Why traditional ECM is a thing of the past and what it means to you Terms amp Definitions Data Big and Small Data Structured amp Unstructured Content AI Machine Learning ML Neural Networks ID: 928785
Download Presentation The PPT/PDF document "Don’t Do IM Without AI" is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
Slide1
Don’t Do IM Without AI
Why traditional ECM is a thing of the past – and what it means to you
Slide2Terms & Definitions
Data
Big and Small Data
Structured & Unstructured (Content)
AI
Machine Learning (ML)
Neural Networks
Natural Language Processing (NLP)
Covalent Neural Networks (CNN)
Algorithms
Information Management
Information Architecture
Records Management
Electronic Content Management
Slide3The Old & New Worlds of ECM
The Changing Landscape of ECM
Slide4The Old World
Business Transformation
High-value processes, infrequently
On-premise first
Little or no use of big data
Human intelligence only
Enterprise-wide, no personalization
ECM Transformation
Monolithic ECM solutions
Multiple content silos
Semi-regular migrations
Irregular cleanups and classifications
Human-managed IA structures
Human-managed classification
Slide5The New World
Business Transformation
Every process, all the time
Cloud- or hybrid-first
(Big) data-driven decision making
Artificial Intelligence AND human intelligence
Enterprise-wide AND
personalizable
ECM Transformation
Cross-platform IM services
Few or no migrations
Regular (daily!) cleanups and classifications (AI “virtual assistant”)
Human- and AI-managed IA structures and classification
Content Services
Intelligent Information Management (IIM)
Slide6Business Transformation
Cloud- or hybrid-first
Big data-driven IM practices
Artificial Intelligence combined with Human expertise
Contextual = enterprise- or even industry-wide standards, while allowing personalization to team or individual level
Constant innovation
Slide7Content Services
Forrester and Gartner concept and endorsed by Microsoft
Content is an
active participant
in daily decision making
Content needs to be
actively managed
as a
key asset
.
Business processes + applications, platforms, and components
No more monolithic ECM “solutions”
Modular capabilities tied to needs of specific business processes
Slide8Intelligent Information Management (IIM)
AIIM agrees with Content Services concept, but goes further
IIM requires an organization to do the following:
Modernize the IM toolkit
Digitize core organizational process
Automate compliance & governance
Leverage analytics and machine learning
Slide9IIM – Modernize IM Toolkit
Modernize the IM toolkit
Cloud content management
Internal and external collaboration platforms
Low- or no-code “self service” development by business users
Content integration & migration tools
Slide10IIM – Digitize Core Processes
Digitize core organizational process
Robotic process automation
Business process management
Multi-channel intelligent capture
High-volume process optimization
Slide11IIM – Automate Compliance & Governance
Automate compliance & governance
Records management & digital preservation
eDiscovery & Legal Hold
Industry & Geographic applications
Organization-specific applications
Blockchain (for provenance)
Slide12IIM – Leverage AI and ML
Leverage analytics and machine learning
AI, content analytics & semantics
Data recognition, extraction & standardization
Metadata & taxonomy management
Document classification & PII identification
All of this relies on Big Data
Slide13AI Basics
High Level Intro to the AI landscape, neural networks, limitations and possibilities
Slide14The Numbers Game
10,000,000,000 MB of new information created every second*
$26 to $39B spent on AI in 2016 (McKinsey)
AI will generate $1.2T of business value in 2018 (Gartner)
Business execs expect to spend 300% more on cognitive computing in 2018 (Forrester survey)
75% of executives say AI will be implemented in their companies by 2020/2021 (the Economist Intelligence Unit)
AI will generate $3.9T of business value by 2022 (Gartner)
* Source: Weforum.org/agenda/2017/01/forget-
ai
-real-revolution-
ia
Slide15AI Landscape
AI is very broad and relatively old field of computing (40s/50s +)
“What is AI?”
Widespread debate about what is or isn’t AI
Minksy
& McCarthy: Any task performed by program or machine that, if a human carried out the same activity, we would say the human had to apply intelligence to accomplish the task
General AI is the kind we see in the movies (HAL)
Real world AI mostly narrow: intelligence organized around fulfilling very specific tasks
Machine Learning AI techniques
Uses math and code learn how to perform a function, and get better at it over time
Most AI approaches these days focus on Machine Learning
Supervised (apply human IM expertise and validate the ML findings)
Unsupervised or semi-supervised (surface AI insights that wouldn’t be found otherwise)
Yann
LeCun
, head of AI research at Facebook: Supervised learning is frosting on the cake, unsupervised learning is the cake itself.
Other AI techniques
Deep Learning: ML that can draw conclusions similar to how a human does, using an artificial neural network
Reinforcement learning: Programmer specifies current state and goal, lists allowable actions, describes environmental constraints – and then the AI figures out how to achieve the goal.
Transfer learning: Let AI figure out how to apply its knowledge to similar problems without explicit programming
Slide16Early Adopters
Potential value of AI ranges by industry sector (McKinsey)
Retail and transport/logistics highest potential value
Aerospace and defence lowest potential value
Characteristics of early AI adopters
Digitally Mature
Larger businesses
Adopt AI in core activities
Adopt multiple technologies
Focus on growth over savings
Executive support for AI initiatives
Slide17Example AI Applications
Input X
Output Y
Application
Car location and speed data using lidar and GPS
Traffic flow patterns
Tesla self-driving car
Voice queries on your mobile phone
Answers and recommendations
Apple Siri
Voice queries on your Windows laptop
Performs computing actions
Windows 10 Cortana
Voice queries on your smart speaker
Answers and recommendations, personal shopping, control smart home devices
Amazon Alexa
Uploaded images
Image tagging your friends
Facebook
Credit card transaction
Purchase approval or alert
Fraud detection
Slide18Artificial Intelligence Techniques
Slide19Conventional Computing vs Neural Networks
Conventional computing is programming by humans
Generally follows an
input -> algorithm -> output
approach
But, the human has to explicitly define all the business rules
In neural networks, the computer “learns” the business rules from data
Neural networks are good at pattern matching, but not math and logic.
Slide20Neural Networks 101
Modelled after neurons in human brains
Individual neuron:
If (Inputs x weights, + bias) = activation function (1 or 0) -> output
Neural Network
Composed of lots of neurons in layers
Inputs -> Hidden layer -> Output
Neurons can be inputs to neurons
Deep networks have > 1 Hidden Layer
Networks can also be taught to “remember” or “forget” inputs & outputs
Training teaches the AI what a “correct” output should have been.
Heavy processing loads – often GPUs (computer gaming graphics cards) are used.
Slide21What AI can and can’t do
AI
can
…
save enormous human time and effort
generate insights humans don’t expect
evolve
AI
can’t
do everything a human can do, but a human can do everything AI can do (but not at scale!)
operate without lots of clean data
perform all the tasks in a business process
be perfect, or magic
Labrador puppy, or chicken nugget?
Slide22Human Involvement
Human involvement is needed!
AI should be used only where it makes sense
Consider using humans as the primary actor and AI as the virtual assistant
High failure rate in AI work – take a hypothesis-based approach
Slide23Interpretability
Can AI understand us? (English is not optimized for machine understanding)
Sometimes AI needs to explain itself to humans (“how did it arrive at that conclusion?”)
Can we understand their results?
Facebook
chatbot example:
Slide24The Impact of AI on IM
The Implications of using Artificial Intelligence for Information Management
Slide25Overall Benefits
Overall Benefits for Information Management
Analyze large unstructured file collections in any location
Auto-classification
Reduce cost and duration of IM projects
Improve ongoing compliance
Recognize user intent
Tireless
Infinitely scalable
Will eventually surpass any individual IM expertise?
Slide26IA Benefits
Benefits for Information Architecture
Analyze large unstructured file collections
File de-duplication
Automatic file synopsis for search indexing and excerpts
Relationship extraction
Propose information architecture and taxonomy based on actual file format and content
Metadata management including auto-classification, and classification accuracy rates
Search results optimization
Governance & compliance reporting
(Eventually) allow individual personalization (“AI folksonomy”)
Slide27RM Benefits
Benefits for Records Management
File de-duplication
Auto-classification based on machine learnings and your file plan
Rapid eDiscovery for audits and legal holds
Relationship extraction (ex. for cases, e-discovery)
Machine translation
Optical character recognition (ex. for physical records, e-discovery)
Improved data quality: Reduce human effort and errors in common RM tasks
User personalization (“AI folksonomy”) – users don’t need to fully understand file plan to be compliant!!!
Ongoing / “instant” compliance
Governance & compliance reporting
Slide28Practical Considerations
“What does this mean for us?”
Slide29Business Considerations
Leveraging AI for Information Management
Strategic Impact
Digital Transformation
Increase business productivity and knowledge
Increase competitiveness
Operational Impact
Impact to day-to-day business operations and processes
Data silos
Governance and compliance requirements
Financial Impact
Significant up-front investment
(
expensive and time-consuming)
Significant ongoing operational costs
Technology Impact
Technology decisions, data management & analysis, AI training/testing/publishing, application development
Executive sponsorship and governance is
required
Slide30Helping AI Derive Meaning
“Garbage in, garbage out”
If a human cannot understand the basics of their data, how can an AI?
AI needs to understand
Organization and application context (“Which industry”, “what does this application do”, “what business process is this for”…)
Domain entities (“department”, “customer”, “widget foo”, “process bar”, “PII report”…)
What is generally being asked for (types of questions, potential valid and invalid responses, search history)
Ontology (information hierarchies, controlled vocabularies, taxonomies)
Therefore:
a good information architecture is prerequisite to AI training
Slide31Technical Considerations
Enormous amounts of processing power required = GPUs
Cloud hosting is therefore beneficial – scale up and down
Pick an AI platform that suits your business purpose:
Azure Cognitive Services (cloud only)
AWS Machine Learning or hosted AI toolkits (cloud or on-premise)
Google TensorFlow (cloud or on-premise)
Others (cloud or on-premise?)
You need large amounts of data storage
Bandwidth considerations uploading and downloading data?
Data sovereignty and encryption requirements (see next slide)
Slide32Data Prep and Classification Considerations
Data Sources
On-premise, cloud, or migrated/integrated (hybrid)
File shares - train in place, or duplicate content into “data lake” for AI?
Non file-shares (aka SharePoint)– how to extract the data sets from their source so it can be used by AI?
Batch or streaming: How / when is new content added? Will this be included in analysis?
Where does the data live? Where will it be analyzed? What are the security or privacy implications?
Data Prep
(60 – 70% of AI project is likely to be data prep)
Remove junk files (.exes, scripts,
etc
).
De-duplicate
What kind of file structure is needed? (Sub-folders, “buckets”, metadata files, manifests)
Slide33Testing Considerations
What are the inputs we are testing for? What is the output we want?
Test, test, test!
Desired accuracy rate
“Trained” or “Supervised” ML needs to be trained and then tested
“Untrained” or “Unsupervised” ML results need to be validated
Test data
must
be different from training data
Create a test harness for determining accuracy & validity
Manual or automated testing?
Test coverage and depth
Test data generator applications, plus Excel spreadsheets for testing
Experimental – take hypothesis-based approach
Slide34RM Considerations
Importance of information management for classification
How to handle record classification
Precondition and Post condition
Iterations
Training and testing data
Slide35IA Considerations
Importance of information management for classification
Precondition and Post condition
Iterations
Training and testing data
Slide36Introducing THEMIS
The Intelligent Information Management Suite
Slide37THEMIS IA
Slide38The Central Role of AI in THEMIS
Imagine a world where you can ask….
…for expert IM advice, 24/7
…for highly accurate risk assessments of your actual compliance gaps
…for a tireless assistant to review and auto-classify every document you have
…for automatic extraction of enterprise taxonomy from your content
…while still allowing personalization down to the individual user (“AI-assisted folksonomy”)
…and this works for any business process, application, or content platform
Slide39Examples of AI Involvement
Slide40How We Approached It
Documented our product goals and objectives
Conducted AI and IM industry research
Identified potential AI “fit”
Planned technical proof of concept: The “art of the possible”
Prepared and organized training and testing files
Created IA for the data files
Focus on limited use case: chatbot with natural language processing and machine learning on IA models we have created for clients
Reviewed the results
Slide41Microsoft AI Platform
We’re a Microsoft Shop, so we assessed their services first
Some examples of what we are using:
Chatbot (Azure Bot Service)
Machine Learning and text recognition (Cognitive Services)
Natural Language Processing (Language Understanding / Luis)
Search suggestions (Bing)
Slide42Future Work
Advanced data pipeline
Extracting testing and training data from content silos including SharePoint
Prep and storage requirements
Data Catalogue and metadata management for data sets
Use AI to aggregate customer IA to produce general and specific models (ex Government of Canada models)
Deep Learning algorithms for unsupervised IM learning
Reinforcement learning for personalized recommendations
Transfer learning to spread IM knowledge amongst all use cases
FAQ with Chat (Microsoft
QnA
Maker)
Slide43Wrap-up
Slide44What’s Next
Goals & Objectives
What are your IM business goals and objectives?
How can you apply AI to them?
Paper exercise or technical proof?
Learn the process? Identify promising approaches? Create the algorithms?
In-house or out-source AI expertise?
Make a business case to obtain executive buy-in
AI Free Trials / Proof of Concepts
Ask for a demo of THEMIS, our AI-driven cloud-hosted Intelligent Information Management Suite
Slide45Thank You
Questions, comments?
john@heluxsystems.com
Ask for a demo of THEMIS, our AI-driven cloud-hosted Intelligent Information Management Suite