You are on page 1of 1

The Last Word

Mind-Reading Computers
By Charles Day

L

ast year, the website of Britain’s Daily Mail newspaper became
the world’s most-visited English-language news source.

Although the Mail’s website owes its popularity to a menu rich
in celebrities, crime, and royals, it offers readers something that
my stuffier hometown newspaper, The Washington Post,
lacks: a top-level section devoted to science.
Granted, the Mail’s science coverage tends toward the
sensational, but it does encompass superluminal neutrinos,
the Higgs boson, and other weighty topics. The story that
led the science section on 1 February 2012 was both sensational and important, as you can tell from the headline:
Mind-boggling! Science creates computer that can decode
your thoughts and put them into words.

The story’s origin lies in an article published in PLoS
Biology by Brian Pasley of the University of California,
Berkeley, and his collaborators.1 Fifteen patients who
suffered either epilepsy or brain cancer agreed to let
Pasley’s team attach an array of electrodes to their brains
while their skulls were opened for surgery. The electrodes recorded signals from neurons located in a part
of the brain, the auditory cortex, that interprets spoken
language.
Before the patients underwent surgery, they listened to
single words and whole sentences. Pasley and his collaborators correlated the electrical recordings with the words’
acoustic spectra. A machine-learning algorithm then derived a mapping that could reproduce an acoustic spectrum
from a neural recording.
Predicting what someone hears based on his or her brain
activity is impressive, but it hardly qualifies as mind reading. However, it turns out that the auditory cortex is also
responsible for encoding speech. When Pasley’s team
asked each patient to think of words without uttering them,
the algorithm accurately predicted what those unspoken
words were. In that sense, the algorithm really did read the
patients’ minds.
Pasley’s algorithm occupies one front in a broad campaign
to understand how the human brain works. On another front,
104

Copublished by the IEEE CS and the AIP

CISE-14-4-Last.indd 104

biophysicists are developing ways to map the topography of
the brain’s interconnected neurons. Given that the human
brain contains on the order of 1011 neurons, each of which
is connected to up to 1,000 other neurons, assembling a
complete neuronal map could turn out to be infeasible—
and perhaps unnecessary.
A detailed map of a single, characteristic neighborhood of the brain might yield enough information to
identify the physical features that underlie thought and
memory. But knowledge of those features alone might
fall short of demonstrating that someone understands
the brain. If that turns out to be the case, then a convincing demonstration might entail building a simulated
brain.
The anatomy and physiology of such a brain wouldn’t
necessarily resemble those of our own. Indeed, the first
prototype could turn out to consist of a building-sized
stack of optical tables where pulsed beams of light—the
information-carrying signals—bounce off mirrors and
pass through prisms. Provided that the simulated brain’s
topology and interconnections are described using the
same mathematical equations that apply to a human brain,
such a demonstration would be valid.
And if that fantasy becomes a reality, simulation would
have attained a new and higher status in science. Rather
than providing a way to calculate a theory’s validation, the
simulation would be the validation.
Reference
1. B.N. Pasley et al., “Reconstructing Speech from Human Audi-

tory Cortex,” PLoS Biology, vol. 10, no. 1, 2012, e10001251;
doi:10.1371/journal.pbio.1001251.
Charles Day is the Web editor at Physics Today. Echo, a seven-monthold Airedale terrier, understands “sit,” “down,” and other commands,
but doesn’t always obey them. 

1521-9615/12/$31.00 © 2012 IEEE

Computing in Science & Engineering

6/12/12 4:24 PM