Online games that tap your brainpower without you noticing can
crack problems that have defeated the most powerful computers,
says Lewis Dartnell
Lewis Dartnell
Get involved with distributed computing with these online games and
downloads
ARISTIDES is a typical 13-year-old boy. He plays basketball after
school, is learning the clarinet, and in the evening sits in front of
his computer playing games. There is one game that he is especially
keen on, however, which marks him out from his peers. Every day he
logs on to www.fold.it, where, under the nickname "Cheese", he
plays a game that involves twisting, pulling and wiggling a 3D
structure that looks a bit like a tree's root system. He manipulates
different lengths of these snaking green tubes until they fit into the
smallest volume possible. It may sound like a rather bizarre game - a
distant 3D relative of Tetris, perhaps - but it is in fact a brilliant
disguise for one of the toughest conundrums facing biologists today:
how do proteins fold?
The structures Aristides plays with are computer representations of
real proteins such as collagen. Without understanding the first thing
about molecular biology, he is contributing to one of the knottiest
problems of modern science. "Predicting exactly how a long protein
chain curls up as compactly as possible, amongst all the myriad
possibilities, is a very hard problem to solve with computers," says
David Baker at the University of Washington in Seattle, who invented
the game.
As the length of a protein chain increases, the number of possible
ways it can be folded increases exponentially. Even for the simplest
chains, it would take a typical desktop computer several centuries to
predict the optimum way a protein would fold. Yet with the help of
60,000 amateur players like Aristides, in the six months it has been
running Foldit has already calculated how scores of proteins would
fold.
Foldit's success is largely down to a clever new twist on an old idea.
The last decade has seen a rise in "distributed computing" projects,
in which massive computational problems are chopped up into smaller
parts and supplied via the internet to millions of desktop computers
around the globe to do the number crunching in parallel. The oldest
and most popular of these ventures was set up as part of the Search
for Extraterrestrial Intelligence project. In s...@home, signals
received by several radio telescopes are analysed by otherwise-idle
computers for signs of a message from aliens. Distributed computing
has since been applied to many other areas, including protein folding,
cracking encryption codes and climate prediction.
Calling all brains
But there are limits to what even a million computers can do. "Despite
computers being very quick and accurate at certain problems, for many
tasks they are still far surpassed by the human brain, such as in
visual processing, spatial reasoning or problem solving," says Aaron
Sloman, who studies artificial intelligence at the University of
Birmingham, UK. So now the idea of distributed computing is being
turned on its head. Instead of harnessing idle machines, researchers
are inventing ways of using the processing power inside the brains of
"idle" computer owners.
There seems to be no shortage of this intellectual power going
begging. Clay Shirky at New York University has calculated that every
weekend in the US alone, 100 million person hours are spent watching
TV adverts - the same amount of time it took to create and edit the
2.5 million encyclopedia entries on Wikipedia. If only a fraction of
this spare brainpower could instead be channelled into simple online
tasks that help science, the contribution would be enormous.
One of the earliest attempts to elicit help from people online was
NASA's Clickworkers website. The project, launched in 2000, asked
volunteers to mark the position and size of craters in photographs of
the Martian surface snapped by the Viking probes in the 1970s. More
recently, volunteers have been classifying different landforms in the
ultra-high-resolution images beamed back by the Mars Reconnaissance
Orbiter. NASA is also hoping to bring this workforce to bear on images
returned by the Dawn spacecraft when it explores Vesta and Ceres, the
two most massive objects in the asteroid belt between Mars and
Jupiter.
There is a significant problem with the Clickworkers approach, however
- it is a solitary experience that simply isn't much fun. It is hard
to see what motivates the home volunteers, beyond a rewarding sense of
contributing to an important cause. This highlights a problem key to
all efforts at human processing: how do you keep the volunteering
brains interested long enough to keep them coming back?
Luis von Ahn, a computer scientist at Carnegie Mellon University in
Pittsburgh, Pennsylvania, says he has the answer. Since 2002, he has
been pushing the use of volunteers, or "human computers", far beyond
the early efforts of Clickworkers. His secret is to transform complex
problems into simple, addictive games.
His first game, called ESP, was designed to create a list of words
people would associate with each image in a database in order to help
train artificial intelligence systems. To do this, the game pairs up
anonymous players and presents them with identical photographs. The
players must try to guess the word their partner would use to describe
the object pictured in the fastest time, gaining points for doing so.
Once agreement is reached, a new image is displayed and the pair
continue racing the clock for maximum points.
The game is now available on the "Games with a Purpose" website
(www.gwap.com), which has so far attracted 120,000 members and had
some significant successes. ESP, for example, has labelled more than
50 million photos, and last year was licensed by Google to help
improve its image search engine. "People love playing this game, and
report feeling a strong bond with the faceless stranger they
collaborate with," von Ahn says.
His latest scheme is the reCAPTCHA project. CAPTCHAs are
computer-generated images in which a word is distorted and presented
against a distracting background to make recognition impossible for a
computer system, but simple for a person. These are now used on sites
such as Facebook and Twitter to confirm a new user is indeed human,
and so prevent hackers from using software bots to set up hundreds of
accounts automatically, to use for generating spam emails or for other
nefarious purposes.
ReCAPTCHA exploits people's ability to identify distorted images. It
uses text from an old manuscript that needs to be converted into
digital form, but which optical character-recognition software has
failed to classify. The unknown word is presented to the user in
addition to a known word, which acts as the real test of humanness. By
correctly deciphering the text of both words, users confirm themselves
to be human and are allowed to sign up to the online account, while at
the same time helping to digitise and preserve an ageing text (see
diagram).
The system has already helped to transcribe more than a billion words,
and is now being used to digitise the 130-year back catalogue of The
New York Times. In the future, von Ahn hopes to use a similar trick to
create transcripts of historical audio recordings.
Using people to process images is not a foolproof method, however.
There have been some unexpected consequences, as the scientists who
run Galaxy Zoo discovered. This project, launched in 2007, has
involved 160,000 volunteers helping to identify about a million
galaxies in images collected by the Sloan Digital Sky Survey.
Volunteers are asked to classify the structure of galaxies as either
spiral or elliptical, and to record the rotation direction of the
spiral galaxies. No expertise is needed, so newcomers can immediately
begin contributing to the project.
One of the things Galaxy Zoo researchers wanted to know was whether
there is any bias in the direction of spin in spiral galaxies. A
previous study had found that the odds of a galaxy rotating clockwise
or anticlockwise depended on where you looked in the sky, suggesting
an unexpected organisation in the universe at large scales. However,
the massive data set produced by Galaxy Zoo's volunteers confirmed
that there is no such bias in the observable universe.
Galaxy Zoo did detect a curious bias in the way that different people
were classifying the same galaxies, though. For some reason, people
turn out to be more likely to classify a spiral galaxy as spinning
anticlockwise than clockwise, although the researchers believe that
this did not taint the results overall; the effect is subtle and only
became apparent because the number of volunteers contributing to
Galaxy Zoo was so large.
Nevertheless, it highlights an interesting concern with human
processing - how can you be sure unconscious preferences or quirks of
human psychology are not colouring your results? The source of the
bias in Galaxy Zoo is unclear. "Personally, I suspect it might simply
be because 'anticlockwise' is the middle button," says Kate Land who
conducted the study into Galaxy Zoo's results.
Galaxy Zoo has produced another surprising result. Last year, Dutch
school teacher Hanny van Arkel spotted a bizarre astronomical entity
in one image. The object looks likes a wisp of bright green smoke,
contains no stars and is unlike any celestial phenomenon ever seen. If
it hadn't been for van Arkel's keen eye, it is unlikely this new
object would have been detected. "The human brain is incredibly good
at identifying the unusual, something that is difficult to program
into the search parameters of an automated visual cataloguing system,"
says Chris Lintott, an astrophysicist at University of Oxford and one
of the Galaxy Zoo team leaders. In recognition of van Arkel's shrewd
observation, it has been named Hanny's Voorwerp (Dutch for "object").
While Galaxy Zoo relies largely on astronomy enthusiasts, and von
Ahn's reCAPTCHAs are used at points where the web user has no choice
but to solve the puzzle, a new generation of projects hopes to entice
the general public to donate their spare time purely for fun. Foldit
is at the forefront of these efforts.
Foldit grew out of a distributed computing project called
rose...@home, also run by Baker. In a similar way to the s...@home
project, rose...@home uses spare computer time on PCs around the world
to search through all the possible configurations to find the most
compact form a protein can fold into. Using the results from the
project, Baker's research team has published a string of papers over
the past year in Nature and Science on protein-structure prediction
and design. Baker and his team found, however, that the raw results
from rose...@home had to be tweaked by hand to be scientifically
useful. He now wants to watch how Foldit users play to improve the
computer programs behind rose...@home.
When it comes to complex problems like protein folding, though, not
everyone is convinced human processing is going to deliver the goods.
Vijay Pande, the lead researcher on another distributed computing
project to simulate protein folding called fold...@home at Stanford
University, is concerned that volunteers will produce results that are
far poorer than those that can be calculated computationally. "Folding
a protein is like learning to play chess well, but with far more
pieces and much more complicated rules, and so an even greater
combinatorial explosion in the number of possibilities that must be
considered," he says. He doesn't believe people will be that good at
taking advantage of all the degrees of freedom. While Pande accepts
that Foldit is a great way to demonstrate the complexity of proteins
to non-scientists, he says he would be very surprised if their results
are as good as those produced by computers.
Unsurprisingly, Baker disagrees. The vast number of different ways
proteins can fold is precisely what flummoxes computers, he says.
Players like Aristides demonstrate how skilful humans can be in
spotting how to tweak a protein to optimise the folding. Within just a
few months of taking up the game, Aristides has already emerged as one
of the top 10 players out of the 60,000 worldwide, and has been
invited onto the top-ranked team in Foldit. He has shown such an
intuitive grasp of the subtle mechanics of the problem that he has
been described by one of the Foldit researchers as a "Foldit savant",
and has recently been flown to Seattle to tell them exactly how he
solves the protein puzzles.
The accuracy of the different human and computational approaches will
be revealed later this year, when the results of a competition called
the Critical Assessment of Techniques for Protein Structure Prediction
(CASP) are announced. The competition compares
experimentally-determined protein structures with predicted
configurations of proteins submitted by different teams using either
computers or people. Initial analysis of results from Foldit are very
promising. "We've already seen solutions from these volunteers that
are very competitive to massive computational efforts, and at times
even outperform them," says Zoran Popovi'c at the University of
Washington, a principal investigator on Foldit. "I expect people to do
much better once we refine the tools to better fit the gameplay and
human puzzle-solving process."
Starting this month, Foldit will be expanded with a new game that
essentially challenges players to design new proteins from scratch.
Players will be able to change side chains, effectively creating
synthetic proteins, and the best ones will be synthesised and tested,
Popovi'c says. And in December, puzzles for growing and moulding
synthetic proteins that bind specifically to a virus will go live.
This could ultimately help produce new drugs and vaccines. "It is here
that I think people will really have an advantage over computers,"
says Baker.
Just as the number of distributed computing applications grew with the
popularity of s...@home, so it is likely that other areas of science
will follow the example set by Foldit. Shirky goes even further: "In
the future there will be so many human processing initiatives that
we'll all feel fierce competition for our idle brain cycles," he says.
Whoever said technology was supposed to give us more leisure time
could not have been more wrong.
Get involved with distributed computing with these online games and
downloads
Be a freelance scientist
Forget Tetris or Solitaire, join one of these science projects and put
your free time to good use
GAMES WITH A PURPOSE (GWAP)
Help train computer vision and artificial intelligence systems.
www.gwap.com
GOOGLE IMAGE LABELER
A spin-off from GWAP, play with an anonymous partner to label objects
as quickly as possible and help improve this image search engine.
www.images.google.com/imagelabeler
FOLDIT
Puzzle your way through folding or designing protein structures and
help biologists design the next generation of drugs. www.fold.it
GALAXY ZOO
Classify images of a million different galaxies to help answer some of
the biggest questions in cosmology. www.galaxyzoo.org
To unsubscribe send a message to [email protected] with
the subject unsubscribe.
To change your subscription to digest mode or make any other changes, please
visit the list home page at
http://accessindia.org.in/mailman/listinfo/accessindia_accessindia.org.in
