More Related Content Similar to The Hive Think Tank: AI in The Enterprise by Venkat Srinivasan (20) The Hive Think Tank: AI in The Enterprise by Venkat Srinivasan1. © Rage Frameworks Inc, 2016. All rights reserved.
Enabling the Intelligent Enterprise
AI in the Enterprise
The Hive Think Tank
Jan 26, 2017
2. © Rage Frameworks Inc, 2016. All rights reserved. | 2
The Resurgence of AI …it’s possible
Source: The Intelligent Enterprise in the Era of Big Data, Srinivasan, Wiley, 2016
Google’s DeepMind wins historic Go content 4-1
The recent accident on a Tesla vehicle in
autopilot mode
3. © Rage Frameworks Inc, 2016. All rights reserved. | 3
AI in the Enterprise
Key Dimensions of Machine Intelligence
…it’s possible
Computer Visioning
Solutions
Non-Visioning
Solutions
Computational
Statistics
Knowledge
Acquisition /
Representation
Computational
Linguistics
Source: The Intelligent Enterprise in the Era of Big Data, Srinivasan, Wiley, 2016
4. © Rage Frameworks Inc, 2016. All rights reserved. | 4
AI in the Enterprise
A Taxonomy of Machine Intelligence Problem Types
…it’s possible
Ad Hoc Search
Clustering
Prediction
[Quantitative data]
Extraction
Classification
[Qualititative, Hybrid data]
Interpretation
[Natural language,
Other data]
Prediction, Classification
Artificial
Intelligence
(Machine Intelligence)
Intelligence Thru
Explicitly
Assumed Models of Data
Learn
from Data
Algorithmically
Learn to Interpret/
Understand
Meaning
Source: The Intelligent Enterprise in the Era of Big Data, Srinivasan, Wiley, 2016
5. © Rage Frameworks Inc, 2016. All rights reserved. | 5
AI in the Enterprise
Machine Intelligence Acquisition Methods
…it’s possible
Source: The Intelligent Enterprise in the Era of Big Data, Srinivasan, Wiley, 2016
6. © Rage Frameworks Inc, 2016. All rights reserved. | 6
AI in the Enterprise
Machine Intelligence Acquisition Methods
…it’s possible
Pragmatics
Automated
Knowledge
Discoverer
Domain
Discourse
Model
Public
Content
Private
Content
RAGE
KnowledgeNet™
WordNet
ConceptNet
FrameNet…
Cognitive Semantic Networks
Deep Parsed Linguistic Maps
Topic Clusters
Syntactic Results
Semantic Roles
Seed Concept (Optional)
Knowledge Type Constraints
Source: The Intelligent Enterprise in the Era of Big Data, Srinivasan, Wiley, 2016
7. |
7
AI in the Enterprise
Machine Intelligence - Functional Architecture
…it’s
possible
Source:
The
Intelligent
Enterprise
in
the
Era
of
Big
Data,
Srinivasan,
Wiley,
2016
8. © Rage Frameworks Inc, 2016. All rights reserved. | 8
AI in the Enterprise
Machine Intelligence VS Intelligent Machines
…it’s possible
Machine Intelligence
Computational
Statistics
Knowledge
Acquisition /
Representation
Computational
Linguistics
Source: The Intelligent Enterprise in the Era of Big Data, Srinivasan, Wiley, 2016
Intelligent Machine
Ingest
Process
Decide
Document
Communicate
Intelligence
Analytics
Integrating into a
Mission Critical
Production
Business Process
9. © Rage Frameworks Inc, 2016. All rights reserved. | 9
Examples of Intelligent Machines in the Enterprise …it’s possible
Wealth Management Active Advising
Commercial Loan Origination
Financial Statement
Spreading
Client Onboarding
Data Quality Monitoring
Real Time Intelligence for
Cap Markets
Knowledge Management
Customer and Market
Intelligence
RAGE KYC Framework
RTITM : Credit &
Supplier Risk
Sales Lead Generation
Automated Contract Review
Customer Service IntelligenceAutomated Billing Reconciliation
Supply Chain Cost Audit
Business
Rules Engine
Model Engine
NLP Engine
Quality
Assurance
Framework
Web Services
Engine
Decision
Tree Engine
Computation
al Linguistics
Engine
Model
Network
Engine
Data Access
Engine
Desktop
Integration
Engine
Connector
Factory
Engine
Questionnair
e Engine
Real Time
Content
Integration
Engine
Assignment
Engine
Message
Engine
External
Object
Engine
Extraction
Engine
Repository
Intelligent
Doc Builder
Engine
User
Interface
Engine
Process Assembly Engine
10. © Rage Frameworks Inc, 2016. All rights reserved.
Enabling the Intelligent Enterprise
Extraction from Semi, Unstructured Documents
Financial Information
11. © Rage Frameworks Inc, 2016. All rights reserved. | 11
RAGE LiveSpread™
Process Flow
…it’s possible
Data extraction from
any format including:
pdf, excel, images,
paper, web scraping
etc.
Extraction Normalization User defined
Normalization
Exceptions and
Quality
Presentation and
Analytics
Integration
Normalized using -
Industry templates;; Pull
from footnotes;;
Footnote interpretation
linked to line items;;
30 plus language
Rules buy Country;;
User defined
normalization ruleset
via self service
screens
Exception handling of
data accuracy, in-
built quality
assurance and
business rule
compliance
Presentation of
spread data and
financial ratio
calculations
Integration into
client’s core systems
Analytics (add on)
Credit score cards and risk monitoring
Equity models
Custom Analytics (M&A deal sourcing, Audit etc.)
Feature snapshot
• Industry specific normalization of data
• Analysis of revolving credit lines
• Auditors’ opinion on the financial statements captured
• Key break-ups from notes to financials
• Industry ID data: NAICS, SIC or GICS codes
• Adjustments for extraordinary/one-time/non-cash items
• Details on operating leases and contractual obligations
• Financial covenant tracking and alerts
• Automated QA checks
• Multiple MRA load/delivery options
13. © Rage Frameworks Inc, 2016. All rights reserved. | 13
Input Form and Format Variability …it’s possible
14. © Rage Frameworks Inc, 2016. All rights reserved. | 14
Normalization Example
Non English Document
…it’s possible
Normalized Output
Italian document
Normalization rules
15. © Rage Frameworks Inc, 2016. All rights reserved. | 15
Example of Extraction from Footnotes …it’s possible
Notes to the financial statements
(note 4)
Final Output - Spreadsheet
Balance Sheet
After pulling
out breakups
from notes to
the financials
Before
capturing the
breakups
Breakups for fixed assets identified and
extracted from notes
16. © Rage Frameworks Inc, 2016. All rights reserved. | 16
Example of Extraction from Footnotes …it’s possible
Key breakups for Operating expenses
were pulled from Operating Leases note
as they were unavailable in the Income
Statement.
Notes to the financial statements
Final Output - Spreadsheet
Income Statement
17. © Rage Frameworks Inc, 2016. All rights reserved. | 17
Normalizing GAAP Rules Across Countries …it’s possible
Final Output - SpreadsheetOriginal Document
Normalized Metadata – Rule File
Canadian GAAP
US GAAP
Bank charges map differently to Interest expenses [As per
Canadian GAAP] and to Other expenses [As per US GAAP]
18. © Rage Frameworks Inc, 2016. All rights reserved.
Enabling the Intelligent Enterprise
Classification with Natural Language Understanding
Customer Service Intelligence
© Rage Frameworks Inc, 2016. All rights reserved.
19. © Rage Frameworks Inc, 2016. All rights reserved.
U.S.
Background on the customer data analytics project
| 19
• The objective: To aggregate all the unstructured data, within Seibel, from
various communication types with the customers, extract, interpret, analyze,
and deliver insights to make decisions rooted in data and insights.
• Key questions for the analysis:
• What are the primary reasons reasons customers are contacted or
customers contact us? How do these reasons rank by volume?
• What are the underlying reason customers are contacted or customers
contact us?
• Do these reasons shed light on the process elements or processes that
may be resulting in repeated customer outreach to us or customer
dissatisfaction?
• Are there any inefficiencies in customer service processes, based on the
service request fulfillment attributes e.g. number of times back and forth
communication with the customers, which can shed some light on the
process inefficiencies?
20. © Rage Frameworks Inc., 2016. All rights reserved
Emails
Semantic
Topic
–
Order
Rescheduling
Semantic
Topic
-‐
Order
Cancellation
Subject: Weston
Pallet
Count
From: kgxxxxx
To: Exxxx Dxxxxx;
Jxxxxx fxxxxx;
Cxxx
CC: Daxx Wxxxx;
Dxxx Rxxxxx
Date: 2014-‐11-‐06
12:25:57
Hi
Eliza,
The
count
for
today
is
1299
@
11:30
am
The
pallet
count
is
high
with
production
requirements.
Please
cancel
Thursday
3
pm
load
4703423658.
Take
Care,
Kexxxx
From: Cxxxx-‐Cxxx
Sent: Thursday,
November
06,
2014
12:42
PM
To: Kxxxx Gxxxxxx;
Exxxx Dxxxxx;
Jxxxxx fxxxxx;
Cxxx
Cc: Daxx Wxxxx;
Dxxx Rxxxxx
Subject: RE:
Weston
Pallet
Count
Good
day,
Please
be
advised
that
PO#4703423658
has
been
changed
to
tomorrow
delivery
at
3pm
as
requested
in
yesterdays
email.
Please
see
the
remaining
orders
for
today/tomorrow;
Thank
you/Merci,
Allxxxx Mcxxxx
Original
Topic
– Pallet
Count
21. Enabling the Intelligent Enterprise
Natural Language Understanding + Extraction
Logistics Cost Audit & Contract Review
© Rage Frameworks Inc, 2016. All rights reserved.
22. © Rage Frameworks Inc, 2016. All rights reserved.
| 22
Intelligent Machine for Cost Audit
How machine learning is applied to deliver insights and speed-to-value?
…it’s possible
Extract Integrate
Interpret and
Categorize
Reconcile and
Analytics
Visualizatio
n
Classify
Extract content from
wide-variety of
document types
Decompose
documents, discover
taxonomy, normalize
taxonomy
Train to interpret in
specific business
context and extract
targeted data for
analytics
Apply data mapping,
business rules,
calculations, models,
and user driven learning
• Yes ML
• Format detection
• Pixel correction
• Character
recognition
• Linguistics
correction
• Numeric
correction
• Yes ML
• Machine learns
from exception
management
performed by
humans
• Yes ML
• Train the
machine to
interpret based
on business
context not rules
• Connect the
information for
the same
provision across
documents
• No ML
• RAGE
configurable
connector factory
is used to rapidly,
non-intrusively
integrate with
hundreds of data
source (SAP,
CRM, TMS,
Legacy etc.).
• Yes ML
• Auto-discover
document
structure, key
provisions, tables
• Auto-discover
key concepts,
and relationships
• Assisted ML to
finalize taxonomy
and target output
Connect with a variety
enterprise/legacy
systems just via
configuration
Customizable user
interface developed
just via configuration
• No ML
• Rapidly configure
custom UI to display
right charts, visuals.
• Can be customized
by users
• Can be changed
very rapidly as the
needs change
Information flow
Machine learningOutput
• Very little to no IT
time needed
• Extract clean
content from
heterogeneous
quality and variety
(PDF types, images)
of documents
• Entire document is
read
• Provisions are
classified based on
language/concept
relationship not key
words and positions
• High accuracy of
content
categorization as
the search is
business context
(e.g. Kroger) driven
• Human based
exception
management
declines
dramatically.
• Custom user
interface to deliver
specific insights that
can be changed
rapidly without
coding
23. © Rage Frameworks Inc, 2016. All rights reserved.
RAGE AI Classification and Categorization Process
Assisted deep learning is deployed for taxonomy creation
…it’s possible
Load Document
Auto Discovery
Filter the Auto
Discovered
Output
Build
Ontology (SI
App)
Upload
Document
not seen by
the system
Execute Contract
Review Process
Output Not
Extracted
Output
Extracted
False Positive
Partial Match
[Low confidence
score]
Accurate
Extraction
Validate the
Output
Document
Decompositions
24. © Rage Frameworks Inc, 2016. All rights reserved.
Classification Process – Document Decomposition
Machine learning automatically identifies document hierarchy and relationships
…it’s possible
PDF Contract Agreement
Domain Discourse Model
Document decomposition helps identify sections, sub-
sections and their relationship with each other
25. © Rage Frameworks Inc, 2016. All rights reserved.
Classification Process – Auto Discovery
Example to discover and related content from tables (e.g. Schedule A and Invoices)
…it’s possible
The engine parses the entire table
content even though there are
multiple variations within a single
table and treat each one of them
separately. The variations are as
follows:
Route Information
Mileage Information
Drop Information
Fees Information
Total
1 2
3
4
5
Document Type: Invoice
1
2
3
4
5
26. Enabling the Intelligent Enterprise
Interpretation with Natural Language Understanding
Real Time Intelligence
Fund Managers/Competitive & Market Intelligence/Customer/Supplier Risk
© Rage Frameworks Inc, 2016. All rights reserved.
27. | 27
RTI systematically interprets and analyzes all publicly and privately available content in
the context of a company, an industry and macro environment, to generate RTI Signal
Heatmaps draw attention to securities with the most change in
their cumulative signal strength highlighting the overall impact
on a company from the market developments around it
For each company, the RTI Signal can be further broken down
by specific business drivers that may be impacting a company
RTI Signal leads the stock price for 30 – 40% of the companies
in RAGE portfolio (Coverage over 8000 companies)
For each company, the cumulative RTI Signal can be tracked
over time with key triggers by date
4. alpha – RTI vs Stock price 3. Company view over time
1. Portfolio View 2. Company view by business drivers
Stock Price (Log)
RTI Cumulative Score
1.66
1.68
1.7
1.72
1.74
1.76
1.78
1.8
1.82
1.84
-‐1.5
-‐1
-‐0.5
0
0.5
1
1.5
2
04/01 04/22 05/13 06/03 06/24 07/15 08/05 08/26 09/16 10/07 10/28 11/18
RTI Stock
Price
(Log)
28. RTI is not a black box: Drill down into the business drivers to see specific content
pertaining to that driver deemed relevant by the RAGE Semantics Engine
| 28
5
Expand the Factors to drill down
into content pertaining to that
factor
29. 29
Impact Network – Wal-Mart Stores, Inc. (WMT) Plans To Unseat Amazon.com, Inc.
(AMZN) Prime (Topic: Expansion and Closure;; Score: 0.3)
1st
Order
Effect
http://learnbonds.com/118763/wal-mart-stores-inc-wmt-plans-to-unseat-amazon-com-inc-amzn-prime/118763/
Topic: Expansion and Closure
Driver: Product Launch
Sector: Retail
Primary Impact: Low Medium Positive
RAGE
SI
Engine
S
1
S
2
S
3
S
4
S
5
Impact Network
[Deep Semantic Interpretational Map]
S
6
S
7
S1: Wal-Mart Stores, Inc.
(NYSE:WMT) plans to rival
Amazon.com, Inc.
(NASDAQ:AMZN) with the launch
of a new delivery system that
costs less.
30. 30
Real Time Intelligence
RTI Signal Leads Stock Price - Wal-Mart Stores, Inc. [WMT.N]
©Rage Frameworks Inc, 2016. All rights reserved.
Business Driver - Same
Store Sales
Jan 7th, 2015 - RetailNext -
Foot traffic dropped 8.3
percent during November and
December versus a year ago at
the specialty stores and large
retailers .
0
5
10
15
20
25
30
35
55
60
65
70
75
80
85
90
95
Alpha Signal Rating
Stock Price
Business Driver – Consumer
Confidence
Oct 15th, 2015 – Bloomberg.com
- Improving views of personal
finances signal the turmoil in
financial markets and slowdown
in hiring is not affecting
consumer psyches,
which bodes well for sustained
gains in consumer spending.
Business Driver - Expansion
July 22, 2015 –
Supermarketnews.com
The new 1.2-million-square-foot
center is part of a "next-
generation" network to support
Walmart's rapidly growing e-
commerce business. It features
state-of-the-art automation and
warehousing systems.
Business Driver – Retail Sales
Jan 20th, 2016 – economywatch.com
Americans spent $626.1 billion in the
holiday season, representing a 3.7
percent increase on a year-over-year
basis when including online sales.
Signal
Stock
Price
32. © Rage Frameworks Inc, 2016. All rights reserved. | 32
AI in the Enterprise
Machine Intelligence Acquisition: Method Fit
…it’s possible
Source: The Intelligent Enterprise in the BigData Era, Srinivasan, Wiley, 2016
n How important is it to start with a high level of accuracy [precision
and recall]? How expensive is a mistake? Both false positive and
false negative.
n How much variability is there in the underlying phenomenon and
therefore data? The larger the variability like unstructured text, the
training sample needs to be extremely large to get reasonable results
n Can you live with a black box? Do you need transparency in the
engine’s reasoning? Do you need to trace its reasoning so you can
understand ‘causality’?
n Random Forests [Breiman] and Natural Language Understanding
[RAGE AI™] are traceable methods. High levels of variability and/or
high cost of mistakes strongly imply traceable and transparent
methods.
33. © Rage Frameworks Inc, 2016. All rights reserved. | 33
Summary …it’s possible
n AI seems to be back in full force and this time getting integrated into the
mainstream
n Big Data. The ability to analyze entire populations vs samples has allowed
assumption-free algorithmic approaches to flourish vs the traditional ‘data
model’. We are letting the data tell us the story vs assuming prior behavior of
data;; but key challenges wrt text are context, language and traceability
n Deep learning with deep linguistic parsing in context will allow us to create
‘natural language understanding’ in machines vs just ‘natural language
processing’
n AI vs Machine Intelligence. AI = Automation including knowledge-based tasks.
Machine Intelligence = embedding intelligence and learning from data and
experts continuously to enable AI.
n With all these advances, enterprise business architecture will change
dramatically. Execution will be largely thru Intelligent Machines. Design will be
machine informed. The rate of change in the role of humans will accelerate.
Source: The Intelligent Enterprise in the Era of Big Data, Srinivasan, Wiley, 2016
34. © Rage Frameworks Inc, 2016. All rights reserved.
Enabling the Intelligent Enterprise
AI in the Enterprise
The Hive Think Tank
Jan 26, 2017