Scientific American - November 2018

ADVANCES

14 Scientific American, November 2018 Illustrations by Thomas Fuchs

GETTY IMAGES

PSYCHOLOGY

Flooding

the Senses

Visualizing climate catastrophes

may spur people to act

Many people view climate change as a
distant, abstract threat. But having them
imagine the tangible consequences of
ßxäø§îž³lß ̧øšîä ̧ß‹ ̧ ̧lä­Dāšx§Ç䚞…î
this perception and encourage proenviron-
mental behavior, a new study suggests.

Researchers asked 93 college students

in Taiwan to read a report on temperature

D³ ̧­D§žxäj‹ ̧ ̧läD³l ̧îšxß`§ž­Dîx`šD³x

ßx§Dîxlxþx³îäîšDîšDþxD†x`îxlîšxžä§D³lÍ

The scientists then asked 62 of the partici-

pants to write down three ways in which

such phenomena might impact their future

lives. Half the people in that group were

instructed to imagine such scenarios in

lxîDž§jž³`§ølž³äÇx`ž‰`ž³lžþžløD§äD³läxî-

tings. The remaining 31 students did not

complete either the writing or imagining

steps, acting as a control group.

All the participants then rated their

perceptions of climate change risks by

responding to prompts such as “How likely

do you think it is that climate change is

having serious impacts on the world?”

They used a scale from 1 (“very unlikely”)

to 7 (“very likely”). The average score was

higher among subjects who had been

asked to envision detailed scenarios than

among those who had not. The results

ÿxßx§Dîxß` ̧³‰ß­xlž³Däx` ̧³lxĀÇxߞ-

ment involving 102 participants.

³lžþžløD§äž³îšx‰ßäîxĀÇxߞ­x³îÿš ̧

šDlþžäøD§žąxlîšxx†x`îä ̧…`§ž­Dîx

change were subsequently more likely to

say they would use air conditioning in an

energy-saving manner. In the second

TECH

Lifelong

Learning

ßxDDßžD§ž³îx§§žx³x

that continues to adapt

What if you stopped learning after grad-
uation? It sounds stultifying, but that is
how most machine-learning systems are
trained. They master a task once and then
are deployed. But some computer scien-
îžäîäDßx³ ̧ÿlxþx§ ̧Ǟ³Dß`žD§ž³îx§§ž-
gence that learns and adapts continuously,
much like the human brain.
Machine-learning algorithms often take
the form of a neural network, a large set of
simple computing elements, or neurons, that
communicate via connections between
them that vary in strength, or “weight.” Con-

sider an algorithm

designed to recognize

images. If it mislabels a

picture during training,

the weights are

adjusted. When mis-

takes are reduced

below a certain thresh-

old, the weights are fro-

zen at set values.

The new technique splits

each weight into two values that

` ̧­Už³xî ̧ž³‹øx³`xš ̧ÿ­ø`š ̧³x³xø-

ß ̧³`D³D`îžþDîxD³ ̧îšxßÍ5šx‰ßäîþD§øxžä

trained and frozen as in traditional sys-

tems. But the second value continually

adjusts in response to surrounding activity

in the network. Critically, the algorithm

also learns how adjustable to make these

weights. So the neural network learns pat-

terns of behavior, as well as how much to

modif y each part of that behavior in

response to new circumstanc-

es. The researchers present-

ed their technique in July

at a conference in Stock-

holm, Sweden.

Applying the tech-

nique, the team created

a network that learned to

reconstruct half-erased

photographs after seeing the

full images only a few times. In

contrast, a traditional neural network

would need to see many more images

before it could reconstruct the original.

The researchers also created a network

that learned to identify handwritten alpha-

bet letters—which are nonuniform, unlike

îāÇxl ̧³xäD…îxßäxxž³ ̧³xxĀD­Ç§xÍ

In another task, neural networks con-

trolled a character moving in a simple maze

î ̧‰³lßxÿDßläÍ…îxß ̧³x­ž§§ž ̧³îߞD§äjD

network with the new semiadjustable

Flooding in the state of Kerala in India

in August 2018. Climate change is mak-

ing extreme weather events more likely.