Model Performance Management with Explainable AI

Model Performance Management with Explainable AI

by Amit Paka, Krishna Gade, Danny Farah
Released November 2021
Publisher(s): O'Reilly Media, Inc.
ISBN: 9781098108670

Read it now on the O’Reilly learning platform with a 10-day free trial.

O’Reilly members get unlimited access to books, live events, courses curated by job role, and more from O’Reilly and nearly 200 top publishers.

Start your free trial

Book description

Artificial intelligence has the potential to provide productive, efficient, and innovative solutions to everyday problems. But it comes with risks. Multiple examples of alleged bias in AI have been reported in recent years, and by the time those issues surfaced, many people were already affected. This could have been avoided if humans had visibility into every stage of the system life cycle.

In this report, Danny Farah and Amit Paka explain the importance of establishing an efficient model performance management (MPM) system in your organization's machine learning workflow. You'll learn how MPM enables CxOs, IT leaders, and AI/ML leaders to gain visibility into every stage of the system life cycle. That includes training ML models to help your system make decisions.

This report covers:

  • MPM and explainability: Explore a data-centric framework for producing high-quality ML and AI models and systems
  • Explainable AI (XAI): Generate explanations from ML models so humans can explain and interpret the overarching AI system
  • The ML life cycle: Follow an ML model on its journey from conception to production
  • MPM in the ML life cycle: Learn how MPM can provide full visibility into issues that arise when training, deploying, and monitoring models
  • MPM and responsible AI: Explore ways to ensure that your AI systems are built with responsibility in mind

Table of contents

  1. Preface
  2. 1. Introduction to Model Performance Management and Explainability
    1. DevOps and Application Performance Management
      1. DevOps
      2. Application Performance Management
      3. Application Performance Metrics
    2. The Rise of MLOps
      1. MLOps Compared to DevOps
      2. Model Performance Metrics
      3. The ML Development Process
    3. Model Performance Management
      1. MPM with Explainability for Full Coverage
      2. Benefits of Using an MPM Framework
  3. 2. Explainable AI
    1. Explainability in Context
      1. How Does XAI Fit into Responsible AI?
    2. Who Needs XAI?
    3. How XAI Can Be Used to Manage Model Performance
      1. Interpretability Versus Explainability
      2. Models That Are Interpretable by Design Versus Explainable After
      3. Understanding Model Predictions
      4. Offline Versus Online Explainability
    4. XAI in Different Domains
      1. Tabular/Structured Data Models (Highest Contributing Features)
      2. Text/Speech Models (Sentiment)
      3. Image/Video Models (Heatmaps)
  4. 3. The Machine Learning Life Cycle
    1. The Three Types of Analytics
    2. Life Cycle Stages
      1. Problem Definition
      2. Data Collection
      3. Data Processing and Storage
      4. Metrics Definition
      5. Data Exploration
      6. Feature Extraction and Engineering
      7. Model Training and Offline Evaluation
      8. Model Integration and Deployment
      9. Model Release and Monitoring
    3. How MPM Fits into the Life Cycle
  5. 4. MPM in the ML Life Cycle
    1. The ML Feedback Loop
      1. Brief on Control Theory
      2. Control Theory and MPM
    2. MPM in the Training and Deployment Stage
      1. Using XAI to Understand Model Bias
      2. Choosing the Best Model to Release
    3. MPM in the Release and Monitoring Stage
      1. Detecting Model Drift
      2. Finding the Root Cause Using MPM and XAI
      3. Live Experiments
  6. 5. Implementing MPM in Practice
    1. Model Performance in Staging Versus Production (Offline Versus Online)
      1. Model Training and Offline Performance
      2. Online Model Performance
    2. An Ideal MPM Framework
      1. Integrated Tools Provided by Cloud Platforms
      2. Specialized MPM Software
      3. Custom-Built Systems
  7. 6. MPM and Responsible AI
    1. What Is Responsible AI?
    2. The Challenges of Responsible AI
      1. Explainability
      2. Reliability
      3. Fairness
      4. Cultural Change
    3. How Model Performance Management Solves These Challenges
      1. Explainable AI
      2. Monitoring in Production
      3. A Single Source of Truth
    4. Model Governance and How MPM Fits into the Life Cycle
    5. The Future of Responsible AI and MPM
      1. New Regulations
      2. A New Role: Chief Ethics Officer
      3. More Tools to Test Bias
      4. MPM with Monitoring Will Become Mission Critical
      5. ML Model Validation Spreads Beyond Banking
  8. About the Authors

Product information

  • Title: Model Performance Management with Explainable AI
  • Author(s): Amit Paka, Krishna Gade, Danny Farah
  • Release date: November 2021
  • Publisher(s): O'Reilly Media, Inc.
  • ISBN: 9781098108670