Theory and Practice of Quality Assurance for Machine Learning Systems An Experiment-Driven Approach

This book is a self-contained introduction to engineering and testing machine learning (ML) systems. It systematically discusses and teaches the art of crafting and developing software systems that include and surround machine learning models. Crafting ML based systems that are business-grade is highly challenging, as it requires statistical control throughout the complete system development life cycle. To this end, the book introduces an “experiment first” approach, stressing the need to define statistical experiments from the beginning of the development life cycle and presenting methods for careful quantification of business requirements and identification of key factors that impact business requirements. Applying these methods reduces the risk of failure of an ML development project and of the resultant, deployed ML system. The presentation is complemented by numerous best practices, case studies and practical as well as theoretical exercises and their solutions, designed to facilitate understanding of the ideas, concepts and methods introduced.

The goal of this book is to empower scientists, engineers, and software developers with the knowledge and skills necessary to create robust and reliable ML software.

1. Introduction.- 2. Scientific Analysis of ML Systems.- 3. Motivation and Best Practices for Machine Learning Designers and Testers.- 4. Unit Test vs. System Test of ML Based Systems.- 5. ML Testing.- 6. Principles of Drift Detection and ML Solution Retraining.- 7. Drift Detection by Measuring Distribution Differences.- 8. Sequential Drift Detection.- 9. Drift in Characterizations of Data.- 10. A Framework Analysis for Alternating Components and Drift.- 11. Optimal Integration of the ML Solution in the Business Decision Process.- 12. Testing Solutions Based on Large Language Models.- 13. A Detailed Chatbot Example.

Samuel Ackerman earned his Ph.D. in statistics from Temple University in Philadelphia, PA, in 2018.  Since then, he has worked as a statistician and data science researcher at IBM Research Israel in Haifa, actively contributing to the development of machine learning (ML) testing and analysis methods and tools.

Guy Barash earned his M.Sc. in computer science with a focus on AI, from Bar Ilan University in 2021. His scientific research examines vulnerabilities of ML software. For eight years, he has been working in the software industry – both corporate and startup – on the design and implementation of reliable ML-based systems.

Eitan Farchi earned his Ph.D. in game theory from Haifa University in Israel, in 2000. He is a distinguished engineer at IBM Research and works on the development of methods, tools and field solutions for quality and reliability of software systems. Recently, he focused on quality and reliability of industrial strength ML-based solutions in the area of intelligent chatbot software.

Orna Raz holds a Ph.D. in Software Engineering from Carnegie Mellon University. Over the years, she has studied the quality of industrial strength software. Recently, she focused on ML-based systems and has conceptualized and developed FreaAI - a slice-based ML software analysis tool that is used for industrial ML software quality analysis.

Onn Shehory is a professor of Intelligent Information Systems at Bar Ilan University (BIU), Israel, where he also serves as the director of the Data Science and AI Institute. He has many years of both academic and industrial experience in the fields of AI and software engineering. In recent years his research focused on ML, its vulnerabilities, and methods for mitigating related risks.