Shaping an AI-Driven Future with
Augmented Intelligence for
Enterprises
Patrice Neff, Squirro – The Insights Company
2Squirro – © 2017 Nektoon AG
The Importance
of Data
Augmented
Intelligence
3Squirro – © 2018 Nektoon AG
WHO AM I?
Speaker’s Introduction
Technologist and entrepreneur with a passion for building user-
friendly applications. Also a father, runner, cyclist and language
nerd.
Patrice Neff
Founder
4Squirro – © 2018 Nektoon AG
MEET SQUIRRO
The Augmented Intelligence Solution
A leading AI-platform
A self-learning system keeping you in the
know and recommending what’s next
AI-driven Apps
Customer Insights
Financial Services Corporate Insurance Other industries
Service Insights
Support
Cognitive Search
Investors
Company
Augmented Intelligence solutions provider
Global presence in Europe, US and Asia
Partners
Global Network of over 20 certified partners
IB CIB AM CI REI
IT-
SM
CS
CI
Squirro Applications
7Squirro – © 2018 Nektoon AG
Bankruptcy Topic Detection Bankruptcy Events are defined as any event related to a company either going bankrupt or an update on
their bankruptcy proceedings
Debt Topic Detection Debt Events are defined as any event related to issuing bonds, loans, or any other debt instrument.
Divestiture Topic Detection Divestiture Events are defined as any event related to a company divesting a portion of its current
operations or a subsidiary
Earnings Topic Detection Earnings Events are defined as any event related to quarterly earnings releases, forecasts, or analyst
estimates about a companies financial performance
Equity Topic Detection Equity Events are defined as any event related to issuing stock, options, or any other equity instrument.
Expansion Topic Detection Expansion refers to the launch of a new business unit, opening a new location, or entering a new market.
IPO Topic Detection IPO Events are defined as any event related to Initial public offerings of stock in a company
M&A Topic Detection M&A Events are defined as any event related to a company merging with, or buying at least a portion of
another company.
Management Change Detection Management Changes are defined as any announcement of changes in employment or leadership at a
company, organization, or group.
Forward Interest Detection Statement about something in the future. Not speculation about things outside of the speaker's control.
Has to be linked to an action that the speaker or their group will take directly
Models
The Importance of (Training) Data
9Squirro – © 2018 Nektoon AG
PROGRESSION OF CLASSIFICATION MODELS
Input
data
Predict
Classify
Recommend
Trigger Automation
Engine action
Ad-hoc
search
From simpler to more complex methods
Naive
Bayes
SVM Random
Forest
Deep
Learning
Smart
Filters
Squirro Machine Learning Service
Ample data
Strong models
SquirroMLS
Enterprise learning cycle
Little data
No models
Models
Libraries
Custom
Models
10Squirro – © 2018 Nektoon AG
IMMEDIATE OR ACCURATE?
Bag of Words / Smart Filters
Less accurate long term
Faster learner
SVM, etc.
More accurate long term
Slower learner
Time
Accuracy
Time
Accuracy
More Data
12Squirro – © 2018 Nektoon AG
EXHIBIT 1
Unsupervised Word Sense Disambiguation Rivaling Supervised Methods, 1995
Plant
Photo Source: Wikipedia
13Squirro – © 2018 Nektoon AG
EXHIBIT 1
Unsupervised Word Sense Disambiguation Rivaling Supervised Methods, 1995
In this current work, the one-sense-per-
discourse hypothesis was tested on a
set of 37,232 examples (hand-tagged
over a period of 3 years)
14Squirro – © 2018 Nektoon AG
EXHIBIT 1
Unsupervised Word Sense Disambiguation Rivaling Supervised Methods, 1995
More impressively it achieves nearly the
same performance as the supervised
algorithm given identical training contexts
vs and in some cases actually achieves
superior performance.
15Squirro – © 2018 Nektoon AG
EXHIBIT 2
Scaling to Very Very Large Corpora for Natural Language Disambiguation, 2001
We collected a 1-billion-word training
corpus from a variety of English texts.
We show learning curves for each
learner.
The curves appear to be log-linear
even out to one billion words.
16Squirro – © 2018 Nektoon AG
EXHIBIT 3
Scene Completion Using Millions of Photographs, 2007
It takes a large amount of data for our method
to succeed. We saw dramatic improvement
when moving from ten thousand to two million
images. But two million is still a tiny fraction of
the high quality photographs available on sites
like Picasa or Flickr.
17Squirro – © 2018 Nektoon AG
EXHIBIT 4
Revisiting the Unreasonable Effectiveness of Data, 2017
Performance increases increases
logarithmically based on volume
of training data. We find there is
a logarithmic relationship
between performance on vision
tasks and the amount of training
data used for representation
learning.
18Squirro – © 2018 Nektoon AG
EXHIBIT 5
Resource-Size matters: Improving Neural Named Entity Recognition with Optimized Large
Corpora, 2018
We report evidence that the inferior quality quality of
German text data and its small size are the major
reasons for the observed lack of progress. To tackle
this problem, we use a larger corpus for training
Training effort
20Squirro – © 2018 Nektoon AG
IMMEDIATE OR ACCURATE?
Bag of Words / Smart Filters
Less accurate long term
Faster learner
SVM, etc.
More accurate long term
Slower learner
Time
Accuracy
Time
Accuracy
21Squirro – © 2018 Nektoon AG
FLEXIBLY CHOOSE THE BEST OPTION
Time
Accuracy
FROM SMART FILTERS…
… TO MODEL TRAINING
Augmented Intelligence
Pragmatic AI
25Squirro – © 2018 Nektoon AG
PRAGMATIC: APPLYING THE 80-20 RULE
Task
complexity
Model complexity
Science fiction
Trivial
80%
80%
The 80%
Zone
26Squirro – © 2018 Nektoon AG
PRAGMATIC AI IN REAL WORLD
Taskcomplexity
Model complexity
Science fiction
Artificial General
Intelligence
Lead recommendation
based on statistics
Lead recommendation
based on support vector
machine model
Chatbot with predefined
responses
Generic virtual assistant
(generated responses)
Automatically discover
main topics in a new
dataset
Self-driving car
Trivial
27Squirro – © 2018 Nektoon AG
HUMAN INTELLIGENCE
Excels at causation
28Squirro – © 2018 Nektoon AG
AUGMENTED INTELLIGENCE
Combine causation with correlation
29Squirro – © 2018 Nektoon AG
Pragmatic or
Practical AI
Implementing Artificial
Intelligence in ways that can be
easily deployed today, with real
data available inside an
enterprise
30Squirro – © 2018 Nektoon AG
SQUIRRO AUGMENTED INTELLIGENCE STACK
Vertical applications for rapid results
Cloud / On-Premise Multi-Instance / Scalable
Corporate
Financial
Services
All verticals All verticalsCorporate
Insurance
Other verticals
IB
CIB
REF
REI
CI IT Service
Management
Customer
Service
Management
SquirroAIPlatform
Admin / Security
IAM
Internal data sources External data sources
Enrich Identify Discover Predict Recommend Automate VisualizeLoad
Gather ActAnalyze
Other industries
Manufacturing
Customer Insights Product Suite Service Insights Product Suite Cognitive Search Product Suite
Enterprise
Search
Web Portal
Search
SquirroApplications
31Squirro – © 2018 Nektoon AG
Thank you!

Big Data LDN 2018: SHAPING AN AI-DRIVEN FUTURE WITH AUGMENTED INTELLIGENCE FOR ENTERPRISES

  • 1.
    Shaping an AI-DrivenFuture with Augmented Intelligence for Enterprises Patrice Neff, Squirro – The Insights Company
  • 2.
    2Squirro – ©2017 Nektoon AG The Importance of Data Augmented Intelligence
  • 3.
    3Squirro – ©2018 Nektoon AG WHO AM I? Speaker’s Introduction Technologist and entrepreneur with a passion for building user- friendly applications. Also a father, runner, cyclist and language nerd. Patrice Neff Founder
  • 4.
    4Squirro – ©2018 Nektoon AG MEET SQUIRRO The Augmented Intelligence Solution A leading AI-platform A self-learning system keeping you in the know and recommending what’s next AI-driven Apps Customer Insights Financial Services Corporate Insurance Other industries Service Insights Support Cognitive Search Investors Company Augmented Intelligence solutions provider Global presence in Europe, US and Asia Partners Global Network of over 20 certified partners IB CIB AM CI REI IT- SM CS CI
  • 5.
  • 7.
    7Squirro – ©2018 Nektoon AG Bankruptcy Topic Detection Bankruptcy Events are defined as any event related to a company either going bankrupt or an update on their bankruptcy proceedings Debt Topic Detection Debt Events are defined as any event related to issuing bonds, loans, or any other debt instrument. Divestiture Topic Detection Divestiture Events are defined as any event related to a company divesting a portion of its current operations or a subsidiary Earnings Topic Detection Earnings Events are defined as any event related to quarterly earnings releases, forecasts, or analyst estimates about a companies financial performance Equity Topic Detection Equity Events are defined as any event related to issuing stock, options, or any other equity instrument. Expansion Topic Detection Expansion refers to the launch of a new business unit, opening a new location, or entering a new market. IPO Topic Detection IPO Events are defined as any event related to Initial public offerings of stock in a company M&A Topic Detection M&A Events are defined as any event related to a company merging with, or buying at least a portion of another company. Management Change Detection Management Changes are defined as any announcement of changes in employment or leadership at a company, organization, or group. Forward Interest Detection Statement about something in the future. Not speculation about things outside of the speaker's control. Has to be linked to an action that the speaker or their group will take directly Models
  • 8.
    The Importance of(Training) Data
  • 9.
    9Squirro – ©2018 Nektoon AG PROGRESSION OF CLASSIFICATION MODELS Input data Predict Classify Recommend Trigger Automation Engine action Ad-hoc search From simpler to more complex methods Naive Bayes SVM Random Forest Deep Learning Smart Filters Squirro Machine Learning Service Ample data Strong models SquirroMLS Enterprise learning cycle Little data No models Models Libraries Custom Models
  • 10.
    10Squirro – ©2018 Nektoon AG IMMEDIATE OR ACCURATE? Bag of Words / Smart Filters Less accurate long term Faster learner SVM, etc. More accurate long term Slower learner Time Accuracy Time Accuracy
  • 11.
  • 12.
    12Squirro – ©2018 Nektoon AG EXHIBIT 1 Unsupervised Word Sense Disambiguation Rivaling Supervised Methods, 1995 Plant Photo Source: Wikipedia
  • 13.
    13Squirro – ©2018 Nektoon AG EXHIBIT 1 Unsupervised Word Sense Disambiguation Rivaling Supervised Methods, 1995 In this current work, the one-sense-per- discourse hypothesis was tested on a set of 37,232 examples (hand-tagged over a period of 3 years)
  • 14.
    14Squirro – ©2018 Nektoon AG EXHIBIT 1 Unsupervised Word Sense Disambiguation Rivaling Supervised Methods, 1995 More impressively it achieves nearly the same performance as the supervised algorithm given identical training contexts vs and in some cases actually achieves superior performance.
  • 15.
    15Squirro – ©2018 Nektoon AG EXHIBIT 2 Scaling to Very Very Large Corpora for Natural Language Disambiguation, 2001 We collected a 1-billion-word training corpus from a variety of English texts. We show learning curves for each learner. The curves appear to be log-linear even out to one billion words.
  • 16.
    16Squirro – ©2018 Nektoon AG EXHIBIT 3 Scene Completion Using Millions of Photographs, 2007 It takes a large amount of data for our method to succeed. We saw dramatic improvement when moving from ten thousand to two million images. But two million is still a tiny fraction of the high quality photographs available on sites like Picasa or Flickr.
  • 17.
    17Squirro – ©2018 Nektoon AG EXHIBIT 4 Revisiting the Unreasonable Effectiveness of Data, 2017 Performance increases increases logarithmically based on volume of training data. We find there is a logarithmic relationship between performance on vision tasks and the amount of training data used for representation learning.
  • 18.
    18Squirro – ©2018 Nektoon AG EXHIBIT 5 Resource-Size matters: Improving Neural Named Entity Recognition with Optimized Large Corpora, 2018 We report evidence that the inferior quality quality of German text data and its small size are the major reasons for the observed lack of progress. To tackle this problem, we use a larger corpus for training
  • 19.
  • 20.
    20Squirro – ©2018 Nektoon AG IMMEDIATE OR ACCURATE? Bag of Words / Smart Filters Less accurate long term Faster learner SVM, etc. More accurate long term Slower learner Time Accuracy Time Accuracy
  • 21.
    21Squirro – ©2018 Nektoon AG FLEXIBLY CHOOSE THE BEST OPTION Time Accuracy
  • 22.
  • 23.
    … TO MODELTRAINING
  • 24.
  • 25.
    25Squirro – ©2018 Nektoon AG PRAGMATIC: APPLYING THE 80-20 RULE Task complexity Model complexity Science fiction Trivial 80% 80% The 80% Zone
  • 26.
    26Squirro – ©2018 Nektoon AG PRAGMATIC AI IN REAL WORLD Taskcomplexity Model complexity Science fiction Artificial General Intelligence Lead recommendation based on statistics Lead recommendation based on support vector machine model Chatbot with predefined responses Generic virtual assistant (generated responses) Automatically discover main topics in a new dataset Self-driving car Trivial
  • 27.
    27Squirro – ©2018 Nektoon AG HUMAN INTELLIGENCE Excels at causation
  • 28.
    28Squirro – ©2018 Nektoon AG AUGMENTED INTELLIGENCE Combine causation with correlation
  • 29.
    29Squirro – ©2018 Nektoon AG Pragmatic or Practical AI Implementing Artificial Intelligence in ways that can be easily deployed today, with real data available inside an enterprise
  • 30.
    30Squirro – ©2018 Nektoon AG SQUIRRO AUGMENTED INTELLIGENCE STACK Vertical applications for rapid results Cloud / On-Premise Multi-Instance / Scalable Corporate Financial Services All verticals All verticalsCorporate Insurance Other verticals IB CIB REF REI CI IT Service Management Customer Service Management SquirroAIPlatform Admin / Security IAM Internal data sources External data sources Enrich Identify Discover Predict Recommend Automate VisualizeLoad Gather ActAnalyze Other industries Manufacturing Customer Insights Product Suite Service Insights Product Suite Cognitive Search Product Suite Enterprise Search Web Portal Search SquirroApplications
  • 31.
    31Squirro – ©2018 Nektoon AG Thank you!