http://www.techradar.com/pro/linux May 2019 LXF249 43
Omoju Miller IntervIew
lXF:Yourtalkhadaninterestingtitle:
“Anewvisionfortheglobalbrain:
Deeplearningwithpeopleinsteadof
machines”.Canyougiveourreadersa
summaryofit?
om:Itrainedasaclassicalcomputer
scientistwhothendidsomeworkon
IntelligentSystems,specificallyon
KnowledgeRepresentation.That’s
knowledgerepresentationandreasoning
ofthesemanticweb.Fromdoingthat
workIwantedtounderstandfurtherhow
humansacquiredcomputation.
Doingworkinthosetwoareashelped
meunderstandthatwecouldactually
abstractawayreasoningfromthe
computationalsubstrate.Itcouldbea
humanbeingcarryingoutthecomputation
oritcouldbesomeothercomputational
device,likeamachine.Ifthat’sthecase
thenweknowwecandothis,andwe’ve
seenexamplesofusdoingthisand
understandingtheframeworkandthe
architectureofadeeplearningsystem.
Werealisedthereisaclosefacsimile
tohowmultinationalcorporationsrun
themselves.Theyhavethishierarchy
withtheCEOontop,allthewaydownto
employees.Andallthesecompanieskind
ofrunthemselvesonObjectivesandKey
Results(OKRs).Sotheyhavesomesortof
goalsthattheywanttoachieve,they’veset
itout,they’vefiguredouthowtomeasure
it,andeachpersondoessomething
towardsmakingsurethatthatgoaloccurs.
Andtheyiterateoverandoveragain.
Youcantakethatsameideaand
extrapolateittoasocialnetwork.It’sthe
samething.Inthesocialnetworkcontext,
werealisedthatyoucanactuallyinfluence
humanbehaviourbylevellingthosesame
constructs.There’ssomethingyouwant
todo,youcanrunexperimentsandthere’ll
besomeoutcomesthatthehumanswill
carryout.Unwittinglymostofthehumans
onthesocialnetworkdidn’trealisethiswashappening.Butnowthatweknow
thatthat’sthecase,insteadofthinking
aboutitassomeevilmonster,something
hideous,whatifweactuallyboughtintoit
andthenco-optedthatsamenetworkand
architecturetosolvedifferentkinds
ofproblems?
Whatifwegeneratedthatsame
architectureanddecidedwewanttouse
ittosolvecertainsocietalproblems?
Maybewhatyouwanttodoiscreate
moreempathy,maybeyouwanttodo
somethingtodosomethingtowards
reducingtheimpactofpoverty,maybe
youwanttodosomethingtowardsclimate
change.Wethenfigureoutwhatistheloss
function,andhowwecanmeasureifwe’re
gettingclosetoachievingourgoal.
Ifwecanfigureouthowtomeasureit
thenwecanback-propagatecertainkinds
ofactionsacrossthisdistributedplatform
andgetusthere.Basicallyit’sanewvision
fortheglobalbrain–deeplearningwith
humansinsteadofmachines.Alsowewant
tonotnecessarilybeafraidoftheglobal
brain–itcanbeusedtodosomethingfar
morepowerfulthatwillnotbedeleterious
tosociety.Weshouldtakeitanduseitto
ourownends.lXF:Thisisfascinating.Iwrotean
introductorymachinelearningfeature
[seelXF236]soit’sinterestingtohear
howfarthesethingscanbeabstracted
beyondrecognition.Ihearyou’rean
advisertotheWorldEconomicForum
onthissubjecttoo?
om: Not officially an adviser, but I’m part
of its AI and Robotics Expert Network
[see http://bit.ly/lxf249ai 1 ]. The idea
is they ask really big questions and want
this group of experts to chime in. I get their
emails, I read them – sometimes I chime
in, sometimes I don’t. But if one has time
it’s a very good place to speak with one’s
peers,totalkonideas,andtoactuallymoju Miller is a Machine
Learning Engineer at GitHub.
She holds a PhD from the
University of California,
Berkeley(which, apart from its academic
excellence, she chose for its proximity
to Skywalker Ranch). While studying for
this part-time, as well being a mother,
she worked as an expert at Google for
its non-profit fund for Computer Science
Education. We were lucky enough to
catch up with her at O’Reilly’s Velocity
conferenceinLondon,October2018.
linuxFormat:Tellusaboutyourdata
scienceworkatGitHub.
omojumiller: At GitHub I am part of the
Machine Learning team, which is a part of
the Platform Engineering team. I work on
building deep learning models. Let me dial
that back, actually...
My job is to use data to supercharge
our services. So we have all these different
data sources and we’re trying to learn
which approaches will be useful to engage
in certain kinds of products. So for
example, the last project I was working on,
and probably am still working on, is about
understanding the fundamental nature of
code and how to represent that.
It’s a bit of applied research, because if
we can understand how to represent code
in some sort of vectorised space, then we
can answer questions like “Is this piece of
code similar to that piece of code?” and we
can do things like code provenance: “Who
first wrote this function?”. All these things
could come from this understanding, so
that’s what I’m working towards.
I work on other things that are more
straightforward – there’s an immediate
application to a product. Some things
are more “We have to do this because
it’s going to help us on our roadmap”,
and if we eventually get there, say in 18
months, then we can really supercharge
our products. So what I do is really Applied
Mathematics plus common sense and
usefulness in the real world.
The master and
o the Padawan.1) https://www.weforum.org/agenda/archive/artificial-intelligence-and-robotics