ISSUE 379|COMPUTERSHOPPER|SEPTEMBER (^2019121)
memoryand process it seriallyon
individualcores. The balanceof
memoryaccess and computationon
SpiNNaker wascarefully designedto
ensurethat all of the coresare kept
busyand that all of the memory
bandwidthto the synaptic
informationis kept fully utilised.
We try nottowasteany available
resource,evenwhen the synaptic
connectivityis sparse.”
Researchparty
So what’sbeen achievedwith
SpiNNaker,and wheremightit
lead in the near future?
According to Davidson,“Wesee
SpiNNaker as apurely research tool
forneuroscientists and roboticists.
Thereare also afew startupsusing
SpiNNaker as an experimental
platform to implementtheir own
algorithms.We don’t expect this
work to lead to the sale of millionsof
SpiNNaker boards, morelikely to the
designof dedicated hardwarethat
hardwiretheir algorithmsafter they
havebeen honedusingSpiNNaker.
“If this happenswe will still be
happy, but thereisthe possibility
that the chip itself may haveviable
standaloneapplications,such as in
drones,self-drivingcars or speech-
recognitiontasks.”
It’sreally hardtohear astatement
likethat withoutthinkingof the future
of neural computingmoregenerally.
Butaccording to Davidson,the
technologyisalreadyaffectingthe
hereand now.
“If we considerthe deepnetworks
that arealreadybeingused by
Googleand others to perform image
classificationand other web-based
services, then one could argue that
the futureisalreadyhere. Consumers
use themtoday,but don’t really
understand whatis goingon under
the hoodbecause theyexist only in
the cloud,”henotes.
“I think whatyouare really alluding
to aredevices basedon ANNsthat
interact directly with peopleand offer
fundamentally newservices. These
would includeself-drivingcars,
autonomousdrones,intelligent
surveillance, and even humanoid
robots to act as homehelps.Ithink
someof thesethingsareimminent,
but others arealong wayoff.”
Butwhile consideringdrones and
surveillancesystems as “relatively
low-hangingfruit”,Davidsonbelieves
that robots in the homearestill at
least 30 years away,and the
much-vauntedself-drivingcars are
equallyfarintothe future.
“The mostcontentious use
of ANNswill be in autonomous
vehicles,whicharegetting alot
of attention and fundingat the
moment. This is againa
controversial view, but Idon’t
expect the currentwaveof
investmentintoautonomous
cars to lead to hordes of
self-drivingvehiclesin our towns
anytime soon,”hesays.
“It mightlead to safety
systems in our cars, applying
brakes and lookingout forhazards,
but we won’tbeletting themtake
overthe driving.Being able to
interpret afast-changingscene with
multiplemoving objects,as we
experiencewhen driving,is avery
complextask. We dedicatemorethan
aquarter of our brain to visual
processing,so that’sover20billion
neurons. We arealong wayfrom
buildingsuch asystem, and Iexpect
the currentattempts to develop
cut-down visualsystems that are
goodenoughto perform automated
drivingwill simplynotbereliable
enoughto meetthe rigorous safety
levels neededbefore theyare let
looseon our roads.”
If youenjoytakingthe wheelof
your prideand joyonthe openroad,
then,it soundsas if you’renot going
to be consignedto the back seat any
time soon.In so manyother areas,
however,the eraofbrain-inspired
computinglooksset to revolutionise
21st-century lifesoonerthan you
mighthavethought.
ABOVE:This
over-simplified
representationof
anANNshows
howneuronsare
arrangedin layers,
fromtheinputsto
theoutputs
INPUTLAYER
INPUTS OUTPUTS
HIDDENLAYERSOUTPUTLAYER
c. jardin
(C. Jardin)
#1