New approaches to machine learning have moved away from
trying to program complex tasks that humans take for granted,
opening up the field to a huge range of applications, according to
Christopher Bishop, assistant director of Microsoft Research
Cambridge.
Presenting the 2005 BCS Lovelace Lecture, Bishop said,
"Computers are great at doing many big calculations a second, but
getting them to read handwriting, for example, is a big challenge.
It is impossible to write an algorithm for recognising handwriting;
many have tried but all have failed. There are too many differences
even in the writing of one person to be able to write a set of
rules or algorithm.
"This is one of a whole range of problems in pattern
recognition, which people are very good at, but which has been very
difficult to get computers to solve. This is one aim of machine
learning.
"With handwriting recognition, for example, we do not try to
program a solution; we give the computer lots of examples of
handwriting, along with the text, and the machine learns to match
the electronic image to the text.
"A key issue in machine learning is uncertainty. We are dealing
with situations which are very variable or where there is a lot of
noise in the signal. We are used to dealing with uncertainty in
everyday life and we have to deal with it in computing too."
Recent approaches to machine learning focused on three key
developments, Bishop said.
The first is the development of work of an 18th century
clergyman and mathematician, Thomas Bayes, who published a theory
that led to a mathematical method of calculating the probability of
future occurrences of an event from past occurrence data.
The second development has been the appearance of graphical
tools that simplify and speed up work on probability modelling.
Finally, dealing with uncertainty has been eased by new
techniques for making approximations as accurate as possible.
Bishop gave some examples of the new approaches, demonstrating
an image editing tool being developed by Microsoft, called
Grabcuts, which accurately cut a flowerhead from a photo and pasted
it into a new file.
The user highlights the object by roughly tracing a box around
it. The system partly uses the colours of the highlighted object to
separate it from the background, even working in individual pixels
at the edge of the main object to cut out the main colour from the
background. This gives a much sharper result than could be achieved
by a human tracing the outline - and in seconds rather than many
minutes.
The Forza car racing game, recently released for the Microsoft
XBox games console, was a totally different example of the use of
machine learning presented by Bishop.
"In previous games the artificial intelligence that drives the
other cars is handcrafted scripts," he said. "People test the game
and then tweak the script and try again. This is very time
consuming. With Forza, instead of handscripting the cars we get
people to use the game and drive them - and capture the data and
try to mimic their style. If you buy Forza you can train the
computer-controlled drivers in your driving style."
Bishop's other examples ranged from image retrieval to tracing
HIV as it mutates in the body, with the long-term aim of developing
a vaccine.
"The range of applications of machine learning has mushroomed in
the last 10 years, from a handful of niche applications to
virtually every area," he said.