HWM Singapore – June 2019

needto storedatain the cloud.Thisis
differentfromsimplyhavinga local
modelon yourphone,and it goesa
step furtherbecausethe trainingis
happeningon yourdeviceas well.
Google’sGboardkeyboardcan
learna new wordlike “Targaryen”
or “BTS”,withoutit knowingwhat
you’retypingat all. Becauseso many
peoplehavetypedthe word,the
sharedpredictionmodelhas been
updatedto reect that.Yourdevice
basicallydownloadsthe current
model,lets it learnfromdataon your
phone,and thensummarizesthe
changesas a small,focusedupdate.
It is this update,and not yourdata,
that is thenencryptedand sentto the
cloud,whereit is averagedwithother
userupdatesin orderto improvethe
sharedmodel.Thismeansthat there
are no identifyingtags or individual
updatesin the cloud,and all the
trainingdatastayson yourphone.

pluggedin, and on a free wireless

connection.But with the right

implementation, federated learning

can be immensely useful, enabling

smartermodels, lower latency, and

lowerpower consumption, while at

the sametime ensuring privacy.

Thatsaid,therehavebeenbig

hurdlesto overcome.For example,

trainingcan’thappenall the time,

and on devicetrainingrequiresa mini

versionof TensorFlow,in additionto

carefulschedulingto ensuretraining

onlyhappenswhenthe deviceis idle,

Trainingonlyhappenswhenthedeviceis

idle,soperformanceisn’taffected.

A distributed

machine learning

approach using

decentralized

data residing on

end devices.

PERSONALIZATION SLEEPING

FEDERATED LEARNING

JUNE 2019 | HWM 81