Å®ÉúСÊÓƵ

THE goal of modern neural computing is to design systems which will
be able to learn in a human-like way. For example, the WISARD pattern recognition
machine, developed at Brunel University in London, can be ‘trained’ by reading
in data from an optical scanner. When the machine is subsequently presented
with an unknown pattern it uses a system of parallel processors to compare
it with the known patterns of the training set, allocating each possibility
a weighting depending on the degree of match. The known pattern with the
highest score is selected as the answer to the problem.

This ability to plump for the best available answer, instead of looking
for an exact match, enables the machine to cope with complex real-world
identification problems, such as reading in printed characters. Standard
algorithmic (rule-based) systems need to be highly sophisticated to achieve
this, because of the extent of variation even within a single typeface due
to imperfections in the original copy. More complicated problems, such as
the recognition of individual faces, are probably impossible to solve by
algorithmic methods, because the number of rules or the database of stored
images is just too huge. This is what the mathematician Sir James Lighthill
called the ‘combinatorial explosion’ – when algorithmic solutions are applied
to complicated, variable problems, the number of possibilities simply becomes
overwhelmingly large.

If slavishly algorithmic methods are a problem for computers, my guess
is that they are even more so for human beings. I only have anectdotal evidence,
but four years spent training people to operate word processors have left
me with a distinct impression that those who have the most difficulty are
the ones who want rigid lists of procedures to follow in every eventuality.
These people ignore as many as possible of the more ‘advanced’ facilities,
thus reducing the learning problem to a manageable size and avoiding their
own personal combinatorial explosion.

This second reaction also has parallels in the field of artificial intelligence.
Algorithmic methods have appeared to succeed where programs are required
to operate in artificially simple ‘toy universes’, but programmers have
failed to produce versions capable of dealing with the unsimplified universe
outside.

Computer users who have become hooked on rule-based learning often don’t
appear to realise that there is any possible solution other than gigantic
feats of memorisation. They marvel at the wonderful memories of successful
computer users and wish that they could do as well. However, the successful
users probably just remember roughly what sort of procedure is needed, then
either look up the details or press buttons and see what happens. Eventually,
of course, what works gets ‘hard-wired’ into our brains for things we do
a lot and we stop having to refer to a crib, but we certainly don’t memorise
long sequences of action right off.

Maybe a return to more rote learning in schools isn’t such a good idea
after all. I’d certainly be interested to know if there’s a difference between
discovery-method and rote-learning education as preparation for learning
to operate complicated systems later on in life.

Rosemary Rodd works in the Literary and Linguistic Computing Centre
at the University of Cambridge.