Document Sample
108 Powered By Docstoc
					The following appeared in the February 2, 2001 issue of ELECTRONIC TELEGRAPH:

Mike Lynch is leading the very 21st-century race to make com-
puters think for themselves. But the formula he's using is 250
years old, the brainchild of a clergyman who died in obscurity.
                                    BY ANDREW ANTHONY

In the corner of a small graveyard, an exhaust        For a quarter of a millennium Bayes's theorem,
plume away from the wake-the-dead roar of             or Bayes's rule as it is sometimes called, enjoyed
London's Old Street roundabout, lies the tomb         a limited and mostly discredited role in statisti-
of an obscure 18th-century clergyman and statis-      cal mathematics. Recently, however, with the
tician named Thomas Bayes. Among his com-             advent of cheap and available computers, its in-
panions in Bunhill Fields, the cramped resting        fluence has rapidly spread beyond the dull
place of many of Britain's most notable religious     grind of statistics to become something akin to a
and political nonconformists, are Daniel Defoe,       philosophical movement, with an almost theo-
John Bunyan and William Blake. If his growing         logical appeal. Yet it's not a system of belief so
band of admirers is right, though, Bayes might        much as a means of measuring belief.
soon gain the kind of international recognition
                                                      Lynch is in no doubt that the clergyman's ap-
that would place his neighbours in the celebrity
                                                      proach to what is known as probability infer-
                                                      ence has at last come of age. 'It has taken 250
Bayes published only two minor works during           years for people to catch up with him,' he says,
his life, one of them anonymously, and neither        'and realise the potential of his ideas.' A 35-year-
was of any lasting consequence. But among the         old billionaire with an expanding company and
notebooks on astronomy and electricity he left in     a receding hairline, Lynch is an enthusiastic ad-
his will, there was a tentative essay entitled        vocate of Bayes. He speaks of the Pauline con-
'Towards Solving a Problem in the Doctrine of         version that Bayesians typically undergo:
Chances', which was published in 1763, two            'People leave the lab one night non-Bayesians
years after his death. The centrepiece of the         and return the next morning convinced Bayesi-
work was an equation designed to predict              ans, and they can't do anything in life — make a
events in conditions of uncertainty. And it is this   cup of tea or pick a girlfriend — without view-
methodology that, according to some informed          ing it from a Bayesian standpoint.'
observers, is now transforming the world of sci-
                                                      Indeed such is the evangelical fervour of born-
                                                      again Bayesians that it seems almost churlish to
'Walk into a research department in Cambridge,        ask what exactly it is that the man came up with.
MIT or Stanford nowadays,' says Mike Lynch,           The answer, as it turns out, is both stunningly
Britain's leading software entrepreneur and a         simple and infinitely complex.
devout Bayesian, 'and you will meet people who
                                                      Not until after 1958 did an entry appear for
will tell you that Bayes is more important than
                                                      Bayes in the Encyclopaedia Britannica, and even
Marx and Einstein put together.'
                                                      then its details were enigmatically sparse. Al-
                                                      most nothing is known for certain about the
man whose speciality was uncertainty; even his             fashion, Bayes's rule was soon viewed by statis-
birth year is subject to debate (it could be 1701 or       ticians as a flawed concept. Only as recently as
1702).                                                     the 1950s was it re-evaluated.
The biographical bones amount to little more               'Everybody is a Bayesian unless they've been
than that he was the son of a nonconformist                trained otherwise,' argues Professor Anthony
minister from London who himself became a                  O'Hagan of Sheffield University's department of
Presbyterian clergyman in Tunbridge Wells,                 probability and statistics, and a leading figure in
Kent. A fellow of the Royal Society, he was a              Bayesian circles. But, as it happened, most statis-
keen Newtonian and he may have attended Ed-                ticians were trained otherwise. The orthodox
inburgh University. And that's about the sum of            school was only interested in how often things
it. Were it not for his friend, Richard Price, an-         happen. Therefore no other criteria but an
other nonconformist minister and amateur stat-             event's frequency were considered when esti-
istician, Bayes would have been lost to history.           mating the likelihood of its happening again.
Price, who inherited Bayes's notebooks along               Gradually     Bayesians,   with     their  all-
with £100, was a natural publicist and spin doc-           encompassing, if 'subjective', outlook on life,
tor. A staunch supporter of the American and               have established themselves as the orthodoxy,
French revolutions, he recognised the earth-               although the more rigid, but 'objective', non-
shaking quality of Bayes's musings. And he did             Bayesians continue to fight a bitter rearguard
not play it down. Replacing Bayes's deprecating            battle.
introduction with his own brand of hyperbole,
Price claimed that the purpose of the essay was            In itself that debate, while engrossing enough
'to confirm the argument... for the existence of           for statisticians, is a trifle arcane for most of us.
the Deity'.                                                The greatest impact of Bayesian thought is on
                                                           computer engineering, a revolution that may
In fact, Bayes had done no such thing. If he was           lead to computers being able to act like us: intui-
interested in the unknown, it was in predicting,           tively. And the Bayesian who has had the great-
not confirming, it. Specifically, Bayes had come           est impact on computer engineering is Mike
up with a means of calculating the probability of          Lynch.
a future event occurring based on previous
events, current conditions and all other known             The progress of Lynch from a graduate student
and related factors. In other words, he had de-            who couldn't get a bank loan to a billionaire
vised a formula for that most unconscious and              businessman, inside a decade, is one that even
subtle of skills: human intuition.                         the most cavalier Bayesian would not have
                                                           dared to predict. 'No,' agrees professor Peter
Every day, whether driving a car or deciding               Rayner, of Cambridge University's engineering
what clothes to wear, we analyse the probability           department. 'Not enough data.'
of events — is it going to rain? Will a truck come
round that corner? — by drawing on sensory                 Lynch worked on post-graduate research into
apprehension and a background of stored in-                probability theory and Rayner was his adviser
formation in such an unthinking manner as to               and Bayesian mentor. 'There were two things I
seem almost instinctive. And as we learn each              noticed about Mike. First, he didn't often read
new factor — the distant appearance of a grey              his lecture notes but he invariably came up with
cloud, the sound of a horn — we are instantly              novel solutions, even if they were not always
able to change our estimation of the likelihood            right. He was clearly a very creative student.
of that rain or truck arriving.                            The other thing was that it was obvious he was
                                                           going to make an awful lot of money.'
In turning that sophisticated process into a basic
equation, Bayes, no doubt unknowingly, upend-              The son of an Essex fireman, Lynch won a
ed science. Traditionally, scientific method is            scholarship to Bancroft's School before going to
based on making the imperfect real world fit in-           Cambridge where he took a degree in infor-
to a perfect model. The beauty of Bayes's rule is          mation engineering. 'Mike had a very sharp
that it accommodates and allows for the myriad             brain,' recalls Rayner. 'But in pure academic
anomalies of real life. Nevertheless, after a brief        terms we've had a number of sharper students.
                                                           He had confidence.'

In the offices of Autonomy, Lynch's internet             knowns (what Bayesians call 'the prior'), listened
software company in Cambridge, I waited in a             to what Lynch had to say ('the evidence') and
large and overheated boardroom to meet this              then took a subjective and intuitively calculated
brilliant, self-confident billionaire. After a few       risk ('the model').
minutes a man in a plaid jacket and jeans                'I think he liked the enthusiasm of youth be-
walked in and started tapping the thermostat. I          cause he couldn't have had a clue what I was
was relieved to see a maintenance engineer deal-         talking about, which was non-linear adapted
ing with the room's uncomfortable stuffiness.            pattern recognition. These days,' Lynch added,
'Hello,' he said. 'Bloody thing's broken.'               'I've learnt the concept of marketing.'
I smiled and then realised that he wasn't the            He set up a company with the cyberpunk name
company spark, but the bright spark who                  of Neurodynamics and, like some sinister, futur-
owned the company. At first sight, Mike Lynch            istic brain from a William Gibson novel, devel-
is an unassuming, indeterminate presence: a              oped a system for matching fingerprints for the
round face with an acute expression, a goofy             Essex police force and another for reading car
smile and quizzical eyebrows, young without              numberplates. The pattern-recognition solutions
looking youthful, and informal but not quite at          Lynch came up with were the result of complex
ease. But when he talks he does so with a com-           Bayesian methods. In 1996, using the same prin-
manding wit and intelligence, so that you want           ciples, he set up Autonomy to exploit the great
to listen even if you don't quite understand.            problem of the early computer era: unstructured
Perhaps it was this conversational talent that in-       data. Or, as it's more commonly known, prose.
spired the manager of a rock band to give him            You only have to use a search engine to realise
£2,000 in a Soho wine bar one night back in 1991.        how cumbersome computers are when it comes
Lynch was completing his PhD and, in his                 to deciphering written text. Type in the word
words, 'Bayesianism was just beginning to                'octopus' and you'll be directed to a whole range
sweep like wildfire.' He wanted to set up a com-         of webpages from Octopus book publishers to a
pany and had enjoyed little success with more            tribute to the Beatles' Octopus's Garden, but
conventional routes for raising finance. 'The            none of these will help you if you are interested
kind of venture capitalists you met in those days        in cephalopods. The problem is that computers
were the idiot sons of idiot stockbrokers,' he re-       can recognise individual symbols but are unable
called. 'Complete waste of time. I went to the           to identify the context of symbols. Autonomy
bank and met an incredibly nice chap who lis-            software enables a computer to analyse the pat-
tened to me and admitted at the end that he              terns of a document, regardless of its language,
normally did loans for people buying newsagent           and then draw conclusions as to its relevance or
shops. We had an interesting discussion about            importance.
confectionery. People always want sweets.'               'We don't need to know the meaning of the link
Then there was an equally forlorn visit to the           between words, we just need the probability dis-
DTI. 'We had tea served in a full china tea ser-         tribution,' Lynch explained. 'So if you see the
vice. We had a long chat and he showed me                word "cat", it's probably about felines. If you see
some of the things that the DTI were involved            the words "cat" and "nine", it's probably about
in. Classic British mistake: they loved anything         felines. Nine lives. If you see the words "tail"
quirky. So it was the sub-bicycle that could be          and "cats", it's probably about felines. Cats have
ridden across lakes.'                                    tails. If we see "cat", "nine" and "tails", it's proba-
In a sense, the bankers and civil servants were          bly about pirates. You start to build up these
non-Bayesians. They simply looked at the fre-            strange combinatory effects.
quency with which specialist software compa-             'The other important thing is conditional proba-
nies had succeeded and made what they                    bility, which is the ability of one event to change
thought was an objective decision against taking         the chances of another one. If we see the word
the risk. The inebriated rock manager, whom              "cat", it's about felines. If we see the word "bur-
Lynch refuses to name, was a natural Bayesian.           glar", it's about your house being broken into. If
He looked at his own record of backing un-               we see "cat burglar", the probability is not half

and half. The meaning of cat has been modified             I asked how much he owned. 'Er,' he said, '19
and we can express that through probability dis-           per cent. But I still eat at McDonald's.' What he
tribution. Therefore you can develop probability           didn't mention was that later on that day he
calculations, which is what Bayesian inference is          would sell one per cent of shares, leaving him
about. The beauty of it is that the belief, the idea       with 18 per cent of the company and £47.4 mil-
of what it's about, is independent of words.'              lion in the bank. As it turned out it was a well-
At that point he caught sight of my dazed ex-              timed measure. The worldwide collapse in high-
pression; a look not seen by anyone, I suspect,            tech shares led to Autonomy losing more than
since my A-level maths teacher attempted to                half of its share value between October last year
share with me his passion for calculus. 'That's as         and January this year. Lynch felt moved to make
clear as mud,' said Lynch.                                 a statement early in the new year to reassure
                                                           shareholders that the future was still bright.
Now Europe's fastest-growing internet software
company, Autonomy is a global concern with                 When not commuting to San Francisco, he lives
offices across America and Europe. Its clients in-         alone — apart from his dog, Gromit — in a tiny
clude the US Department of Defence, a number               Suffolk village with a Viking tower. There is no
of media conglomerates, General Motors and the             room or time yet for a steady girlfriend. But he
BBC. But in truth, Autonomy has not yet                    savours the slow pace of the English country-
achieved anything really sexy, unless you're ex-           side. 'In London or San Francisco you go into a
cited by the idea of a computer automatically              restaurant and overhear people talking about C-
replying to an e-mail. But it is a giant leap on the       programming or Java. In Suffolk conversation is
way to computers making decisions. And Mi-                 about how to catch moles. It's the great unsolved
crosoft is just one of a number of companies that          problem, more so than Fermat's last theorem.'
are following Autonomy's lead.                             He also keeps koi carp in a pond and maintains
Lynch still finds time to get along to gatherings          a small-scale railway in the garden. 'Pretty much
of fellow Bayesians to discuss probability theo-           the exact replica of a steam engine. I went to an
ry: 'It's like going to a cheese-and-wine do, but          agricultural auction, a real string-round-the-
there are blackboards around the room. You                 trouser-leg job. I managed to buy one for next to
have to be careful because you can find yourself           nothing.'
wiping someone's equation off the board with               That £100 and the notebooks that Thomas Bayes
the back of your jacket.'                                  left Richard Price turned out to be some inher-
                                                           itance. What is the probability distribution that
It's a typical Lynch comment, comic and demys-
                                                           in the future coachloads of superwealthy geeks
tifying. Like Bayes himself, who argued that
                                                           from Silicon Valley will be paying homage in a
maths was a matter of amusement rather than
                                                           cemetery just near the roundabout at Old Street?
high seriousness, Lynch finds much of what
passes for rational thought utterly absurd. 'One
of the things I find hilarious is modern man-
agement textbooks. The idea that you've got this
human brain that understands all the subtleties
of employees, markets, what's going on — gen-
eral feeling — and instead you're supposed to
make decisions by looking at some stupid dia-
Certainly few management consultants could
quibble with his results. A few weeks after I met
him, I called him up to ask, among other things,
if it was true that he was a billionaire. He told
me people used the term because they divided
the market price of his company by the percent-
age of his shares in it. Autonomy was valued at
£4.7 billion.


Shared By: