1 / 3100%
Weh haveh spenth someh timeh inh theh discussionsh learningh abouth Bigh Datah
platformsh andh tools,h soh Ih thoughth Ih wouldh switchh gearsh andh starth ah discussionh
abouth operationalizingh AI/MLh andh Bigh Datah solutions.h Ash ah formerh cloudh
operationsh engineerh andh administrator,h Ih canh appreciateh theh levelh ofh efforth
andh challengesh aroundh operationalizingh anyh concept.
Oneh shouldh haveh manyh questionsh toh answerh beforeh movingh forwardh withh anyh
planh toh produceh ah concept.h Ourh agencyh likesh toh deployh ah potentialh solutionh ash
ah Proofh ofh Concepth (POC)h toh understandh ah potentialh solution'sh movingh parts.h
Ourh agencyh likesh toh knowh whath worksh well,h whath doesh not.,h andh whath ish
involvedh inh developing,h deploying,h andh supportingh theh concept.h Weh likeh toh
knowh howh receptiveh theh targetedh userh communityh willh be,h howh muchh theh
solutionh willh cost,h andh whath ourh ITh Serviceh Deskh needsh toh knowh toh supporth
theh solution.h Theseh areh buth ah fewh questionsh thath oneh probablyh needsh toh
answerh beforeh operationalizingh AI/ML.
First,h ash weh allh knowh fromh ourh studiesh inh theh MSDAh programh hereh ath UMGC,h
datah ish theh fuelh thath powersh AI/MLh models,h muchh likeh gasolineh ish theh fuelh
thath powersh ourh automobiles.h However,h datah qualityh ish alwaysh ah concern.h Ifh
oneh pumpsh lowh octane,h ethanol-richh gasolineh intoh theh tank,h oneh mayh noth geth
theh optimalh powerh oneh needsh toh geth fromh originh toh destination.h Theh sameh
goesh forh poorh qualityh inh data:h garbageh inh equalsh garbageh out.h Ash ah buddingh
datah analyst,h oneh canh appreciateh howh Mr.h Johnh Parkinsonh putsh ah twisth onh theh
wholeh "garbageh in/garbageh out"h idiomh inh hish articleh onh managingh Bigh Datah –h
"youh haveh toh solveh forh garbageh in/goldh outh andh preventh goldh in/garbageh out"h
(Parkinson,h 2021,h para.h 4).h Noticeh howh Mr.h Parkinsonh pointsh outh ah desiredh
stateh orh garbageh in/goldh out.
Inh datah science,h oneh seeksh toh explore,h transform,h prune,h andh cleanh fromh theh
garbageh pileh soh thath oneh canh findh theh goldenh egg.h Oneh wantsh toh offerh valuableh
insightsh toh theh userh communityh toh increaseh profits,h predicth disease,h reduceh
customerh churn,h etc.h Toh operationalizeh AI/MLh withh Bigh Data,h oneh musth
formulateh ah strategyh forh handlingh massiveh amountsh ofh datah andh addressingh
someh earlierh questions.
Second,h oneh musth oftenh chooseh betweenh businessh priorities.h Doesh theh
businessh needh toh maintainh ah real-timeh viewh ofh theh datah originatingh fromh
multiple,h disparateh sourcesh withh looseh convergence?h (Parkinson,h 2021).h Doesh
theh businessh needh toh maintainh ah comprehensive,h historicalh viewh ofh theh datah
availableh inh modernh datah warehouseh systems?h (Parkinson,h 2021).h Thish
priorityh choiceh ish ah criticalh decisionh pointh sinceh theh decisionh oftenh drivesh datah
storageh andh acquisitionh architectureh andh theh AI/MLh operationalh plan.h Forh
example,h anh AI/MLh modelh thath predictsh ah potentialh airplaneh crashh inh real-
timeh wouldh likelyh needh anh entirelyh differenth AI/MLh operationalh planh thanh anh
AI/MLh modelh thath predictsh customerh churnh forh ah popularh big-boxh retailer.h h h
Ah discussionh ofh theseh twoh challengesh mayh makeh oneh questionh whyh weh needh
anh AI/MLh operationalh plan.h Oneh musth understandh theh AI/MLh useh casesh oneh
ish attemptingh toh operationalizeh andh AI/MLh capabilities.h Differenth AI/MLh useh
casesh consisth ofh differenth Bigh Datah platforms,h services,h andh toolchainsh toh
supporth theh useh case.h Forh example,h ah conversationalh (chat)h both thath leveragesh
Naturalh Languageh Processingh (NLP)h toh determineh userh intenth basedh onh
utterancesh wouldh likelyh needh ah differenth operationalh planh thanh ah supervisedh
AI/MLh modelh thath classifiesh customerh sentimenth fromh onlineh customerh
reviews.
Theh chatboth useh caseh wouldh needh ah websiteh orh channelh toh interacth withh it.h
Mosth chatbotsh thenh needh toh communicateh toh anh NLPh serviceh toh identifyh userh
intent.h Onceh theh NLPh serviceh determinesh userh intent,h theh serviceh musth sendh
oneh orh moreh responsesh backh toh theh chatbot.h Theh moreh questionsh theh chatboth
needsh toh answer,h theh moreh datah needsh toh beh storedh (andh theh datah needsh toh beh
ofh highh qualityh soh thath theh chatboth respondsh withh theh appropriateh answer).h
Oneh musth considerh howh toh maintainh differenth versionsh ofh theh NLPh model,h
whichh determinesh userh intenth andh whenh newh versionsh ofh theh NLPh modelh areh
madeh availableh toh theh chatbot,h andh henceh theh userh community.h Beforeh
deployingh ith toh productionh andh documentingh testh results,h oneh musth considerh
testingh theh NLPh modelh forh accuracyh andh completeness.h Oneh mayh needh toh
adapth theh chatboth rapidlyh toh addressh unansweredh questions,h incorrecth
answers,h orh serviceh performanceh andh reliabilityh thath impacth users.h Oneh mayh
needh ah specifich Integratedh Developmenth Environmenth (IDE)h toh importh C#h
programmingh languageh librariesh withh varioush methodsh toh supporth theh
solution.
Theh customerh sentimenth analysish wouldh needh toh feedh theh texth inh oneh orh moreh
customerh reviewsh toh ah supervisedh AI/MLh modelh thath thenh labelsh theh texth
basedh onh theh trainedh model.h Occasionally,h oneh mayh needh toh retrainh theh modelh
toh accounth forh newh wordsh orh slangh inh languageh thath helpsh labelh theh sentimenth
ash positiveh orh negative.h Oneh musth considerh buildingh andh maintainingh anh
AI/MLh pipelineh thath ingestsh theh reviews,h storesh theh data,h andh labelsh theh data.h
Oneh mayh decideh toh trainh andh evaluateh theh AI/MLh modelh inh ah cloud-basedh
notebookh thath utilizesh Pythonh orh R.
Theseh areh twoh examplesh ofh AI/MLh useh casesh thath canh helph oneh understandh
thath noth allh AI/MLh operationalh plansh areh createdh equal.h Thereh areh manyh
factorsh andh nuancesh toh considerh whenh developingh anh AI/MLh operationalh
plan,h soh oneh musth proceedh withh cautionh andh anh eyeh towardh theh operatingh
costsh andh valueh generatedh forh theh organization.h Withh thath said,h oneh canh turnh
attentionh toh maturingh operationalh plansh toh supporth varioush AI/MLh withh Bigh
Datah useh cases.
Oneh wayh toh matureh AI/MLh operationalh plansh ish toh adopth anh MLOpsh approachh
toh supporth varioush AI/MLh useh cases.h Theh chatboth useh caseh mayh requireh anh
NLPh serviceh managedh byh oneh cloud-servicesh provider.h Inh contrast,h theh
sentimenth analysish mayh requireh ah texth analyticsh engineh managedh byh anotherh
cloud-servicesh provider.h Therefore,h oneh shouldh considerh adoptingh anh MLOpsh
approachh noth tiedh toh anyh language,h framework,h platform,h orh infrastructureh
(Machineh Learningh Operations,h n.d.).
Inh conclusion,h ith ish noth enoughh toh acquireh data,h buildh models,h testh modelh
accuracy,h andh explainh modelh efficacyh toh stakeholders.h Oneh musth alsoh
considerh deployingh andh supportingh AI/MLh modelsh inh production.h Ih lookh
forwardh toh discussingh MLOpsh inh moreh detailh inh futureh discussionsh buth
wantedh toh hearh whath othersh thinkh abouth MLOpsh andh itsh importanceh inh
supportingh AI/MLh andh Bigh Datah operations.h Doesh anyoneh haveh experienceh
withh specifich vendorsh withh ah goodh MLOpsh storyh orh solution?h Hash anyoneh inh
ourh classh hadh experienceh deployingh actualh AI/MLh modelsh inh production?h Canh
anyoneh shedh lighth onh commonh pitfallsh toh avoidh orh lessonsh learned?h Ih wouldh
likeh toh hearh abouth anyh experienceh withh MLOps.
References:
Machineh Learningh Operationsh (n.d.).h MLOps.org.h Retrievedh fromh https://ml-
ops.org/
Parkinson,h J.h (2021,h Mayh 12).h Managingh Bigh Data:h Sixh Operationalh
Challenges.h CIOh Insight.h Retrievedh fromh https://www.cioinsight.com/news-
trends/managing-big-data-six-operational-challenges/
less
Students also viewed