across disciplinary boundaries, and sometimes it’s worth it.

“But the grander potential for syn-
ergy that’s often spoken of at the level
of federal funding agencies probably
doesn’t happen as much as people
think would be best for science,” Masys
continues. “You can’t push that rope
all that well, because it depends on the
art of the possible with respect to tech-
nologies and the vision of the scientists
doing the work.”
Tom Mitchell, chairman of the ma-
chine learning department at Carn-
egie Mellon University (CMU), concurs
with Masys’ assessment. “I think it
starts from the bottom up and at some
point you’ll see commonalities across
domains,” he says. As an example, he
cites time series algorithms being de-
veloped by CMU colleague Eric Xing
that may also be useful for brain imag-
ing work Mitchell is undertaking.

“There’s an example I think is prob-
ably pretty representative of how it’s
going to go,” Mitchell says. “People en-
counter problems and have to design
algorithms to address them, but time
series analysis is a pretty generic prob-
lem. So I think bottom up it will grow
and then they will start connecting
across [different disciplines].”
Vanderbilt’s Masys is about to be-
gin a collaboration with computation-
al biologists from Oak Ridge National
Laboratory. Masys says the Oak Ridge
scientists’ optimization of Vander-
bilt’s fundamental algorithms and the
lab’s teraflop-capable architecture will
likely speed processing of problems
involving multiplying “several million
genome data points by several thou-
sand people” from five days to three
hours—a prime example of focused
intradisciplinary collaboration and
leading-edge hardware.

new Perspectives on Data

Both Mitchell and Randal Bryant, dean of the school of computer science at CMU, cite the influence of commercial companies for helping to expand the concept of what kind of data, and what kind of data storage and computational architectures, can produce useful scientific results.

“The commercial world, Google and its peers, have been the drivers on the data side, much more than the traditional sciences or universities,” says Bryant, who cites the example of a Google cluster running a billion-word index that outperformed the Big Iron architecture of the “usual suspects” in a 2005 language-translation contest sponsored by the U.S. National Institute of Standards and Technology.

The availability of such large datas-
ets can lead to serendipitous discover-
ies such as one made by Mitchell and his
colleagues, using a trillion-word index
Google had originally provided for ma-
chine translation projects. “We found
we could build a computational model
that predicts the neural activity that will
show up in your brain when you think
about an arbitrary noun,” Mitchell says.
“It starts by using a trillion-word collec-
tion of text provided to us by Google,
and looks up the statistical properties
of that word in the text; that is, if you
give it the word ‘telephone’, it will look
up how often ‘telephone’ occurs with
words from a long list of verbs—for
example, how often does it occur with
‘hug’, or ‘eat’, and so on.

ubiquitous;Computing
Intel’s Friendly, Smart Machines

Context-aware computing, in which devices understand what a user is doing and anticipate his or her needs without being asked, are the next step in the evolution of smart machines, says Justin rattner, Intel vice president and chief technology officer.

In his keynote address at
IDF2010, the recent Intel
Developer Forum in San
Francisco, rattner laid out a
vision in which computers use a
variety of sensors—microphones,
accelerometers, and global
positioning systems (gPSs)—
combined with “soft sensors”
such as calendars and social
networks, to track people’s
activity and figure out how the
devices can help. For instance, a
device might locate someone at
her office, hear the sound of
human voices, crosscheck her
calendar, and conclude she’s in a
business meeting, then suggest
to the husband trying to call her
that this wouldn’t be a good time
to interrupt.

about what’s going on around
it, could learn to recognize
which person is holding it—
based on how the user moves,
what angle he holds it at, and
how fast he presses the
buttons—then make
personalized recommendations
for shows, based on past
preferences. a prototype
Personal Vacation assistant,
developed with Fodor’s Travel,
uses gPS location, time of day,
and past behavior to
recommend restaurants and
tourist sites. Data, collected over
time and shared among devices,
is run through an inference
algorithm that examines the
input and generates confidence
scores to determine what is
likely going on.