ML Centrepiece for Digital Transformation

© 2019, Amazon Web Services, Inc. or its affiliates. All rights reserved.S U M M I T
AWS Machine Learning
Centerpiece for digital transformation
Mohammed Jamal
Aramex
Clive Charlton
Lead Solutions Architect
AWS, Sub-Saharan Africa

40% of digital transformation initiatives
supported by AI in 2019
—IDC 2018
InnovationDecision
making
Customer
experience
CE N TE RP IE CE FO R D IG ITA L TRA N SFO RM A TIO N
Business
operations
Competitive
advantage

Our mission at AWS
Put machine learning in the
hands of every developer

W H Y A W S FO R M L?
200 new features and services
launched this last year alone
Unmatched flexibility
Broadest and
deepest set of AI
and ML services
70% cost reduction
in data-labeling
10x faster performance
75% lower inference cost
Accelerate your
adoption of ML
with SageMaker
Built on the most
comprehensive cloud
platform optimized for ML
AWS holds the top spots
on Stanford’s benchmark,
for fastest training time, lowest
cost, lowest inference latency

10,000+ customers | 2x the customer references | 85% of TensorFlow projects in the cloud happen on
AWS

FRAMEWORKS INTERFACES INFRASTRUCTURE
AI Services
Broadest and deepest set of capabilities
T H E A W S M L ST A C K
VISION SPEECH LANGUAGE CHATBOTS FORECASTING RECOMMENDATIONS
ML Services
ML Frameworks + Infrastructure
A M A Z O N
P O L L Y
A M A Z O N
T R A N S C R I B E
A M A Z O N
T R A N S L A T E
A M A Z O N
C O M P R E H E N D
& A M A Z O N
C O M P R E H E N D
M E D I C A L
A M A Z O N
L E X
A M A Z O N
F O R E C A S T
A M A Z O N
R E K O G N I T I O N
I M A G E
A M A Z O N
V I D E O
A M A Z O N
T E X T R A C T
A M A Z O N
P E R S O N A L I Z E
Ground Truth Notebooks Algorithms + Marketplace Reinforcement Learning Training Optimization Deployment HostingAmazon SageMaker
F P G A S
A M A Z O N
E C 2 P 3
& P 3 D N
A M A Z O N
E C 2 G 4
A M A Z O N
E C 2 C 5
A W S
I N F E R E N T I A
A W S I o T
G R E E N G R A S S
E L A S T I C
I N F E R E N C E
D L C O N T A I N E R S
& A M I s
The
picture
can't be
display
ed.

FRAMEWORKS INTERFACES INFRASTRUCTURE
AI Services
Broadest and deepest set of capabilities
T H E A W S M L ST A C K
VISION SPEECH LANGUAGE CHATBOTS FORECASTING RECOMMENDATIONS
ML Services
ML Frameworks + Infrastructure
A M A Z O N
P O L L Y
A M A Z O N
T R A N S C R I B E
A M A Z O N
T R A N S L A T E
A M A Z O N
C O M P R E H E N D
& A M A Z O N
C O M P R E H E N D
M E D I C A L
A M A Z O N
L E X
A M A Z O N
F O R E C A S T
A M A Z O N
I M A G E
A M A Z O N
V I D E O
A M A Z O N
T E X T R A C T
A M A Z O N
P E R S O N A L I Z E
Ground Truth Notebooks Algorithms + Marketplace Reinforcement Learning Training Optimization Deployment HostingAmazon SageMaker
F P G A S
A M A Z O N
E C 2 P 3
& P 3 D N
A M A Z O N
E C 2 G 4
A M A Z O N
E C 2 C 5
A W S
I N F E R E N T I A
A W S I o T
G R E E N G R A S S
A M A Z O N
E L A S T I C
I N F E R E N C E
D L C O N T A I N E R S
& A M I s
The
picture
can't be
display
ed.

Bringing machine learning to all developers
A M A Z O N SA G E M A K E R
Collect and
prepare
training data
Choose and
optimize your
ML algorithm
Set up and manage
environments
for training
Train and
tune model
(trial and error)
Deploy
model in
production
Scale and manage
the production
environment

Collect and
prepare
training data
Choose and
optimize your
ML algorithm
Set up and manage
environments
for training
Train and
tune model
(trial and error)
Deploy
model in
production
Scale and manage
the production
environment
Pre-built
notebooks for
common problems

Collect and
prepare
training data
Choose and
optimize your
ML algorithm
Pre-built
notebooks for
common problems
Built-in, high
performance
algorithms
• K-means clustering
• Principal component analysis
• Neural topic modelling
• Factorization machines
• Linear learner (regression)
• BlazingText
• Reinforcement learning
• XGBoost
• Topic modeling (LDA)
• Image classification
• Seq2Seq
• Linear learner (classification)
• DeepAR forecasting

Collect and
prepare
training data
Choose and
optimize your
ML algorithm
Set up and manage
environments
for training
Train and
tune model
(trial and error)
Deploy
model in
production
Scale and manage
the production
environment
Pre-built
notebooks for
common problems
Built-in, high
performance
algorithms
One-click
training

Collect and
prepare
training data
Choose and
optimize your
ML algorithm
Set up and manage
environments
for training
Train and
tune model
(trial and error)
Deploy
model in
production
Scale and manage
the production
environment
Pre-built
notebooks for
common problems
Built-in, high
performance
algorithms
One-click
training Optimization

Collect and
prepare
training data
Choose and
optimize your
ML algorithm
Set up and manage
environments
for training
Train and
tune model
(trial and error)
Deploy
model in
production
Scale and manage
the production
environment
Pre-built
notebooks for
common problems
Built-in, high
performance
algorithms
One-click
One-click
deployment

Collect and
prepare
training data
Choose and
optimize your
ML algorithm
Set up and manage
environments
for training
Train and
tune model
(trial and error)
Deploy
model in
production
Scale and manage
the production
environment
Pre-built
notebooks for
common problems
Built-in, high
performance
algorithms
One-click
One-click
deployment
Fully managed with
auto scaling, health checks,
automatic handling
of node failures,
and security checks

One-click
model training
and deployment
Train once
run anywhere
10x
better algorithm
performance
2x
performance increases from
model optimization with
Amazon SageMaker Neo
70%
cost reduction for data
labeling using Ground Truth
75%
cost reduction for inference with
Amazon Elastic Inference
REDUCE COSTS INCREASE PERFORMANCE EASE-OF-USE
CU STO M M A CH IN E LE A RN IN G FO R Y O U R B U SIN E SS
AMAZON SAGEMAKER

• Fully-managed training and hosting
• Near-linear scaling across 100s of GPU
• 75% lower inference costs with
Amazon Elastic Inference
• 3x faster network throughput with Amazon
EC2 P3
T H E B E ST P LA C E T O R U N T E N SO R FLO W
Amazon SageMaker is the best place
to run TensorFlow in the cloud
65% Stock
TensorFlow
AWS-optimized
TensorFlow90%

Modernize your contact center to improve customer service
P U T M L T O W O R K FO R Y O U R B U SIN E SS
conversational chat bots | call transcription | intelligent routing | sentiment analysis
VoC analytics text-to speech | multilingual omni-channel communication
AMAZON
POLLY
AMAZON
TRANSCRIBE
AMAZON
TRANSLATE
AMAZON
COMPREHEND
AMAZON
LEX

Meaningful
customer interactions
Liberty Mutual uses Amazon Lex and AI services
to develop natural language-driven conversational
apps to allow customer service agents to respond
to customer requests with real-time and contextual
intelligence, improving response time,
and quality of service.

Use AI services to strengthen safety and security
accurate facial analysis | identity protection | metadata extraction
AMAZON
REKOGNITION
IMAGE
AMAZON COMPREHEND &
AMAZON COMPREHEND MEDICAL
AMAZON
REKOGNITION
VIDEO

Real-time
identity verification
Aella Credit uses Amazon Rekognition to
analyze images to verify an individual’s identity
in real-time without human intervention,
allowing it to provide instant loans to eligible
customers through its mobile app.

Automate media workflows to reduce costs and monetize content
content moderation | contextual ad insertion | searchable media library
custom facial recognition | multi-language metadata search
AMAZON
REKOGNITION
IMAGE
AMAZON
REKOGNITION
VIDEO
AMAZON
COMPREHEND
AMAZON
TRANSCRIBE
AMAZON
TRANSLATE
AMAZON
TEXTRACT

Scaling video
indexing
C-SPAN uses Amazon Rekognition to
automatically index video news footage
for search. With Amazon Rekognition, C-SPAN
reduced indexing time per video from one
hour to 20 minutes and uploaded 97,000
images in under two hours.

Reduce localization costs and improve accuracy
custom vocabulary | timestamp generation | secure real-time translation | language identification
AMAZON
POLLY
AMAZON
TRANSCRIBE
AMAZON
TRANSLATE
AMAZON
COMPREHEND

Scaling real-time
translation
Using Amazon Translate, Lionbridge is able to
scale machine translation in order to localize
content faster and in more languages. Using
Amazon Translate, Lionbridge was able to reduce
translation costs by 20%.

Understand the voice of your customer
sentiment analysis | app localization | translation services | transcription services | cataloging media | accessibility
AMAZON
REKOGNITION
IMAGE
AMAZON
REKOGNITION
VIDEO
AMAZON
TRANSLATE
AMAZON
TRANSCRIBE
AMAZON
COMPREHEND

Targeted
customer acquisition
VidMob uses Amazon Rekognition and
Amazon Transcribe for metadata extraction
and sentiment analysis, to help marketers
understand which videos resonate with
audiences. This allows marketers to promote
targeted content to acquire new customers.

Personalize customer experiences with targeted recommendations
recommendation technology used by Amazon.com | context-aware recommendations
sentiment analysis | VoC analytics
AMAZON
PERSONALIZE
AMAZON
REKOGNITION
IMAGE
AMAZON
REKOGNITION
VIDEO
AMAZON
COMPREHEND

Personalizing
customer experiences
Domino’s uses Amazon Personalize to customize
and scale relevant marketing communications to
customers based on time, context, and content,
thereby improving and enhancing their
experience with the Domino’s brand.

Accurately forecast future business outcomes
forecasting technology used by Amazon.com | multiple time-series data
forecast scheduling and visualization | supply chain integration
AMAZON
FORECAST

Increase efficiency with automated document analysis
optical character recognition (OCR) | automatic data extraction | natural language processing
intelligent search | text analytics
AMAZON
TEXTRACT
AMAZON COMPREHEND &
AMAZON COMPREHEND
MEDICAL

1
Create
the loop
Connect technology initiatives
with business outcomes
2
Assess your structured and
unstructured data sources
Advance your
data strategy
?
3
Put machine learning in the
hands of your developers
Organize
for success

H O W W E C A N H E L P
• Brainstorming
• Custom modeling
• Training
• Work side-by-side with Amazon experts
ML Solutions Lab
• Practical education on ML for new
and experienced practitioners
• Based on the same material used
to train Amazon developers
Machine Learning
Training and Certification

Aramex
• Leading provider of courier delivery, logistics,
and e-commerce services in the Middle East and
globally
• Operations span across 66 countries and
employing over 17,000 professionals
• Enhancing the customer experience through
rapid digitalization is critical to our success
• Big data & AI is one of the key strategic focus
areas

How AI is transforming Aramex
CUSTOMS DUTY
PREDICTION
Customs Duty: $25
O
…
P…
O
…
P…
CAPACITY PLANNING

Objective
Predict the date when the shipment will be
delivered to the consignee
Business Case
• Better customer experience
• Improved capacity planning
• Reduction in inbound calls at call center
Challenges
• Seasonality – DOW, HOD, Month, Country, etc.
• Variations in entity capacity at different points in time
• Managing customer perception

Data science is all about failing fast
So let’s first build a quick MVP…

Transit time v0 (quick MVP)
Weight COD & customs
value
Seasonality
Origin &
destination
Product
type
Pre-
paid/COD?
XG Boost ML Model
3 September,
2019
+ Pickup
date
Key notes:
• Training period: Three months
• Type: Regression
• Output unit: Hours
At origin Flight
transit
In custom Delivery+ + +

Transit time problem statement
Fail fast but learn faster
So what did we learn from our MVP?

Transit time lessons learned
Lesson 1: Destination City
• Incorrectly interpreted “Destination Entity” as the final destination
• Each entity was further divided into multiple cities, which could essentially have very
different delivery times
JFK Jeddah JFK Jeddah
Taif
Madinah Tabuk…

Madinah Tabuk…
Model Errors
O
rigin
Transit
Custom
O
rigin
Transit
Custom
Lesson 2: Additive vs End-to-End Model
• The base model consisted of four sub models, each predicting the transit time for each
leg; finally the four predictions were summed to give the end-to-end transit time
• The problem with this approach was that the error induced by each of the models
added up to a more significant end-to-end error
• Resolved this by training a single end-to-end model; this model was able to implicitly
learn the transit time for each leg, as well as the end-to-end transit time

Lesson 3: Prediction Confidence
• The base model was a regression model that predicted a numerical value. This model
however did not indicate the confidence in the prediction.
• We moved to a classification model that assigned probabilities to a range of possible
days (1-10 days).
• Doing so enabled us to compute delivery intervals with probability thresholds.
Day 0 1 2 3 4 5 6 7 8 9 10
Probability 0.05 0.05 0.1 0.4 0.2 0.2 0.0 0.0 0.0 0.0 0.0
Probability Threshold = 0.8
Cumulative probability = 4 days
Interval = 3 – 5 days

Transit time problem statement
Putting learning into practice…

Transit time deep learning model architecture
Weight, PCS, Value
Origin Destination Product Type
...
LSTM LSTM LSTM LSTM
t - 30 t - 29 t - 28 t
Ʃ
Categorical Input Layer
Time Series Feature Layer
Ʃ
Softmax Layer
Prediction: Transit Time
Dense Hidden Layers
Classification Layer
Time Series (open shipments, stats, etc.)

Data architecture
Amazon S3
AWS DynamoDB
Maker
Data Lake/Data Warehouse
On-premises
SQL Server

ML Centrepiece for Digital Transformation

Recomendados

Recomendados

Más contenido relacionado

Similar a ML Centrepiece for Digital Transformation

Similar a ML Centrepiece for Digital Transformation (20)

Más de Amazon Web Services

Más de Amazon Web Services (20)

ML Centrepiece for Digital Transformation