2019-09-01_Computer_Shopper

ISSUE 379|COMPUTERSHOPPER|SEPTEMBER (^2019119)
is able to perform dozens of different
functions,dependingon the software,
aneuron cancarry out just one type
of operation.Aneuron has several
inputsand asingle output.Put simply,
it multiplieseach of its inputsby a
weighting factor,adds themtogether,
and usesthisweightedsumto generate
asingle output,dependenton
whether the sum exceeds athreshold.
We canimaginethat acircuit to
carry out the simplefunctionof a
neuron needonly haveahandfulof
transistors, perhapsup to several
thousandif thereare lots of inputs.
The processorin adesktop PC has as
manyashalf abillion transistors per
core whenaveragedacross the chip.
Given this, it’seasytoappreciatewhy
there’s such adifferenceinarithmetic
performance, even if pattern
recognitiontells adifferentstory.
Butwe’re notcomparinglikewith like.
Amicroprocessorcould have
somewherebetween acoupleof
coresand 64, and the rate at which
datacan flowbetween themis limited
because it has to use ashared
communicationchannel.The human
brain is very different. Neurons may
be simple,but thereare vast numbers
of them–86billion, according to
someof the mostrecent research.
What’smore, theydon’t communicate
via somesort of shared databus, but
by point-to-pointlinks, and herethe
numbersarestaggering.Each neuron
canconnectto as manyas10,000
other neurons, and thereare an
estimated 1,000trillionof these
so-called synaptic connections
throughoutthe brain.
This is surely the ultimatein
parallelism,and it mightbehelpful
to draw somecomparisonswith
parallelismin conventional
microprocessors.The first
microprocessorhad asingle core.As
the limitsof whatcould be achieved
this waywerereached, thesewere
supplantedbymulticorechips,
containingtwotoeightcores formost
desktop designs,with each core being
hugelypowerful but with limited
communicationbetween them.
In parallel with this, graphics
processingunits (GPUs) were
designed,initiallyforhigh-performance
graphics,but morerecently they’ve
beenused foravariety of demanding
applications.Theyhavemanymore
cores–over5,000 in sometop-end
chips–but, sincethey’redesignedto
streamlineonly certain specifictasks,
thosecoresare muchsimplerthan
thosein standardmicroprocessors,
perhapscontainingjust afew million
transistors instead of 500 million.
However,because thosecorescan
work pretty muchin isolationfrom
the other cores, inter-core
communicationperformanceisstill
very modest.We can, in someways,
viewthe humanbrain as the result of
taking this trend to the extreme.
ANNBasics
The foundationsof the neural
network approach were laid at about
the sametime as thoseof the digital
computer,eventhoughthe latter
has beenthe dominantforce in
computingpretty muchever since. A
blow-by-blowaccountof80years of
developmentwould be tediousbut, to
setthe scene forthe state-of-the-art,
let’slook at the basicprinciples.
We’vealreadyreferred to the
underlyingprincipleof abiological
neuron as producingan outputthat
dependson the sum of its weighted
inputs.Artificialneurons do muchthe
sameelectronically,and areusually
arrangedin layers, includingat least
an inputlayer and an outputlayer.
THEAFFORDABLE
FACEOFANNs
The SpiNNakersupercomputer at the Universityof
Manchesterwouldsetyou back afair bit, but there
is an affordable faceto neuralcomputing. For less
than£100,you could avail yourselfof Intel’s Neural
Compute Stick 2(NCS2), aneuralcoprocessorthat
plugsinto aUSB slotonany PC.
Intel says it has beendesignedto buildsmarter
AI algorithmsto allowdeep neuralnetworktesting
and tuningin applications such as computer vision,
therebyhelpingdevelopersmove from prototype to
production. The companysaysthe NCS2offers
benefits in the developmentofproducts including
smartcameras,drones,industrialrobots and smarthome devices. However, it’s surelyalso a
goodwaytolearnabout this up-and-comingtechnology.
Accordingto the firm, “The Intel NCS2ispoweredby the latest generation of Intel VPU–the
Intel MovidiusMyriadXVPU. This is the first to feature aneuralcompute engine:adedicated
hardwareneuralnetworkinferenceaccelerator deliveringadditional performance. Combined
with the Intel Distribution of the OpenVINOtoolkitsupporting morenetworks,the Intel NCS2
offersdevelopersgreater prototyping flexibility.”
If you’renotsure howaVPU compares with aCPU or aGPU, it standsforvision processing
unit, and it’s aspecific type of AI accelerator designedformachine vision applications.
BELOW:Microsoft’s
workonANNshas
resultedin theapp
SeeingAI,which
assiststhevisually
impairedby
describingscenes,
readingtextand
muchmorebesides
ABOVE:AddaIntel NeuralComputeStickto
yourPCtojointheneuralrevolution