A holistic and real-world approach to operationalizing artificial intelligence in your company
In Operating AI, Director of Technology and Architecture at Ericsson AB, Ulrika Jägare, delivers an eye-opening new discussion of how to introduce your organization to artificial intelligence by balancing data engineering, model development, and AI operations. You'll learn the importance of embracing an AI operational mindset to successfully operate AI and lead AI initiatives through the entire lifecycle, including key areas such as; data mesh, data fabric, aspects of security, data privacy, data rights and IPR related to data and AI models.
In the book, you’ll also discover:
- How to reduce the risk of entering bias in our artificial intelligence solutions and how to approach explainable AI (XAI)
- The importance of efficient and reproduceable data pipelines, including how to manage your company's data
- An operational perspective on the development of AI models using the MLOps (Machine Learning Operations) approach, including how to deploy, run and monitor models and ML pipelines in production using CI/CD/CT techniques, that generates value in the real world
- Key competences and toolsets in AI development, deployment and operations
- What to consider when operating different types of AI business models
With a strong emphasis on deployment and operations of trustworthy and reliable AI solutions that operate well in the real world - and not just the lab - Operating AI is a must-read for business leaders looking for ways to operationalize an AI business model that actually makes money, from the concept phase to running in a live production environment.
Table of Contents
Foreword xii
Introduction xv
Chapter 1 Balancing the AI Investment 1
Defining AI and Related Concepts 3
Operational Readiness and Why It Matters 8
Applying an Operational Mind- set from the Start 12
The Operational Challenge 15
Strategy, People, and Technology Considerations 19
Strategic Success Factors in Operating AI 20
People and Mind- sets 23
The Technology Perspective 28
Chapter 2 Data Engineering Focused on AI 31
Know Your Data 32
Know the Data Structure 32
Know the Data Records 34
Know the Business Data Oddities 35
Know the Data Origin 36
Know the Data Collection Scope 37
The Data Pipeline 38
Types of Data Pipeline Solutions 41
Data Quality in Data Pipelines 44
The Data Quality Approach in AI/ML 45
Scaling Data for AI 49
Key Capabilities for Scaling Data 51
Introducing a Data Mesh 53
When You Have No Data 55
The Role of a Data Fabric 56
Why a Data Fabric Matters in AI/ML 58
Key Competences and Skillsets in Data Engineering 60
Chapter 3 Embracing MLOps 71
MLOps as a Concept 72
From ML Models to ML Pipelines 76
The ML Pipeline 78
Adopt a Continuous Learning Approach 84
The Maturity of Your AI/ML Capability 86
Level 0 - Model Focus and No MLOps 88
Level 1 - Pipelines Rather than Models 89
Level 2 - Leveraging Continuous Learning 90
The Model Training Environment 91
Enabling ML Experimentation 92
Using a Simulator for Model Training 94
Environmental Impact of Training AI Models 96
Considering the AI/ML Functional Technology Stack 97
Key Competences and Toolsets in MLOps 103
Clarifying Similarities and Differences 106
MLOps Toolsets 107
Chapter 4 Deployment with AI Operations in Mind 115
Model Serving in Practice 117
Feature Stores 118
Deploying, Serving, and Inferencing Models at Scale 121
The ML Inference Pipeline 123
Model Serving Architecture Components 125
Considerations Regarding Toolsets for Model Serving 129
The Industrialization of AI 129
The Importance of a Cultural Shift 139
Chapter 5 Operating AI Is Different from Operating Software 143
Model Monitoring 144
Ensuring Efficient ML Model Monitoring 145
Model Scoring in Production 146
Retraining in Production Using Continuous Training 151
Data Aspects Related to Model Retraining 155
Understanding Different Retraining Techniques 156
Deployment after Retraining 159
Disadvantages of Retraining Models Frequently 159
Diagnosing and Managing Model Performance Issues in Operations 161
Issues with Data Processing 162
Issues with Data Schema Change 163
Data Loss at the Source 165
Models Are Broken Upstream 166
Monitoring Data Quality and Integrity 167
Monitoring the Model Calls 167
Monitoring the Data Schema 168
Detecting Any Missing Data 168
Validating the Feature Values 169
Monitor the Feature Processing 170
Model Monitoring for Stakeholders 171
Ensuring Stakeholder Collaboration for Model Success 173
Toolsets for Model Monitoring in Production 175
Chapter 6 AI Is All About Trust 181
Anonymizing Data 182
Data Anonymization Techniques 185
Pros and Cons of Data Anonymization 187
Explainable AI 189
Complex AI Models Are Harder to Understand 190
What Is Interpretability? 191
The Need for Interpretability in Different Phases 192
Reducing Bias in Practice 194
Rights to the Data and AI Models 199
Data Ownership 200
Who Owns What in a Trained AI Model? 202
Balancing the IP Approach for AI Models 205
The Role of AI Model Training 206
Addressing IP Ownership in AI Results 207
Legal Aspects of AI Techniques 208
Operational Governance of Data and AI 210
Chapter 7 Achieving Business Value from AI 215
The Challenge of Leveraging Value from AI 216
Productivity 216
Reliability 217
Risk 218
People 219
Top Management and AI Business Realization 219
Measuring AI Business Value 223
Measuring AI Value in Nonrevenue Terms 227
Operating Different AI Business Models 229
Operating Artificial Intelligence as a Service 230
Operating Embedded AI Solutions 236
Operating a Hybrid AI Business Model 239
Index 241
