[EM7407 Thomas Serre firstname.lastname@example.org re relationals]
Apr 4th 2007 Economist Magazine
Science & Technology
Easy on the eyes
A computer can now recognise classes of things as accurately as a person
[by Satoshi Kambayashi]
NEVER underestimate a computer. Never overestimate one either. For many
years Garry Kasparov, a world chess champion, said that a computer would never
beat him (or, indeed, any other human in his position). In May 1997 he had to
eat his words. Deep Blue, an invention of IBM, did just that.
This was impressive, but it demonstrated processing power
rather than intelligence. Computers are generally good at solving specific
problems, not specifically good at solving general ones. Deep Blue did not
learn to play chess from experience. It was painstakingly programmed with
thousands of “tactical weighting errors” devised by human experts. So whenever
it selected a move, it used these to work through multitudes of possible
options and their possible responses. No one is quite sure how Mr Kasparov's
processor operates but it certainly does not do that. One theory goes that the
human brain recognises strategic positions in a general way, and that this
helps to reduce the problem to a manageable size.
Thomas Serre and his colleagues at the Massachusetts
Institute of Technology have built a computer processing system that tries to
work in this general way. Among the tasks that computers are bad at is
recognising broad categories of images. Tell one to search for something
specific, such as a rectangle or even a human face, and it can make a
reasonable fist of the task. Ask it to find “animals” among photographs of
dragonflies, trees, sharks, cars and monkeys, and it falls over. Indeed a
monkey—or even a human baby—would leave it in the dust.
That, at least, was how it used to be. But as Dr Serre
describes in this week's Proceedings of the National Academy of Sciences, his
computer handles this problem rather well. In a recent test it even did a
little better than humans.
Given the briefest of glances at a picture, most people believe they have not
had time to recognise anything in it at all. Ask them whether they saw an
animal and they consider themselves to be making a futile guess. Yet those
guesses are right much more often than they are wrong. That is because the
brain can carry out immediate visual processing even when it does not have time
for any cognitive back-chatter. A neuroscientist trying to understand how
people recognise objects would thus start with this simplest of systems.
That is the purpose of Dr Serre's computer. His project is
nothing less than an attempt to reverse-engineer the relevant part of the
brain. That part is the ventral visual pathway. Anatomy shows that it is
organised into numerous areas. Experiments on monkeys, in which researchers
have recorded what excites individual nerve cells in each of these areas, give
strong hints about how it works.
The pathway is hierarchical. Signals from the retina flow to
the most basic processing area first; the cells in that area fire up others in
the next area; and so on. Those in the first area are fussy. They react to
edges or bars in particular orientations. By combining their signals, however,
cells in the second area can respond to corners or bars in any orientation. And
so the system builds up. Cells in the final area can recognise general things,
Dr Serre considered his computer's processing units analogous
to nerve cells, and he organised them into areas, just as they are in real
brains. Then he let the machine learn in much the same way that babies do.
First he mimicked early development when nerve cells are plastic. At this stage
babies' brains tune their nerve cells to visual features according to how
common those features are in the world around them. That is why kittens raised
so that they see only vertical lines have brains that look different from those
raised in an environment with purely horizontal ones. Dr Serre's processor
developed sensitivities in a similar fashion when he showed it lots of
photographs. That stage complete, he then told the computer when what it “saw”
contained an animal, and when it did not.
The result was a model that closely imitates the ventral
visual pathway. Processing units in each area are sensitive to the same set of
features as nerve cells in the brain's analogous areas, and they are linked
together as they are in the brain. This artificial recognition system correctly
distinguishes photographs containing animals from those without creatures 82%
of the time; Dr Serre's students get it right 80% of the time. Moreover, his
computer and his volunteers tend to slip up on the same images—and turning
photographs on their sides makes poorer animal-recognisers out of both, by
roughly the same amount.
A system like this has obvious applications (it may, for
instance, soon be put to use searching for child-pornography sites on the
internet). But it also brings more subtle benefits. Based as it is on how
brains work, it may give insights into what happens when they go wrong. Real
neuroscientists rely on lesions (that is, damaged areas of a brain) to help
them understand what is going on in brains by seeing what happens in response
to particular sorts of damage. Dr Serre has therefore “lesioned” his computer
system in similar ways. So far, this has demonstrated the importance in visual
recognition of the rare connections that bypass a unit or two in the hierarchy.
A computer chess-player could not have told you that.