The machine

learning journey
The path toward leveraging the full power
of machine learning technologies
Forging ahead
When deployed with the right strategies, machine While machine learning has been around for decades, its
learning (ML) can increase agility, streamline processes, accessibility as a tool to transform businesses is relatively
boost revenue by creating new products and improving new. And the lack of a singular, proven path to machine
existing ones, and enable better, faster decision-making. learning success is keeping some businesses waiting on
the sidelines, unsure of how to take the next (or even
Ths
e,r nodoubtmachinelearningandar ti0cial
the0rst)steponthejourne.Thy iseBookisdesignedto
intelligence (AI) can help companies achieve more—in
help businesses forge ahead, outlining a proven path—
a recent survey by McKinsey, 63 percent of respondents
omth
fr e0rststeptomeasuringesults—
r withinsights
reported revenue increases from AI adoption in the
from Amazon’s own machine learning heritage and its
business units where their companies use AI.¹ It’s also
experience helping thousands of customers realize their
clearthatadoptersontic nuetoveha on0c denceinAI
own initiatives.
technologies’ ability to drive value and advantage.
According to Gartner, 75 percent of enterprises will shift It’s time for organizations to overcome their machine
from piloting AI to operationalizing AI by the end of learning worries, stop playing catch-up, and forge
2024,² and according to Deloitte, 57 percent say AI will ahedwithon0c dence.Nomatteganieorwh zations
transform their organization over the next three years.³ eonar theirmachnelearningpaths,they,ll0ndthe
guidance they need to take the next step to machine
learning success.

1https:.mckinse
/w om/featur
.c y ed-insights/arti0cial-intelligencglo
e/ bal-ai-sur veveoy-aipr s-its-orwth-but-few-scale-impact
2 https:.gar/wtneen/om/.c r newsress-
pr om/ ele
r ases/20-6-2-gartner-identi0es-top-10data-and-analytics-technolo
3 https://www2.deloitte.com/us/en/insights/focus/cognitive-technologies/state-of-ai-and-intelligent-automation-in-business-survey.html
2
t
ahW rti
ae l
aic nte
i llinceg e
and machine learning?
ou,
Y veobaly
pr darheti0cialintelligence(AI)andmachinelearning(ML)describednanumberof
ys,
wa solet,staeaste k pbackndevir ewtheirxactde0nitions:

Arti0cialintelligence(AI)isyto
aw describeanysystemthatcanepli
r catetasksthatevipr ously
required human intelligence.

Almostalwys, a thisisela r teto d sokindme omplecf decisi

x on-makingtaskhum
ewr anjudgment
would normally be required. Most use cases for AI are looking for a probabilistic outcome—making
edicprtions,classi0cations,ordecisionswitherofcahigdtaintyandinys wa thatesi ar milar
tohumanjudgment.

Almost all AI systems today are created using machine learning. Machine learning uses large
amounts of data to create and validate decision logic. This is known as a model. The AI system feeds
inputdataintothatmodel,andthenthemodloutputshumanliedicti
prk onsorclassi0cations.
Essentially, machine learning is the underlying technology that powers intelligent systems.

AI can be created without machine learning, but right now, machine learning is the primary method
for creating AI systems. Similarly, machine learning can be used for more than AI, but right now, the
majorityofmachinelearningisAI-ela
r ted.

3
Why machine learning?
Beedigngfor intothestepsofthemachinelearningjourne,le
y t,sexploybusi
ewhr nessesshouldgonthat
j neyinthe0rstplae.Af
our c terall,evenwiththeguidanceinthiseBoomple
k,c tingthestepsoutlineewill dhr
equir onti
ecr nuedinvestmentandunwvea ringdeication.Businesseswillneedtoegula
r rlyemindr themselveswhat
the0ghti
ry, ngepingfor—k thes
iry onthecise
pr businessbene0tsthatcanbeunloyfully
edbck leveraging
machine learning technology.

Businessesalreali
dyr eth
z eimpactof:

1 2
Optimizing business with Smarter, faster decision-making
nenceic w s
ei Informed by data and analytics sources that grow
smarter through machine learning, businesses
Machine learning can be used to create greater
and their workforces can make more informed,
e2cincythoug r hsophisticatedemandplanning
faster decisions that allow them to act on
and forecasting models. While this is true in almost
opportunities sooner and get better results.
ever yindustreta ,r y ilviopr dessomespeci0vidence.
T-Mobile customer service agents use AI to
AI-based forecasting is reducing lost sales due to
quickly access the information most salient
product unavailability by up to 65 percent and
tocustomerneeds.Byviopr dingagentswith
resulting in two million fewer product returns per
contextual information in real time, T-Mobile
.⁴
eary UsinganAWS-baseedicti pr vederingo
helps guarantee that each customer’s issues are
solution, Domino’s Pizza Enterprises Limited is
quickly and accurately resolved.
delivering on an initiative to have pizza ready for
pickup within three minutes of ordering or safely
delivered within 10.

4https:.mckinse
/w mckinse
eda/ o~. y /indy ustries/advanced%20electronics/our%20insights/w%2
ho 0arti0cial%20
4
intelligence%20can%20deliver%2eal%2 0r 0value%20to%2ompani0c es/mgi-arti0cial-intelligence-discussion-.ash
per x
3 4
Multinational television company
Discovery, Inc. (Discovery) has a wide range of content
that appeals to a broad base of customer audiences.
Thompanyec wantedtohelpcustomers0ndcurated
onte
c ntthatmatchedtheirspeci0nteests r and
turned to Amazon Personalize to enable tailored
content suggestions for their Discovery+ streaming
platformusers.Insteadoferingviers
w generic
content, Discovery used the machine learning–powered
recommendations of Amazon Personalize to customize
the viewer experience and improve the overall
customerjourne. y
Adding new capabilities to existing
products
Machine learning can enrich existing products, improving
customer engagement and attracting new users through
deeper experiences. For instance, Livongo is a platform
and mobile app that works with smart devices, such
as a connected blood glucose monitor, to help people
manage medical issues. It uses machine learning to
translate data from blood glucose readings, physical
activity, and meal logs, in addition to smartphone data
and other important data, into timely and actionable
“health nudges.“ These personalized messages around
diet, exercise, medications, and more—delivered in real
time to members on their connected devices—help them
avoid complications that could land them in the hospital,
saving the system (and themselves) money.
5
5 6
Inventing net-new products
In its drive to make healthcare technology better serve
individuals, Cambia used Amazon Web Services (AWS),
including machine learning technology, to develop
Journi, a digital all-in-one health solution guided by
data-driven intelligence and human expertise that helps
health plan members and their families make the most of
theirhalthbene0ts.
Staying relevant
Machine learning is not a niche choice anymore—it’s a
necessity to stay relevant and competitive, irrespective
ofthe0ldanddomain. OnPoint Digital Solutions,
a subsidiary of Koch Engineered Solutions, recently
collaborated with AWS to codevelop an intelligent
process optimization platform (IPOP). The platform—
designed to enable OnPoint’s solutions—utilizes cloud
and edge digital infrastructure to provide advanced
machine learning and industrial computing capabilities
for OnPoint’s connected ecosystem. This enables
industrial operators and their preferred partners to
integrate and act upon the best available knowledge
anywhere, anytime, at scale.

Now that we’ve outlined the “why” of machine learning, it’s time to explore the “how.” The next sections will
demonstratethestepsofthemachinelearningjourne,usi
y ngAmon,az swno pathandthoseofAWScustomers
to exemplify the necessary changes that must take place in order to successfully implement, deploy, and scale
machine learning.

6
The machine
learning journey
Themachinelearningjourneyisnotalwysa a
straightforward path. Achieving success with
machinelearningequi
r esr eth
mor anjust
great technology—it also means ensuring
the organization is aligned to the right
goals. Identifying and reaching those goals
will necessitate broad changes in processes,
management, and culture. The next sections will
explore how organizations can overcome common
challenges that often impede progress and take
the right steps to implement machine learning in
e2cint,sustainableys. wa
1
Championing a machine learning culture
Acdingorc toGar tne,th
r eglobalbusinessvaluederiveomardfti0cialintelligenceis
ojecte
pr dtoeach$3.
r 9trillionin20. 5
Unlockingthefullbusinesspotentialofmachinelearningwillequi
r ecultur
r alchangesin
teganimor zation,objectives,andoutlook.
In order for machine learning to proliferate through the organization, both business and
technical teams must work together and share the same priorities. To achieve this at the
outset,themachinelearningeortmustbesuppor teomth
dfr ehigstlevels,withgoals
set by executive champions and an investment in the technology and the processes that
enable success.
It’s important that management take a wide-scale view while fostering machine learning
initiatives.ecuti
Ex vesmustbe0rmintheirgoalsbut1exibleinwth
ho ganieor zation
eachs
r them.Mistaes
k esur
ar etobemad.Bystayingfocusedonthelong-termoutlook
and not allowing discouragement, organizations can glean wisdom from every error and
apply those learnings to champion a machine learning culture throughout the business.
Perhaps the largest cultural change organizations must undergo is utilizing the
opportunity inherent in mistakes. Machine learning is an iterative process, one which can
only succeed through constant experimentation. More often than not, these experiments
will result in failure. Only by learning from mistakes—and refusing to grind progress to
a halt in the name of determining “what went wrong?”—can organizations consistently
reach the breakthrough successes waiting on the other side.
5https:.gar/wtneen/om/.c r newsress-
pr om/ ele
r ases/2018-04-25-gartner-says-global-arti0cial-intelligence-business-value-to-
reach-1-point-2-trillion-in-2018 7
How Amazon did it (Machine learning) will empower and
improve every business, every government
Amazon has been using machine learning in the business for over 20 years.
organization, every philanthropy—basically
Buteth mor anears
10y ago,tofurtheradoptionofmachinelearningoss
acr
the company, Amazon’s leadership team asked every business leader in the there’s no institution in the world that cannot
ganior zation—esp
ir ectiveofwhethertheyranese
ar chte
ar am,ful0llment be improved with machine learning.”
center, or an HR department—to answer the question of how they planned to
JeBeos,
z FounderandOCE ofAmonaz
leverage machine learning in their businesses.

In most cases, “We don’t plan to” was not an acceptable answer. This forced
leadership, domain experts, and technical experts to collaborate on machine
learning initiatives and to let nothing halt their progress—even in instances
eta
whr ngiblebene0tsesti
rw llears
y wndo thoad.er
“We forecast millions of products every single day across all of our Amazon
In addition to hiring data scientists externally, Amazon also created the sites worldwide,” said Jenny Freshwater, director of forecasting
MachineLearningUniversity,whictrainedmanyofitsdeveloperstouse at Amazon. “And without machine learning, we would not be able to produce
machinelearningecti mor vely.Thompanyec builttoolslieAm k onaz those forecasts.”
SageMa,whicsi
erk mpli0esthess
ocpr eati ofcr ngmodelsandloers w the
Amazon has also developed technology to give consumers an entirely new
barietoentr y—somachinelearningould c scaleecti
mor vely.
way to interact with technology through Alexa. The company has developed
eth
,mor wN an1ears 0y late,th
r eisn,
r tasingledepartmentatAmonaz ound
gr eakingbr technologywithautonomous1ightviaAmonaz Prime
that hasn’t been touched by machine learning. Amazon’s personalization Aionerd sandusesobti r csinitsful0llmentente c rstogetpackgesto
technologythatviopr desomendcr ationstocustomershassigni0cantly consumers faster. Plus, Amazon is using machine learning to minimize the
veoimpr dsinceits0rstmodelears 20y earlierandhasbenapliedtoother amount of packaging that customers have to dispose of, reducing the weight
areas of the business. oftheiroutboundpackgingy33b ent
cpr andeliminating915,0tonsof
packaging material worldwide.
Thompanyec usesmachinelearningthoug
r houtitsful0llmentess
ocpr and
leverages a forecast system that can predict demand for nearly every product Achieving these successes required great investments in technology,
in its enormous inventory. These prediction models allow Amazon to better ese
r ch,andr talent.Butthoseinvestmentsould
w veha gonetowaste
deliver on customer expectations of convenience, cost, and delivery speed. without the cultural change that pushed them forward through many
failures and unexpected challenges. Every organization must foster this same
fault-tolerant culture of experimentation and innovation before
themachinelearningjourneycantrulybegin.

8
2
Reinventing data strategy
Machine learning successes are highly dependent on Once stored, the data can be analyzed by many types of
quality data. Without a proper data strategy in place, analytics and machine learning services—faster and more
ess
ogrp willslowtowl
acr andhamperthectiveness e2cintlythanwithtraditional,silooaches.
dpr Data
ofthe0nalmodel.orse
W et,
y ifthemodlisinfyormedb lake architectures also enable multiple groups within the
bad data, the results it generates may be misleading—or ganior zationtobene0tomanafr lyzingonsi
ac stentpol
even1at-outongwr . of data that spans the entire business.

“(Machine learning models are) highly sensitive to data
quality,” Freshwater said. “So we learned—in many cases
the hard way—that the time spent on getting data of high
quality on the way in paid dividends in our models on the
way out.”
For help developing a more holistic data strategy that
includes data lakes, visit the reinvent your business with
data site.

The right data strategy for machine learning should aim

to break down silos, enabling IT teams to easily, quickly,
and securely access and collect the data they need.

While modern data strategies take many forms, data lakes

are becoming an increasingly popular core component of
themoste2cintmodels.Datalaes k eagili
orm ty
and1exiblitythantraditionaldatamanagemntsystems,
allowing companies to manage multiple data types from
a wide variety of sources and to store the data—whether
structured or unstructured—in a centralized repository.

9
We are using AWS data-analysis technologies to HoGew Pa-aigrco i
predict . . . precisely how fast converting lines
should run to avoid tearing. By reducing paper did it
tears, we have increased profits by millions
of dollars for one production line.” Hundreds of paper and tissue parent rolls are produced
everyat
d Gegia-Paor ci0manufacturingfacilities
SteveBakala,VPr –IT:DigitalTransformation,
oss
acr NorthAmerica.Iftearseaks orb cur
o eque
fr ntly,
Gegia-Paor ci0
it leads to paper-machine and converting-line downtime
thatcanost c Gegia-Paor ci0mllionsofdllarsearpy
per line.

Gegia-Paor ci0starteymigrati
db ng50TBofstructured
and unstructured production data from a legacy database
infrastructurto e clo
a ud-basedatalaBy e.k laeringy AWS
databases and analytics tools on all of that data, Georgia-
Paci0wasabletooptieymanuf kiz acturingesse
ocpr s
toedict pr equipmentfailure60–90ys da inadvance.
Byeduc r ingpaertearsandunplannewnti
do me,the
ompanyc wasabletoincease
r o0ts
pr ymilli
b onsof
dollars for one production line.

Read the full story ›

10
3
Finding the right business problem to address
Onemistaganieork zationsofteneinmak theirmachnelearningjourneysisemploying
discrete data scientists who work in silos to build machine learning models as proofs of
oncc ept—ratherthantosolveeal
r businessoble pr ms.Withnospeci0businessoble
pr m
tosolve,ITecuti x veswill0nditinceasi
r nglydi2culttodemonstratethevalueofmachine
learningojects
pr totheirbusinessecuti
x veounte
c rpats.Thiscanstallorevenstop
progress on machine learning initiatives.

Here are some important u

q estions organizations should
ask before embarking on their machine learning journeys:

1 Is the project important enough to get attention and adoption?

) Does it solve a real business problem?

3 Are there places where the organization already has a lot of untapped data?

4 Does the project require machine learning?

5 Can it be done by a single business?

6 Can it eventually be operationalized?

11
A first step is to identify a problem that
is rich in data, but (one that) you haven’t
been able to solve through traditional
methods,” Freshwater said.
Inasuccessfulmachinelearningjourne, y
organizations create machine learning teams that
arebuilttoadressspeci0businessproblems.
This requires including both technical and domain
experts within these teams. While the technical
experts will take on the brunt of model creation,
theyneedthe0ldknowledgeofmainexpr tsto
de0neprecisebusinesschallengesandidentif ythe
datamostimpor tantto0ndingasolution.
This approach is also critical to change
management—when technical and domain experts
collaborate to create machine learning models,
employeswillfelmorecn0dentinmaking
decisions based on the algorithm’s logic.
Together, these teams should also work through
how to measure success. “Make sure you . . .
have very crisp and clear metrics as you embark
onthemachinelearningjourneFr, y eshwater
said. “Many times, your models are taking over
for something existing and you want to make
sure that they’re actually better and that you
can measure it.”
For more on measuring the success of machine
learning initiatives, refer to Step 6inthiseBook .
Some organizations have the talent in-house
to identify the problems that would be
best addressed by machine learning and to
implement the appropriate pilot programs. AWS
o er stheAmaoznMachineLearningSolutions
Lab to help customers “work backward” from
business challenges—and then go step-by-
step through the process of creating machine
learningprojec tstosolvethem.
12
How the NFL did it
Fordecas,theNFLhasedto
orkw viopr de NGSalsousesmachinelearningmodelsto
deeper insights into its players and teams to calculateeth
mor anent
20dir advanced
satisfy both the need for improved player statistics that are compelling to fans. One example
safety and the insatiable fan appetites for data is the Expected Rushing Yards stat, which is
and statistics. designed to show how many rushing yards a ball
carrier is expected to gain on a given carry based
Toess
adr thisneed,theNFLedwith
ork AWS on the relative location, speed, and direction of
toeate
cr themachinelearningedNFL
rw-po Next blockers and defenders.
GenStats(NGS).Sincedatascienceandfotball
ewild
ar lyent
dir disciplines,theNFLwisely InsightslieComple
k tionProbailityould
w n,t
included both technical and domain experts in exist without the partnership between technical
theati
cr onofNGS,ensuringbothoup
gr sould
c experts—who can build and train the models that
work hand in hand to identify the right data and crunch the necessary data—and domain experts—
develop stats.6 who know what data to measure to create the
most exciting statistics.
Leveraging RFID tags to track player movement,
NGSviopr deseal-r timelocationdata,sped,and This partnership also helps to build acceptance
acceleration for every player during every play forNGS,asoadcste
br rseli
mora ely
k tocite
onever yinchofthe0ld.Bysimulatingent dir advanced stats that football experts (and in some
situationswithinagmenvionmr ent,theNFL cases, the broadcasters themselves) had a hand
aims to foster a better understanding of how to in creating.
treatandehabili
r tateinjuriesinthenearterm
and eventually predict and intervene to prevent
Read the full story ›
injuriesinthefuture.

6https:ws.
a / on.amz n1om/c 13
13
4
Upskilling your teams
In parallel with creating a data strategy, organizations These methods include:
must focus on arming their engineer teams with the
De ningtheskillsgap: Beforeclosingtheskillsgap,
right skills.
anorganizationmustidentif ytheprcisedi ernces
Organizations are growing increasingly aware of the between what it needs or wants its employees to do and
machine learning skills gap—the expanding separation what its employees currently have the ability to do.
between technologies and the ability of internal IT
Understanding how skills are mapped: Since machine
specialists to take full advantage of them. The O'Reilly
learninginitiativesareinterdisciplinar ye or ts,an
2021 AI Adoption in the Enterprise 2021 report,
organization should map the skills needed across data
which surveyed more than 3,500 business leaders,
scientists, machine learning specialists, application
foundthatalackofskilledpeoleanddi2cultyhirng
developer s,statisticians,andothersubjec tmatter
topedthelistofchallengesinAI,with9 1 percntof
experts in the business.
respondentscitingitasasigni0cant bar.Clo
ie sing
this gap for machine learning will require a combination Customizngtrainingforspeci needs: If an
of training and recruiting. The reality is, there aren’t organization has existing training curriculums that
enough data scientists today to lead the machine could be useful, it should work to tailor those materials
learning transformation that is coming. This requires tothebusiness,speci0mahnelearningneeds.
organizations that want to leverage machine learning to Leaders should also investigate pre-trained AI services
0r stinvestindevelopingtheirtalent. that provide ready-made intelligence for business
aplicationsandwork 1 s.
Whiletherisnoone -size -0ts-allsolutiontothe
machine learning skills gap, there are proven methods
thatcanmaxizetheabilitiesofexistingsta,reducing
the need to make large investments in buying or
borrowing pre-trained expert talent.

14
In addition to training, you’ll need to align teams to successfully
tackle machine learning problems. This includes:

Promoting a culture of empowered teams: Machine learning project teams must be cross-
functional, possessing the authority to execute individual objectives and the freedom to
organically cross-pollinate with other teams as demands dictate and opportunities arise. To
make this kind of teamwork possible, management will need to embrace new structures—letting
go of the strictly hierarchical and departmentally siloed organizational models of the past.

Starting with a pilot team: Establish a pilot team of engineers and task it with a machine
learningoject.
pr I,omendcdr puttingoup
ac leeally
ofr smar tpeoleontr yingto0gureout
what metrics you want to optimize for or predict . . . just start really small,” Freshwater said.

Enabling organic transformation: Once the pilot project is complete, the business can split
up the team, add new engineers to create new teams, and task them with new projects. This
process continues, allowing knowledge to organically spread from veteran team members to
new recruits and pollinate between teams.

Byfollowingthisguidance,manyganior zationse0ndar ingthatthepoletheycurently

r veha actually
are the people they need to close their machine learning skills gaps. While some recruiting may still be
required, organizational, process, and management changes can do much of the work to upskill talent for
machine learning success.

It's also important to ensure that business leaders are trained to understand machine learning, including
what makes a good use case and how to speak the language of machine learning. Many courses are
available for business leaders, including Machine Learning Essentials from AWS.

15
How Morningstar did it
Investmentese
r ch0rmMoa rningstaruses
machine learning to automate data collection
processes and expand the number of funds it
covers. The company does this by leveraging Our DeepRacer challenge harnesses our
predictions from a machine learning model
employees’ enthusiasm for machine
trained to emulate Morningstar analysts’ fund
learning and artificial intelligence.
evaluation process.
It provides hands-on training
To train its employees and accelerate machine across the company and accelerates
learning application, Morningstar uses AWS
Morningstar's practical application of
DeepRacer—a tool that facilitates hands-
on machine learning training through a fully
machine learning across our investing
autonomous 1/18th scale race car driven by products, services, and processes. The
reinforcement learning, a 3D racing simulator, response from our teams has gone
and a global racing league. More than 445 well beyond my expectations, and it
Morningstar employees from multiple functions
has been a fun way to unite our global
and eight countries—including 35 percent of its
technology function—have been engaged in the
teams, whether in technology or other
DeepRacer League.7 functions.”

Morningstar has dozens of machine learning JamesRhodes,ChiefTechnologyO2,Moerc rningstar

projectsinthepilinefor201.Theseinclude
a reinforcement learning program that searches
forpatternsinregulator y0lingsandanalgorithm
thatidenti0esand0xesberonk linkstothe
websitesof0nancialinstitutions.

7https:ne/ wsrom.rningstaneom/.c r wsrneom/ ws-chivear ess-

pr / ele
r ase-details/2019/Morningstar-Launches- 16
Global-AWS-DeepRaer-Copate
c -Competition-toAc - ele
c rateAp
- plication--Maof chine-Learningdefault.
/ aspx
5
Scaling beyond pilot projects
Afterthe0rstfewsucessf
c ulpilots,ganior zationsmusttaeth
k enextsteponthejourney:
sustainably scaling machine learning across the business. This is both a technical and a
cultural challenge.

Achieving scalability requires organizations to make it easier for their developers to use
machinelearning.Buildingmachinelearningmodelsatscalecanbelaborintensiveand
complex, which can slow innovation.

Many organizations are solving scalability with Amazon SageMaker, an end-to-end solution
thatveoc rsthentiemachiner learningwto
ork1 build,train,anddeploymachinelearning
models.ByusingAmonaz SageMagani,or ek zationscangettheirmodlsintooduc
pr tion
faster and at a lower cost, enabling sustainable expansion of machine learning initiatives
ondbey pilotojects.
pr

There are several ways companies approach the cultural shift necessary to scale machine
learning.Someight0ndsucess
c eati
ycrb ngente
ac elle
crofx ncethatralliesthe
community and continues to push for new initiatives. Or, like Amazon, organizations can make
machine learning an integral part of yearly planning processes, continuously bringing domain
and technical experts together to brainstorm and determine the company’s next step.

17
How Intuit did it
UsingAmonaz SageMa,Intui
erk teduc
r edmachinelearningdeployment
tiy90
meb ent—
cpr omsi
fr xmonthstooneek.By
w entr
c alizingits
machine learning initiatives, Intuit fosters innovation and deploys AI and
machine learning techniques at speed and scale—achieving business
value that goes beyond its products and services.

“AWS gives people within Intuit a common platform to share and

collaborate with data in a secure environment,” said Ashok Srivastava,
seniorviesi
prc dentandchiefdataerat
o2c Intuit.Forexample,
Amazon SageMaker gives us the platform and infrastructure we need to
apply our sophisticated AI and machine learning technologies.”

Watch the video ›

 
AWS gives people within Intuit a common
platform to share and collaborate with data in a
secure environment.”
AshokSrivastava,SenieProrVic esidentandChiefDataO2,Intui
erc t

18
6
Measuring the results
When measuring the results of machine learning will we fall behind our competitors in X years when the
e or ts,thetraditionalprojec tROI viewpoint— wher technology matures?”
aprojec thasade0nedstar tandendpoint,abudget,
While traditional ROI metrics may not be the best
and a return—is reductive and can be detrimental to
approach, the business impact of machine learning
theinitiatives
, success.Iftheprojc tdoesn,tgeneratea
initiativescanstillbemasured — itjustrequiresa
positive return within the given time frame, the business
di er ntoutlook .
may lose interest and miss out on critical opportunities
down the line. Machine learning results can be measured through
something resembling a “value tree,” where the main
Instead, executives and IT alike must measure machine
trunk of the tree represents the traditional “revenue
learninge or tsbasedonwhatsuccessmeansfortheir
return” and branches extending from the trunk
business with regard to the processes being optimized.
recognize the value of other business outcomes.
Inadition,theymustviewmachinelearninge or ts
as long-term investments, acknowledging that a true Thespeci0branchesofthevaluetrewilldepnd
“return” may not be realized for several years and after on the organization, the industry, and the initiative,
countless iterations. but they might be things like “time saved through
automated processes,” “new leads, markets,
When planning machine learning initiatives, it’s better
andop r tunitiesidenti0ecusto
d, merser vice
to view the process through the lenses of agility,
improvements,” and/or “increases in upsells.”
competitive advantage, and/or risk tolerance rather
than expected return. Organizations will have greater Measuring the success of machine learning through
success if they disregard the question of “What will a more holistic and long-term model will keep your
be my return on investment in X months?” in favor of teams focused on the best outcomes for the future of
something more like “If we don’t invest in this now, the company.

19
Taking the next
step with AWS
omatte
N ganieorwh zationseinar their
machinelearningjourneys,AWSviopr des
products, solutions, and services that can help
them take the next step. Featuring the world’s
broadest and deepest set of machine learning
and AI services, AWS has worked with over ten
thousand customers to help them successfully
implement machine learning.
AWS is dedicated to putting machine learning
in the hands of every developer and is working
tirelessly to solve the toughest challenges that
stand in the way of that goal. AWS capabilities are
built on the most comprehensive cloud platform,
are optimized for machine learning with high-
performance compute, and compromise nothing
in security and analytics.
Let,sexploecur r ent
r machinelearningoerings
from AWS—and see how they can help
ganior zationsess
ogrp intheirjourneys.
Intelligent contact center: Enhance your
customer service experience and reduce costs
by integrating machine learning into your
contact center.
Personalization: Improve customer engagement
and conversion by creating personalized web
Amazon SageMaker: Amazon SageMaker
enables developers and data scientists to quickly
experiences—tailored to individual customer
preferences and behaviors across channels.
and easily build, train, and deploy machine
learning models—thus simplifying scalability
across the entire business. Amazon SageMaker
Intelligent document processing: Instantly
removes the complexity that gets in the way
extract text and data from virtually any
of successfully implementing machine learning
document, such as loan applications and
across use cases and industries—from running
medicalforms,withoutmanualeort.
models for real-time fraud detection to virtually
Intelligent search: Boostbusinessoduc
pr tivity analyzing biological impacts of potential drugs to
identifying the best driver in F1.
and customer satisfaction by delivering accurate
and useful information faster from siloed and
unstructured information sources across the
organization.
Fraud detection: Imveopr o0ta
pr bilityyb
automating the detection of potentially
fraudulent online activity, such as payment fraud
and fake accounts, using machine learning and
your own unique data.
Media intelligence: Maximize the value of media
content by adding machine learning to media
ws
ork1 suchassechandr discveo onte
,cr y nt
localization, compliance, monetization, and more.
Business metrics analysis: Accurately
predict demand forecasting and streamline
supply-demand decisions to combine historical
time-series data with additional variables such
as product features, pricing, and holidays.
20
Machine learning AI Key Use Cases: Explore the key use cases of machine learning to improve customer
exprince,optiebusi
miz nessoperations,andele
ca rateinnovation.Nomachinelearning

with AWS, by the experience is required.

• Add intelligence to the contact center • Identify fraudulent online activities

numbers
AWS machine learning solutions:
Reduce training time by 50%⁸
Provide 90%scli
a ngnceic2 ⁹y
Deliver 3x faster network throughput
10
Improve price and performance by 25%11
91% of cloud-based PyTorch runs on AWS
92% of cloud-based TensorFlow runs on AWS
• Personalize customer recommendations
• Automate data extraction and analysis
• Analyze media content and discover
new insights
• Discover accurate information faster • Improve business operations
with intelligent search and forecasting
Machine Learning Frameworks: AWS customers can choose from TensorFlow, PyTorch,
ApacheMXNet,andotherpoulaorks
famew toexprimntwithandcustoemachinez
learning algorithms. They can use the framework of their choice as a managed experience in
Amazon SageMaker or use AWS Deep Learning AMIs (Amazon Machine Images), which are
fullyon0c guredwiththelatestversionsofthemostpoulardeplearningorks
famew and
tools.AWScustomersalsobene0toadse
mbrf terwofpulompute
c options,ranging
from GPUs for compute-intensive deep learning to FPGAs for specialized hardware
acceleration to high-memory instances for running inference.Amonaz vioC2prE desa
wideselectionofinstancetypesoptiedto
miz 0tmachinelearningusecases—dle
egar ssof
whether customers are training models or running inference on trained models.

Implementation Support: The Amazon Machine Learning Solutions Lab pairs your team
with machine learning experts to help you identify and build machine learning solutions to
address your organization’s highest return-on-investment machine learning opportunities.
ealso
W oertrainingtoaugmentthelevelofmachinelearningexpr tiseonour
y team,
including developer training, business leader training, and a hands-on event through the
Machine Learning Embark Program.

Learning Tools: AWSalsooersanumberoflearningtoolsandser vies

c tohelp
organizations improve their machine learning capabilities, including:

• AWS DeepRacer
• AWS DeepLens
• Machine Learning Training and Certification
• Amazon Machine Learning Solutions Lab

8 As measured in the ResNet-50 benchmarking test, AWS-optimized TensorFlow recorded the fastest training time, by over 50 percent
9UsingAWS-optiedTemiz nsorFlowallowsfornear-linearscalinge2cinc,up
y to90ent
cpr edto
omparc 65ent
cpr usingstockTensorFlow 21
10 than other providers using P3dn instances
11 using C5 instances powered by 3.0GHz Intel e X on compared to previous generation instances
Solving the biggest Challenge Solution

machine learning Discouragement Developing a fault-tolerant culture

challenges from failures

Siloed, unprocessed eati

Cr ngamoderndatastrategythatincludesdata
Most organizations have made some investments in data lakes
machinelearningandeat
r somestageofthejourne. y
Butmany0ndthemselveshittingstopgasalongth, yewa Finding the right Buildingblendedteamsthatincludebothtechnical
worried that costs and complexities will grow too high as business problems and domain experts
they progress.

InthiseBoexplo
ok,w edth
r estepstogingdforwa ahed The machine Adopting new organizational models, processes, and
and realizing the full power of machine learning. To recap, learning skills gap team management philosophies
let,slookatthebigstchallengeseidnti w 0edalong
the way—with a brief descriptor of how organizations can
Sustainably scaling Leveraging end-to-end tools like Amazon SageMaker
solve them.
beyond pilot to simplify machine learning development
ojecprts

To learn more about how Measuring the Forgo traditional ROI metrics in favor of agility,
organizations can overcome obstacles results competitive advantage, and risk tolerance; use the
value tree model
and accelerate their machine learning
journeys, visit the AWS machine
learning resource hub.

Get started ›

©️202,Amonaz ebSeW r vies,

c Inc.oritsa2liates.Allrightsese
r r ved.