brain. Attempts (successful to have computers learn with neural networks. - 'connectionists'. [example of van taught to drive itself]
But how does learning happen - we don't know.
[example of identification of tanks in photographs !!!]
[example - NetTalk]
Prospects of neural net successes long term - problems with scaling-up experiments.- explosion in learning-time with larger nets.

Brain has solved the problem - appears to be collection of integrated special-purpose (language, vision, movement) mechanisms - as opposed to a general purpose machine.

"So far AI is mainly a history of fascinating failures" - although we have:

• chess-playing systems which can beat all but the very best.
• Expert and Knowledge-based systems.
• rudimentary robots
• computers which can read books to the blind
• limited translation machines

Doug Lenat's CYC program still going strong - trying to capture the common-sense of the world piece by piece and develop a dis-embodied mind. [example of CYC searching for inconsistencies in its knowledge overnight]. CYC appears to be an 'alien' intelligence.

If it succeeds (it is a high-risk high-payoff bet) CYC will empower human beings in undreamed of ways.