Risks of AI & Responsible AI - MMPA · 3. Ecosystem 5. Deployment 6. Operate and Monitor Corporate...
Transcript of Risks of AI & Responsible AI - MMPA · 3. Ecosystem 5. Deployment 6. Operate and Monitor Corporate...
Dr. Anand S. RaoGlobal AI Lead, PwC
Risks of AI & Responsible AI
GLOBAL ARTIFICIAL INTELLIGENCE LEADDr. Anand S. Rao
www.pwc.com
PwC– AI Lab
Today’s discussionExcitement Around AI
AI based Automation and Augmentation: Finance & Accounting
01
02
2
Risks of AI03
PwC – AI Lab
Excitement around AI
3
01
PwC – AI LabPwC New Services and Emerging TechConfidential information for the sole benefit and use of PwC’s client.
”In a way, AI is both closer and farther off than we imagine. AI is closer to being able to do more powerful things than
most people expect -- driving cars, curing diseases, discovering planets, understanding media. Those will each
have a great impact on the world, but we're still figuring out what real intelligence is.”
4
Mark Zuckerberg, "Building Jarvis", Facebook, December 19, 2016
PwC – AI LabPwC New Services and Emerging TechConfidential information for the sole benefit and use of PwC’s client.
AI is closer than you think…
5
PwC New Services and Emerging TechConfidential information for the sole benefit and use of PwC’s client.
6
Deep Blue beats Garry KasparovMAY 11, 1997
PwC New Services and Emerging TechConfidential information for the sole benefit and use of PwC’s client.
7
FEB 16, 2011
Watson beats Jeopardy Champions
PwC – AI Lab
AlphaGo beats Lee Sedol
MARCH 15, 2016
8
PwC – AI Lab
AlphaGo v. AlphaGo Zero vs AlphaZero (2016-2017)
• Trained with data from human Go Players
• Uses data from playing with itself
• Generated ’new’ moves that humans had not used
• AlphaGo Lee beats Go Grandmaster Lee Sedol 4-1 in March 2016
• Uses just the rules of the game with no human data
• AlphaGo Zero beats AlphaGoLee in 3 days of training in 2017
• The same program now uses the rules of Chess
• AlphaZero AI beats Stockfish(best Chess program) 64-36
• System was trained in 4 hours using 5,000 TPUs
9
PwC – AI LabPwC New Services and Emerging TechConfidential information for the sole benefit and use of PwC’s client.
AI is farther than you think…
10
PwC New Services and Emerging TechConfidential information for the sole benefit and use of PwC’s client.
11
Deep Learning is NOT EQUAL to Deep Understanding
Alexa: What is the weather?
The weather in Lexington is …
Alexa: What was my previous
question?
Here is something I found on Dictionary.com. Previous question…
12
Two Paths to AIEnterprises are realizing value along two distinct paths from Digitization to AI
Digitization
Artificial Intelligence
Productivity Experience Profits
Simplification
Standardization
Automation
Revenues
Autom
ation P
ath
An
alytics Path
Data (Volume, Velocity, Variety, Veracity, Value)
Cognification
Personalization
Analytics
PwC – AI Lab
Statistics Econometrics Optimization Complexity Theory
Computer Science
Game Theory
FOUNDATION LAYER 13
AI that can act…▪ Robotic process automation
▪ Deep question & answering
▪ Machine translation
▪ Collaborative systems
▪ Adaptive systems
AI that can sense…▪ Natural language
▪ Audio & speech
▪ Machine vision
▪ Navigation
▪ Visualization
AI is defined as the theory and development of systems that sense the environment, make decisions, and act that would normally require human intelligence.
Hear
SeeSpeakFeel
AI that can think…▪ Knowledge & representation
▪ Planning & scheduling
▪ Reasoning
▪ Machine Learning
▪ Deep Learning
Physical
Creative
Cognitive
ReactiveUnderstand
PerceivePlan
Assist
PwC – AI Lab
AI based Automation and Augmentation: Finance & Accounting
14
02
PwC – AI Lab
AI is being applied in four distinct ways progressing from automated to assisted to augmented to autonomous intelligence
15
Automated Intelligence1
Assisted Intelligence
2
Augmented Intelligence3
Autonomous Intelligence4
Hardwired / specific systems
Adaptivesystems
No human in the loop
Human in the loop
+
PwC – AI Lab
Finance in the digital age will deliver better value with reduced resources and will also be a proactive partner to business
16
Advances in automation and augmentation technology offer Finance the opportunity to move into more value-adding roles by improving its own operations as well as bringing technological advances to the business
Challenges?
Introducing pioneering technology to reduce costs and increase efficiency to the rest of the organisation
Improving quality and control of transactional activity using a digital workforce
Bringing forward-looking ‘cockpits’ to business to perform ‘what-if’ analysis
Improving strategic and financial planning for disruptive and transformational initiatives
PwC – AI Lab
So what is artificial intelligence in a business context and what are its limitations?
17
Find patterns in apparently random data or apply structure to unstructured data
Planning and thinking ahead
Learn through repeated exposure to particular problems
And what it can’t do….( not yet anyway) …What AI can do…
Make sense of human communication (speech or text), interpret and identify rich media (e.g.. music and images)
It can’t place its work in context i.e. see the bigger picture
Context aware
It’s only as good as the data it’s been trained on
Self-improvement
It cannot reason from first principles
Common sense reasoning
Its capability is limited to the purpose for which it is built
Multi-tasking
PwC – AI Lab
Augmentation of finance role by AI technologies – natural language processing, machine learning, and intelligent agents
18
Producing reports and narratives with data driven insights
Extracting and processing data from multiple sources and formats such as invoice processing
What-if analysis of financial and
operational goals
Detection of fake articles or journals for
publication, or suspicious internal
activity
Reporting and Insight
Business Model Simulation
Smart Extraction
Fraud Detection
PwC – AI Lab
Case Study: Use of NLP and Machine Learning in document processing
19
PwC – AI Lab
Case Study: Business model simulation – Finance in M&A and Partnerships
20
Pilot
Build
Scale
StrategyPwC assists the Director of Strategic Initiatives of a large Global Auto Manufacturer, in evaluating the ride share/car share business model and potential disruptors in the personal mobility sector
Go-to-Market ModelPwC uses advanced AI techniques (agent-based simulation, big data technologies) to model over 200K go-to-market scenarios modelling consumer behaviour, pricing, adoption, and competitive response
City SelectionClient decides to set up a separate business unit focused on Personal Mobility. PwC assists in city selection model and go-to-market strategy for each city
Acquisitions & PartnershipsClient leverages the personal mobility adoption model to make targeted acquisitions and partner with new Auto-tech players in different cities
AV Operational ModelPwC assists Client in developing an operational model for autonomous vehicles and electric vehicles involved in Personal Mobility
Operational RolloutClient is rolling out personal mobility services in major US cities and aims to collect additional operational data to feed the personal mobility models
EU ExpansionPwC is assisting the Client’s European Division to evaluate and select cities for expanding personal mobility services
PwC – AI Lab
Risks & Challenges of AI
21
03
PwC – AI Lab
Large organizations and regulators have been voicing concerns about the risks associated with AI and the importance of understanding AI ‘Black Box’
22
PwC – AI Lab
AI risks that need to be assessed, mitigated and managed can be categorized into six categories that impact consumers, businesses, societies and nations
23
Performance• Risk of errors• Risk of bias• Risk of opaqueness• Risk of performance instability
Security• Adversarial attacks• Cyber intrusion risks• Privacy risks• Open source software risks
Ethical• Lack of values risk • Value alignment risk
Societal• Reputational risk• Autonomous weapons
proliferation• Risk of intelligence divideEconomic
• Job displacement• Liability risk • Risk of “winner takes all”
concentration of power
BUSINESS-LEVEL RISKS
NATIONAL-LEVEL RISKS
Risks
• Lack of human agency in AI supported processes
• Inability to detect/control rogue AI
Control
PwC – AI Lab
PwC’s Responsible AI toolkit covers the five fundamental aspects that make AI responsible
24
ETHICS & LEGAL Ensure AI development is in line with major local and global regulations, both enacted and emerging; and allow the business to evaluate the ethics of an AI system and how to operationalize ethics in the organization
INTERPRETABILITYEnable human users to understand, appropriately trust, and effectively manage the emerging generation of AI
BIAS & FAIRNESSUncover bias in the underlying data and model development process and enable the business to understand what process may lead to unfairness
ROBUSTNESS & SECURITY Assess the performance of AI over time to identify potential disruptions or challenges to long term performance
GOVERNANCEIntroduce enterprise-wide and end-to-end accountability for AI applications and consistency of operations to minimize risk and maximize ROI
Ethical & Societal
Performance & Security
Control
PwC
Apply Ethics & Legal throughout
25
...helping clients understand the ethical implications of their use of AI
Business Implications
• Ensure AI development is in line with major local and global regulations, both passed and in discussion.
• Allow the business to evaluate the ethics of an AI system and the impact to employees and customers.
• Enable the business to deploy AI with confidence.
How it will operate
• The general and contextualized principles will help in defining the organization’s AI strategy, as well as guiding the development and operations of AI models
• The Legal framework Repository of discussed and passed regulation by territory and industry
Ethical Principles Traceability Matrix
Full Coverage Partial Coverage No Coverage n Principle ID
PwC
Ensure end-to-end Governance
26
...ensuring auditability through the allocation responsibility, accountability and controls for AI
Business Implications
• Introduce enterprise-wide accountability for AI applications and consistency of operations.
• Enable and empower clients to effective control and manage AI applications across the business.
How it operates
• Standardized controls framework
• Centralized dashboard for monitoring of AI across the organization
• Auditable documentation of data, model, and human interaction with AI
Model Level
1. Strategy
3. Ecosystem
5. Deployment
6. Operate and Monitor
Corporate Strategy
Industry Standards & Regulations
Internal Policies & Practices
Operational Support
Compliance
Portfolio Management
Program Oversight
Delivery Approach
Technology Roadmap
Sourcing
2. P
lan
nin
g
Transition & Execution
Ongoing monitoring
Data Extraction
4. D
evel
opm
ent
Change Management
Evaluation & Check-in
Model Integration
Solution Design
Business & Data Understanding
Pre-Processing
Model BuildingWho Benefits:• C-Suite• Risk & Compliance• Process Owners• Data Scientists• Consumers• Regulators
PwC
Uncover potential Bias Business Implications
• Uncover bias risks in the underlying data and model development process.
• Enable the business to understand what influences a model’s decision that could lead to unfairness to the individual or the group.
How it operates
• Bias framework and diagnostic to identify potential concerns in the data, model development, and usage process
• Automated assessment of data and model bias, customized to the needs of the organization, including detection and correction of proxies.
27
...identifying aspects of data, model, and human interaction that lead to unfairness of AI
Statistical ParityConditional Statistical
Parity
Predicted & Actual Outcomes
Test FairnessWell-Calibration
Balance for Positive ClassBalance for Negative Class
Statistical Measures
True PositivesFalse PositivesFalse NegativesTrue Negatives
Predictive ParityFalse Positive Error Rate BalanceFalse Negative Error Rate Balance
Equal OpportunityConditional Accuracy
Overall AccuracyTreatment Equality
Positive Predictive ValueFalse Discovery RateFalse Omission RateNegative Predictive
Value
True Positive RateFalse Positive RateFalse Negative RateTrue Negative Rate
Similarity-based MeasuresCausal Discrimination
Fairness Through Unawareness
Fairness Through Awareness
Causal DefinitionsCounterfactual Fairness
No Unresolved Discrimination
No Proxy DiscriminationFair Inference
Predicted Outcomes
Predicted Probabilities
Decision
Fairness Definition
Protected Attributes
Dataset
PwC
Uncover potential Bias Business Implications
• Uncover bias risks in the underlying data and model development process.
• Enable the business to understand what influences a model’s decision that could lead to unfairness to the individual or the group.
How it operates
• Bias framework and diagnostic to identify potential concerns in the data, model development, and usage process
• Automated assessment of data and model bias, customized to the needs of the organization, including detection and correction of proxies.
28
...identifying aspects of data, model, and human interaction that lead to unfairness of AI
Data Reporting
Model Bias Detection
Bias Intervention
Input data User can upload dataset on the platform or leverage synthetically generated data
Users can perform different exploratory data analysis
Distributions analysis
Multicollinearity
Outliers detection
Missing values analysis
Identifying Proxies
Choose bias intervention techniques
Threshold computation to achieve individual or
group fairness
Leverage fair algorithms that optimize accuracy by
weighing in fairness
Tuning the effect of Proxies on the model by
various techniques
Creates a report that generates summary details of class imbalance and missing values, along with identifying proxy variables
Generates a report summarizing the success/failure details of sensitive attributes by fairness measures.
PwC
Introduce Interpretability Business Implications
• Introduce explainability and interpretability to AI systems, while maintaining a high level of learning performance.
• Enable human users to understand, appropriately trust, and effectively manage the emerging generation of AI.
How operates
• Global Interpretability reports help data scientists to understand black box models
• End-users are provided with comprehensible reports explaining how their data was used to reach a decision
• Automated exploration of decision boundaries, including explanations of what needs to change in order to shift a prediction
29
...helping clients increase their understanding of AI systems and tackle the black box problem
Scratch Dent Crack
Panel Separation Missing Piece Non-damaged
Possible Damaged Parts:Detected Parts:
PwC
Robustness & Security
30
...safeguarding an AI system by through long-term stability and high performance Business Implications
• Assess the performance of AI over time to identify potential disruptions or challenges to long term performance.
• Enable the business to gain confidence in AI through simulation.
How it operates
• AI benchmarks to assess long term performance
• Real-time and continuous monitoring of model decision making and flagging of deviated activity
• Sensitivity analysis & stress testing
Synthesize
Perturb
AI Model
Input Data &Specify the Structure
Learn the Data
Test Model Robustness
Test Model Sensitivity
PwC 31
Are we entering a new AI-inspired arms race?
PwC 32
The key elements of National AI Strategies must address six policy categories
Reskilling• Workforce reskilling• Digital fitness• University education
Basic AI R&D• Moonshot projects• University funding• Business incentives
Business Protection• Local companies• Specific industry sectors• Algorithmic governance
Specialized AI Tech.• Drones• Autonomous vehicles• Service robots
Consumer Protection• Data security• Income security• Digital anonymity
Ethics• Citizen monitoring• Autonomous weapons• Beneficial use of AI
PwC – AI LabPwC New Services and Emerging TechConfidential information for the sole benefit and use of PwC’s client.
33
PwC | A practical guide to responsible AI – Interpretability & ExplainabilityPwC’s Digital Services
Thank you.
© 2019 PwC. All rights reserved. Not for further distribution without the permission of PwC. “PwC” refers to the network of member firms of PricewaterhouseCoopers International Limited (PwCIL), or, as the context requires, individual member firms of the PwC network. Each member firm is a separate legal entity and does not act as agent of PwCIL or any other member firm. PwCIL does not provide any services to clients. PwCIL is not responsible or liable for the acts or omissions of any of its member firms nor can it control the exercise of their professional judgment or bind them in any way. No member firm is responsible or liable for the acts or omissions of any other member firm nor can it control the exercise of another member firm’s professional judgment or bind another member firm or PwCIL in any way.
Dr. Anand S. RaoGlobal AI Lead
[email protected]@AnandSRao