Heading in Verdana Regular - Bilişim Zirvesi · Deloitte aiStudio 2 AI in Risk Management Your Speaker David Thogmartin Deloitte. Risk Advisory Leader aiStudio Tel: +49 211 8772

Informatics Summit 2019 Instanbul

Artificial Intelligence: Capabilities and Selected Use Cases

aiStudio

21 November 2019 - David Thogmartin – Deloitte GmbH Risk Advisory aiStudio

Deloitte aiStudio 2

AI in Risk Management

Your Speaker

David Thogmartin

Deloitte. Risk Advisory

Leader aiStudio

Tel: +49 211 8772 2336E-Mail: [email protected]

Deloitte aiStudio 3

A Collection of Technologies

Allowing computers to perform tasks previously reserved for humans

Optimization / Calibration

Prediction / Decisioning

Interaction / Orchestration

Writing / Talking

Association / Interpretation

Reading / Listening

Selection / Categorization

Deloitte aiStudio 4

User Perspective

• Human-like capabilities – perception of the world, comfortable interaction

• Helps us to use the flood of available data

• Not particularly smart, not curious, not easily transferrable

Not general intelligence, but “Narrow AI”AI is a Tool to Helps Us Manage Complexity

Practitioner Perspective

• Trained by data vs instructed by rules... … well suited for complex relationships

• Multiple approaches-a) technologies –b) and algorithms –c) derived from mathematics –d)

• Iteration requires substantial computational power

(d- linear algebra, calculus, stastistics(a- supervised, unsupervised, reinforcement

(c- Nearest neighbor, naïve Bayes, decision trees, linear regression, logistic regression, SVM, ANN

(b- ML, DL, NLP, NLG, CV…

Deloitte aiStudio 5

Understanding and Managing Complexity

To understand the world, we need tools that can handle the complexity

n D2D linear 2D non-linear 3D spacial 4D spacial

Deloitte 2019 6

AI is Important

Not a future technology, it is here, it is now

Largest Spending

• Customer service ($2,9 bn)

• Threat intelligence & prevention ($1,9 bn)

• Sales process recommender systems ($1,7 bn)

• Preventative maintenance ($1,7 bn)

$24 bn

$78 bn

Source: Study 19 Sept 2018International Data Corporation 2018 2022

Worldwide Spend on AI & Cognitive

84% 75%63%

Competitive advantages

New business, new entrants

Required to reduce costs

Source: Stanford University AI Index 2018

Fastest Growth

• Pharmaceutical – Research (46,8% CAGR)

• Product recommender systems (46,5%)

• Enterprise digital assistants (45,1%)

• Cognitive robotic process automation (43,6%)

Big Spenders: Banking & Retail at $4 bn each

Deloitte aiStudio 7

The Conditions for AI are Right

Trillion-fold increase in computing power

(a- FLOPS = floating (point) operations per second

1 xCray-1 Supercomputer

350 MFLOPS-a)

10 x

(b- ZB = Zetabytes=1,000,000,000,000,000,000,000 bytes = 1 trillion GB

Apple 1 (Series 1)

4 BFLOPS

=

Meagre CPU-based Computing

Isolated Machines / Networks

Limited (accessible) Data (0,1 ZB-b) 2006)

Artificial Neural Networks

Rule-Based (Theories)

Static

Academic

Future

Powerful CPU + GPU-based Computing

Internet & Cloud

Abundant Data (45 ZB 2019)

Deep Learning Neural Networks

Data Driven (Facts)

Evolving

Economic

Present

Deloitte aiStudio 8

The aiStudio

An international coalition to lead in the field of AI

Talent

+

Focus

+

Creativity

+

Collaboration

International mix of bright minds across Deloitte

firms

Specific use cases, dynamic, application vs

research

Critical mass + building on

collective successes

Own location, start-up vibe, fresh ideas, agile sprints

Deloitte aiStudio 9

3 in development

The aiStudio’s Ambitious Agenda

On track to deliver 9 applications within first year

6 MVPs in 6 Months

DNAi• 15.2.2019

• ML rapid

prototyping

• Proprietary tool

to build tools

Guardian• 15.2.2019

• Detects bias in

ML models

• AI/Ethics

RoadRunner• 15.7.2019

• Assesses

condition of

infastructure

• Image detection

Lucid [ML]• 30.8.2019

• Explains ML

models

• Whiteboxing

Consistency• 15.9.2019

• Detects anomalies

in data (time series)

• Unsupervised

learnng

IntelliMap• TBD

• Suggests match

between data tables

• Data lineage

Green Label• TBD

• Automated quality

control of human

labeled data

• Quality

CashCast• TBD

• Cash flow forecasting

on complex data

• Time-series prediction

TableMiner• 15.2.2019

• Extracts tables

from pdfs

• Mutliple ANN

Deloitte aiStudio 10

Turbo-Charging Analytics with AI

Text Mining

Typical Challenges

• Quality of Documents (Inputs)

• Multiple Languages / Writing Styles

• Rigidity of RegEx Key-Word-Search

• Many irrelevant hits – or related passages overlooked

Text Mining

Speed Quality Cost

Solution Approaches

• Stemming – simplification of words into root concepts – more robust against varying styles

• Tokenization – partitioning a text into components (words or n-gram sequences)

• Vectorization – representation of texts in tabular data form, suited for machine learning analysis & combination with other data



Table Extraction

Typical Challenges

• Valuable employee time spent on the menial task of data transfer

• Manual data transfer often error prone, requiring four-eyes checks and other time-consuming controls

• Tight deadlines leave little time for analysis

Speed Quality Cost

Solution Approaches

• Automatic recognition of tables within PDF documents – also multiple tables per page

• Extract tables verbatim from documents into spreadsheets for later analysis / processing

• Multiple document types: e-doc to “dirty scans”

Table Extraction



Enrichment & Model Design

Typical Challenges

• Labor-intensive preparation - quality check and enrichment of data to produce useful features

• Intuition to find optimal ML algorithm

• Slow, iterative optimization… leaving little time for analysis, testing, piloting of results

Enrichment & Model Design

Speed Quality Cost

Solution Approaches

• Rapid Prototyping – automating preparatory / tuning steps to minimize iteration cycle times

• Selection / parametrization of the ML-algorithm to achieve best predictive power

• Employing the most modern and effective methods, such as „chromosomal algorithms“



Model Interpretability

Model Interpretability

Speed Quality Cost

Solution Approaches

• Use multiple mathematical techniques to reveal the underlying drivers of sophisticated models

• Quantify their relative importance, enabling construction of simpler surrogate models – with tolerable trade-offs in predictive power

Typical Challenges

• Models built on neural networks are often more accurate, but difficult to understand

• Regulators demand transparency into decision-aiding models

• Since GDPR, consumers have the right to know why discriminated against



Dynamic Visualization

Typical Challenges

• Out-of-date information

• Limited analysis possibilities / interaction and exploration of analysis results

• Reliability / robustness

Speed Quality Cost

Solution Approaches

• Integration of ML-engines / applications to data-sources via API

• Design of real-time data-stream

• Fit-for-purpose and flexible visualization of predictions and driving features (inputs)

Dynamic Visualization



Technology Integration

Typical Challenges

• Valuable employees spend their precious time on menial tasks

• Repetitive tasks cause fatigue, careless work and manual error

Technology Integration

Speed Quality Cost

Solution Approaches

• Apply robotic process automation not only to automate routine tasks, but to trigger advanced AI engines at decision nodes

• ML classifiers to automate decisions based on the same data inputs as the robot’s human counterpart

• Integration of text-mining and table-extraction tools into the process allows for higher degree of automation without manual intervention


Enabling More Effective Risk Management

By improving capabilities to anticipate, measure, and respond to risk

Caught by surprise, scramble in crisis-mode to contain risks

Complexity & urgency leave little time to quantify & understand

Crude measures to half-understood problems can misfire

Time-series

forecasting

with ML

models

Anticipate

Leverage data for early warning indicators – before issues surface

Big-data

mastery,

identified

leading KRI

Measure

Reliance on data & correlation provide a quantitative approach

Rapid &

easily

updatable

modeling

Respond

Prepared responses to analyzed scenarios & interactions


Enabling More Effective Business Management

By improving ability to manage constraints to speed, quality and cost

Results delivered too late to be of practical use

Resources required to deliver make us uncompetitive

High rates of error, misguided decisions on incorrect information

Robotic

process

automation

& NLG

Speed

Insights delivered nearly instantaneously

Deep

Learning

models +

explainable

surrogates

Quality

Greater prediction or classification accuracy, rapidly updatable

Cognitive

process

automation

Cost

Multiplier - augmenting the workforce – flexible and always on


The Spectrum of Use Cases for AI

Multiple technologies at work across a wide variety of specific tasks

CV = computer vision, NLP = natural language processing, NLG = natural language generation, RPA = robotics process automation, DL = deep learning, Unsupervised = statistical methods

Anomaly Detection

Classification / Decisioning

Prediction / Forecasting

ProcessAutomation

• Manufacturing Defects (CV)

• Isolating Non-Standard

Contact Terms (NLP)

• KYC - Customer Data

Validation (Various)

• Trade Surveillance

(Unsupervised)

• Data Quality Check &

Enhancement (Various +

Imputation)

• Fraudulent Transactions, Cyber-

Attacks (Unsupervised)

• Detection of Malignant Tumors

(CV, DL)

• Risk Segmentation for

Underwriting, Reserving (DL)

• IAM - User Entitlements (ML)

• Enhanced Imaging for X-Ray /

other Scans (DL)

• Early Warning / Risk Sensing (NLP)

• Challengers to Classical Models /

Model Calibration (DL)

• Forecasts / Sensitivity Analysis

(Time-Series DL)

• Detect Bias in Data & Models

(Unsupervised)

• Predictive Maintenance (ML + NLP)

• Automated Reports (NLG)

• Avoid Errors /

Inconsistency Data

Consolidation / Controls

(RPA)

• IAM - Facial Scanning,

Surveillance (CV)

• Homomorphic Encryption

(DL)


Sample Use Cases: Banking Industry

Revealing deeper insights from available data to improve decision-making

Underwriting – ML models offer a substantial upgrade over classic scorecard models driving credit decisions, as they make use of wider datasets and recognize complex patterns to achieve greater accuracy in discerning good from poor credit risks.

Defaults – ML models effortlessly blend granting with repayment behavior data to judge likelihood of default, providing timely guidance to appropriate collection strategies as well as to provisioning.

Anti-Financial Crime –multiple unsupervised learning methods spot anomalies in transactions that raise red flags to potential fraud or money laundering activity.

Risk-Based Pricing – by combining multiple data sources and imputing on missing data, ML models can identify client payment obligations vs incomes, providing an advanced means of loan affordability checks

Cross-Sell – by finding inconsistencies between client profile fields and even transaction behavior, ML models can identify outdated client profiles, improving customer service and targeted marketing (mailing) effectiveness

Identity & Access Management – ML offers several approaches to automate the granting of system privileges (training on past mappings) to optimization (role-mining)

License Compliance – as avid users of software tools, banks are under the scrutiny of providers to ensure license compliance. Missing information (software edition) can put the bank at risk of fines. ML can impute missing information with a surprisingly high degree of accuracy.

Challenger Model – validate models used for provisioning, securities, derivatives using approaches completely independent to the original model under evaluation – whether statistical or supervised learning


Sample Use Cases: Insurance Industry

Driving efficiencies for insurers to compete in increasingly commoditized markets

Technical Accounting - Statement of Account are in several formats – read them (NLP) in and standardize them (NLG). Output: either a structured DB or a standard report (NLG). Possibly directly input into system (RPA).

Life Insurance – sensitivity analysis of mega-trends impact to actuarial (life expectancy) tables

Risk Management – optimize the reinsurance program (combination of different policy coverages and their thresholds – quota share, excess of loss, stop loss, …)

Claims – automatically check validity of insurance claims vs policy coverage (whether covered, to what extent)

Data Enrichment (imputation) – predict size / occurrence window of natural catastrophes (infrequent and few data points) with highly frequent weather data (challenging, but potentially highly rewarding)

Asset Management – early warning indicators from several sources (NLP, web-scraping) automatically triggered by a process, in order to trigger decisions for investment managers / update dashboards (example: bond pricing)

Reinsurance Buy – automatically check for consistency between coverage T&C on sold policies (inwards business) vs purchased reinsurance / retrocession (outwards business)

Challenger Model – to text proper implementation of classical actuarial methods – and to identify optimization potential


Sample Use Cases: Audit Analytics

Technology How it Works Benefits

Text Mining / NaturalLanguage Processing

Regular expressions + “bag of words” + lemmatization to search unstructured text in a manner tolerant of inexact matches

Exhaustive audit vs statistical samples. ++ + +

Table ExtractionNeural networks to identify tables, OCR to convert “dirty scans” to readable text. Transfer tables from PDFs to spreadsheets.

Eliminate manual data transfer – slow and error-prone.

+++ + ++

Anomaly DetectionMultiple unsupervised learning (statistical) techniques to isolate data points that are likely to be outliers.

More rigorously derived, risk-based judgmental samples.

+ ++

Automated AssociationAn array of statistical and transformation techniques to shortlist probable duplicate information between tables.

Accelerate the work of mapping tables to produce wider dataset.

+ + ++

Robotics Process Automation

Highly configurable software agents to operate other software, simulating human interaction, orchestrating processes.

Reduce or eliminate manual intervention. +++ ++ ++

Natural Language Generation

Proprietary text engines that interface seamlessly with datasets to produce reports mimicking human expression.

Automate generation of draft audit reports. +++ ++

Deep Learning Prediction

Range of optimized algorithms to identify complex patterns hidden within wide datasets not visible to human eyes.

More accuratepredictions (of failure modes, brewing issues)

+ ++ +++


This presentation is for internal distribution and use only among personnel of Deloitte Touche Tohmatsu Limited, its member firms and their related entities (collectively, the “Deloitte Network”). None of the Deloitte Network shall be responsible for any loss whatsoever sustained by any person who relies on this presentation.

Deloitte refers to one or more of Deloitte Touche Tohmatsu Limited, a UK private company limited by guarantee (“DTTL”), its network of member firms, and their related entities. DTTL and each of its member firms are legally separate and independent entities. DTTL (also referred to as “Deloitte Global”) does not provide services to clients. Please see www.deloitte.com/de/UeberUns for a more detailed description of DTTL and its member firms.

Deloitte provides audit, risk advisory, tax, financial advisory and consulting services to public and private clients spanning multiple industries; legal advisory services in Germany are provided by Deloitte Legal. With a globally connected network of member firms in more than 150 countries, Deloitte brings world-class capabilities and high-quality service to clients, delivering the insights they need to address their most complex business challenges. Deloitte’s approximately 286,000 professionals are committed to making an impact that matters.

Collaborate. Innovate

aiStudiowww.deloitte.com/de/aistudio

http://www.deloitte.com/de/aistudio