Tuesday, 22 March 2016

Are Artificial Intelligence programmes really so clever?

Lee Se-dol lost three games in a row
When he played against slick Alpha-Go
Well he won number four
But the program won more,
'Cause mere humans are really too slow.

Draughts was the first of the classic board games to succumb to the power of computers, then Chess, and now Go. So congratulations to all the people who have worked on the design of the Alpha-Go software and the computer on which it runs.

But what does it really tell us about intelligence. It tells us that humans are persistent and given a very clear goal - however artificial and unimportant in any real sense - they will work hard to achieve that goal. It might be a personal challenge such as swimming the English channel, or winning a gold medal at the Olympic games - or a technological challenge, such as sending a man to the moon, or building an unbeatable game-playing machine. Many people have spent time thinking about how to write programs to play Chess and Go and world-wide many millions of man-hours must have been expended on refining the task, making improvements as a result of feedback, using ever more powerful computers. At last they have succeeded.

But surely the whole point about human intelligence is that we are flexible and can operate with incomplete and uncertain information, while the machines we are building are very specialised. Most of the modern AI systems make use of some form of big data - vast amounts of well formatted information processed with sophisticated statistical tools. Alpha-Go was not only supplied with information on large numbers of human against human games of Go - and also played millions of games against itself. Of course Alpha-Go did loose one game - but the reason has been noted by the clever human programmers - and this weakness is being addressed.

Do the computer programs inside Alpha-Go tell us anything about human intelligence and which humans are "clever"? Probably not. Let us assume that we wanted to learn about how humans walk and rather than concentrate on everyday activities we decided that we would start building a machine tot get to the top of Mount Everest in less time than it takes a human being skilled in mountain climbing. In choosing this task we deliberately ignore the 99% of humanity who could never climb to the top - but researchers want to start by choosing something they feel is spectacular. But of course a machine which tool the form of a helicopter - or had hands with fingers which could drill holes in the solid rock to ensure a firm grip - or which was not affected by the rarefied atmosphere, tell us absolutely nothing about the human ability to climb mountains.

My contention is that Alpha-Go uses powerful mathematical tools, and significant computing power, in a way which means that the word ARTIFICIAL in Artificial Intelligence needs to be highlighted because it has little to do human intelligence.

My approach, based on my research with CODIL is say that any model most be meaningful in evolutionary terms, working upwards from very simple pattern marching systems. It must be able to switch between a wide variety of open-ended and poorly defined mathematically unsophisticated tasks. It must also be fit in with the known limitations of the human brain - such as:
  1. The limit of about six concepts at any one time in the conscious short term memory.
  2. The ability of "automate" frequent learned patterns so that they can be processed unconsciously.
  3. Logical flaws which lead to problems such as confirmation bias.
  4. An unreliable long term memory which has gaps - which it creatively fills when trying to remember what happened.
  5. The brain processes patterns and not predefined rules operation on well-formatted sates (i.e. it is not a stored program computer).
The point I want to make is that there are many very complex applications that computers can do - which are impractical if carried out by the unaided human mind, because humans would be too slow, or too inaccurate, or could not handle the volume of data. For historical reasons some of the least practical applications are called "artificial intelligence" while many others, which help keep modern society running, are not. We should be very careful not to confuse attempts to model how humans process information with much of the work which goes under the name of "artificial intelligence."

No comments:

Post a Comment