machine learning, market structure & competition
TRANSCRIPT
Carl Shapiro and Hal Varian
NBER Economics of Artificial IntelligenceToronto, Canada
13 September 2017
Machine Learning, Market Structure & Competition
Quick Review of Machine Learning
Slide 2
What is Machine Learning?
âą Predict Labels as Function of FeaturesâȘ Classic Approach: Construct Numerical
Features, Construct RulesâȘ Deep Learning: Use Raw Data, Learn Directly
âą Images: Pixelsâą Translation: Paired Documentsâą Transcription: Voice and Text
âȘ Requires Labeled Data (OpenImages), Hardware (GPU, TPU), Software (TensorFlow), Expertise
âą Optimize Using Reinforcement LearningâȘ Multi-Armed BanditsâȘ Chess, Go, Atari Games etc.
Slide 3
What Can Machine Learning Do?âą Kaggle PredictionsâȘ Passenger Threats; Home Prices; Traffic to
Wikipedia Pages; Personalized Medicine; ImageNet; Taxi Trip Duration; Product Purchases; Clustering Questions; Rental Listing Interest; Lung Cancer Detection; Click Prediction; Inventory Demand
âą Demand: Match Customer & Productâą Supply: Reduce Cost and Wasteâą Substitute and Complement HumansâȘ Reduced Demand: Cashiers, Drivers, TranslatorsâȘ Increases Demand: Analytic Skills
Slide 4
What ML Inputs Are Scarce?
à Data Infrastructure: Critical Prerequisite§ Collection, Manipulation, Storage & Retrieval§ System Integrators Can Play Big Role
Ă Software: Open Source & In CloudĂ Hardware: Can Be Purchased in CloudĂ Expertise: Scarce But Growing RapidlyĂ Firm-Specific Labeled Data: Key Input
Slide 5
Obtaining Labeled Data
âą Multiple Ways to Obtain Needed DataâȘ As By-Product of OperationsâȘ By Offering a Service (GOOG411, Flickr)âȘ Hiring Humans to Label DataâȘ Buying Data from ProviderâȘ Sharing Data (Perhaps Mandated)âȘ Data from Governments and/or Consortia
âą Data is Non-Rival, Partially ExcludableâȘ Rights, Permissions, Licensing, RegulationâȘ âOwnershipâ Too Narrow a Concept for PolicyâȘ Example: Who Control Driverless Car Data?
Slide 6
Big Data, ML and Public Policy
âą Does Access to Data Give Incumbents a Major Competitive Advantage?âȘ Entrants Must Build or Acquire Necessary DataâȘ But: Entrants May Have Data From Adjacent Markets
âą Incumbents Also Learn How to Improve Algorithms and Business ProcessesâȘ Shape of the âMachine Learning Learning CurveââȘ Domain Knowledge Can Be Important
âą Apply Essential Facility Doctrine to Data?âȘ Scope of âEssentialâ Data that Must be Shared?âȘ How to Regulate Terms & Conditions of Data Access?
Slide 7
Machine Learning Meets Good Old Industrial Organization
Slide 8
Adoption of ML Technology
à Which Firms and Industries Will Successfully Adopt Machine Learning?§ Large Heterogeneity in Timing of Adoption &
Ability to Use ML EffectivelyĂ Can Later Adopters Imitate Early Adopters?
§ Patents & Trade Secrets; Firm-Specific Routinesà Role of Geography in Adoption Patternsà Very Large Competitive Advantage for Early,
Successful Adopters § Large Firms? New Firms? Disruptive Aspects
Carl Shapiro Slide 9
Evidence on AI Adoption
Ă McKinsey Global Institute Survey§ 3000 âAI Aware C-Level Executivesâ in 10 Countries§ 20% Are âSerious Adopters⧠40% are Experimenting or are âPartial Adopters⧠28% Feel Their Firms Lack the Technical
Capabilities to Implement AI
à Key Enablers of AI Adoption§ Leadership, Technical Ability, Data Access
Carl Shapiro Slide 10
AI Adoption by Industry (McKinsey)
Carl Shapiro Slide 11
Key Research Question: Machine Learning & Vertical Integration
à How Will Machine Learning Tools and Data Be Combined to Create Value?§ Within or Across Corporate Boundaries?
à Will ML Users Develop Their Own ML Capabilities or Purchase ML Solutions from ML Vendors?§ Classic Make vs. Question, Key for Industry Analysis
à One-Stop Shopping in the Cloud is Happening§ Data Labeling, Software, Algorithms, Consulting§ Special-Purpose Hardware: Tensor-Processing Units
(TPUs) Create Cost Advantage
Slide 12
Machine Learning and Vertical Integration: Some Public Policy Questions
à Privacy Regulations May Limit Ability of ML Vendors to Combine Data from Multiple Sources§ Limits on Transfer of Data Across Corporate
Boundaries and/or Sale of Data§ Privacy Concerns vs. Growth of Markets for
Data Used for Machine LearningĂ Mandated Data Sharing May Promote
Vertical DisintegrationĂ Treatment of Vertical Mergers Between ML
Vendors and ML UsersSlide 13
Machine Learning Vendors
Slide 14
Structure of the ML Industry
à ML Vendors Offer Several Services§ Data Centers, Containers, Dockers§ Labeling Services, System Integration, Consulting
§ ML Vendor Could Specialize in ML and Purchase Data Processing/Storage in Cloud
à Industry Structure is Oligopolistic§ Leaders: Amazon, Google, Microsoft, Salesforce§ Other Suppliers: IBM? Who Will Be Next?
§ Will Industry Become More Fragmented?§ Specialists by Industry?
Slide 15
Diminishing Returns to Scale
Slide 16
Source: Stanford Dogs dataset
ImageNet Progress: 2010-2015
Slide 17
Data Size Held Constant Improvement Due to Hardware and Software
Source: Stanford ImageNet
Pricing of ML Services
à Large Fixed Costs, Low Marginal Cost§ Undifferentiated Services & Bertrand Trap?§ Size of Customer Switching Costs§ Containers & Dockers
Ă Learning by Doing for ML VendorsĂ Multi-Product Offerings and Bundling
Slide 18
Pricing of ML Services: Google
Slide 19
Pricing of ML Services: Amazon
Slide 20
Impact of Machine Learning on Downstream Markets
Slide 21
Impact of ML on Minimum Efficient Scale?
à Will ML Generally Increase Minimum Efficient Scale by Transforming Variable Costs into Fixed Costs?§ Fixed Cost of Developing a ML Solution§ Substitutes for Variable Labor Costs
à Not if the Fixed Costs of ML are Small§ Off-the-Shelf Generic ML Capabilities vs. Need
to Develop a Specialized Solution§ See: Pricing Structure for ML Solutions
à ML Could Lower Minimum Efficient Scale§ Reduce or Eliminate Certain Fixed Costs
Slide 22
How to Start Up a Startup
Ă Fund Your Project on KickstarterĂ Hire Employees Using LinkedInĂ Purchase Cloud Computing Services from AmazonĂ Use Open Source Software: Linux, Python, TensorflowĂ Set up a Kaggle Competition for Machine LearningĂ Communicate Using Skype, Gmail, Google DocsĂ Use Nolo for Legal Documents Ă Market Your Product or Service Using AdWords Ă User Support Provided by ZenDesk
Slide 23
Use of ML for Downstream Pricing
à Far Greater Price Discrimination?§ Yield Management Goes Bananas§ Auctions and Other Mechanisms§ But: Customers Can Use ML Counterstrategies
Ă Group Discrimination â Many Groups!§ More Data on Which to Condition Prices§ Blurs Line Between Individual and Group Pricing
à Self-Selection & Product Differentiation§ Customized Products§ But: Competition + Low Consumer Search Costs
Slide 24
Algorithmic Collusion: Economist Catnip
Ă Classic Issue of Dynamic Oligopoly Pricing Ă Rapid Response Equilibria
§ In Markets with Transparent Prices§ Firms Move Far Faster Than Consumers
à Evolution of Machine Cooperation§ Can Machines Find a Better Way to Coordinate?§ Taking MFNs & MCCs to the Next Level?
à Instructive Examples§ NASDAQ; ATPCO; Spectrum Auctions§ Machines Learning Cryptographic Code
Ă Antitrust Implications: Who Goes to Jail?
Slide 25