Astronomy

WWW.ASTRONOMY.COM 35

Galaxy Zoo, the public is asked to iden-
tify the type of galaxy shown: Is it a disk?
Is it edge-on? Is there a central bulge?
These features can be quickly identified
by eye, but natural variations can make
them exceedingly difficult for computers
to recognize and categorize.
“Humans are actually very well
designed to picking out serendipitous
discoveries in image datasets,” Fortson
says. “By virtue of evolution, humans
have developed this amazing visual cor-
tex that can differentiate the unknown
unknowns from the knowns.”
Of course, using the untrained public
doesn’t come without its challenges.
People make mistakes. Luckily, the large
number of people involved in the identi-
fication can be used to create averages
and a group consensus, which, over the
long run, can be even more accurate
than a single scientist’s identification. In
Galaxy Zoo, 40 different individuals
examine each galaxy to create a trusted
identification. By carefully processing
the results, individual people can even
be weighted differently depending on
their identification success rate. In this
way, people whose identifications gener-
ally don’t agree with the group consen-
sus can be f lagged for rejection, so they
don’t skew the end results.

Rise of the machines
Once the masses have identified and
categorized thousands of images,
significant work remains to analyze
the data. This is where computers finally
come in. These machines are the heavy
lifters, allowing for complex calculations
and comparisons that the
human brain would be hard-
pressed to match on its own.
While machines historically
can only do exactly what they
are told, a subset of comput-
ers are being taught to think
on their own.
Astronomers are using a
type of artificial intelligence,
called machine learning, to
get computers to teach them-
selves how to find patterns in
the data. A specific method of machine
learning known as artificial neural
networks was designed based on how
the brain functions. These neural
networks draw connections in vast
webs of data, just as the human brain
does. To create these networks, a scien-
tist starts by showing the computer a

“training set,” which is a series of exam-

ples containing what the computer is

looking for — such as spiral galaxies.

Over time, and with enough examples,

the computer will become adept at

identifying spiral galaxies, despite

their wide range of appearances. At

this point, the scientist can provide

the computer with a sample of unidenti-

fied galaxies, and the machine will

return those that fit the criteria it

has assessed.

Machines can also be taught a much

more difficult task: assessing how objects

and their characteristics relate to one

another. For example, scientists have used

artificial neural networks to

investigate how galaxies form

clusters and how that group-

ing affects the numbers of

stars the galaxies produce.

Only with the assistance of

computers are the scientists

able to compare the many

physical properties at play,

such as galaxy mass, distance

between galaxies, and previ-

ous interactions between gal-

axies. And by comparing

many hundreds of thousands of galaxies,

scientists are able to make broad conclu-

sions about our universe that are unbi-

ased by small irregularities.

When encoded properly, artificial

neural networks can provide profound

insight to scientists; however, they can

also be easily misused. For example, if the

training set is not extensive enough, the

computer will draw the wrong conclu-

sions. Or, as astronomers are fond of

repeating, “Garbage in, garbage out.”

The other drawback to artificial neu-

ral networks is that they require vast

datasets to “learn” from. Luckily, in the

era of large-scale surveys, vast datasets

are common. This means that artificial

neural networks can quickly turn the

problem of too much data into an advan-

tage. The larger the training set — which

citizen scientists can help bolster — the

better the results.

The future of

unexpected discoveries

“Our ability to collect these humongous

datasets is developing in parallel with

our ability to interpret these huge data-

sets,” says Ivezić. “Both directions are

important — people who collect data and

people who develop tools to analyze and

interpret. Otherwise we’d just be stuck

with a huge pile of zeros and ones we

couldn’t make sense out of.”

With the combination of large-scale

surveys, a legion of citizen scientists, and

new machine learning techniques, it

seems many new unexpected discoveries

will soon emerge from the darkness. But

as for the nature of those discoveries?

Only time can tell.

Mara Johnson-Groh is a science writer and

photographer who writes about everything

under the Sun, and even things beyond it.

Galaxies, clusters

of galaxies, and

clusters of clusters

of galaxies join

with dark matter

to form a grand,

weblike structure

called the cosmic

web, a slice of

which is shown

here. With the

help of artificial

neural networks,

astronomers

hope to run

simulations like

this to investigate

the cosmic web in

much greater detail

than previously

possible. NASA, ESA,

AND E. HALLMAN (UNIVERSITY

OF COLORADO BOULDER)

LSST will

collect as

much as 30

terabytes of

data every

clear night.