Interpretable AI
Building explainable machine learning systems
Ajay Thampi
  • MEAP began June 2020
  • Publication in Summer 2021 (estimated)
  • ISBN 9781617297649
  • 275 pages (estimated)
  • printed in black & white

I think this is a valuable book both for beginners as well for more experienced users.

Kim Falk Jørgensen
AI models can become so complex that even experts have difficulty understanding them—and forget about explaining the nuances of a cluster of novel algorithms to a business stakeholder! Fortunately, there are techniques and best practices that will help make your AI systems transparent and interpretable. Interpretable AI is filled with cutting-edge techniques that will improve your understanding of how your AI models function. Focused on practical methods that you can implement with Python, it teaches you to open up the black box of machine learning so that you can combat data leakage and bias, improve trust in your results, and ensure compliance with legal requirements. You’ll learn to identify when to utilize models that are inherently transparent, and how to mitigate opacity when you’re facing a problem that demands the predictive power of a hard-to-interpret deep learning model.

About the Technology

How deep learning models produce their results is often a complete mystery, even to their creators. These AI "black boxes" can hide unknown issues—including data leakage, the replication of human bias, and difficulties complying with legal requirements such as the EU’s "right to explanation." State-of-the-art interpretability techniques have been developed to understand even the most complex deep learning models, allowing humans to follow an AI’s methods and to better detect when it has made a mistake.

About the book

Interpretable AI is a hands-on guide to interpretability techniques that open up the black box of AI. This practical guide simplifies cutting-edge research into transparent and explainable AI, delivering practical methods you can easily implement with Python and open source libraries. With examples from all major machine learning approaches, this book demonstrates why some approaches to AI are so opaque, teaches you to identify the patterns your model has learned, and presents best practices for building fair and unbiased models. When you’re done, you’ll be able to improve your AI’s performance during training, and build robust systems that counteract errors from bias, data leakage, and concept drift.
Table of Contents detailed table of contents

Part 1: Interpretability Basics

1 Introduction

1.1 Diagnostics+ AI – An Example AI System

1.2 Types of Machine Learning Systems

1.2.1 Representation of Data

1.2.2 Supervised Learning

1.2.3 Unsupervised Learning

1.2.4 Reinforcement Learning

1.2.5 Machine Learning System for Diagnostics+ AI

1.3 Building Diagnostics+ AI

1.4 Gaps in Diagnostics+ AI

1.4.1 Data Leakage

1.4.2 Bias

1.4.3 Regulatory Non-Compliance

1.4.4 Concept Drift

1.5 Building a Robust Diagnostics+ AI

1.6 Interpretability v/s Explainability

1.6.1 Types of Interpretability Techniques

1.7 What will I learn in this book?

1.7.1 What tools will I be using in this book?

1.7.2 What do I need to know before reading this book?

1.8 Summary

2 White-Box Models

2.1 White-Box Models

2.1.1 Diagnostics+ AI – Diabetes Progression

2.2 Linear Regression

2.2.1 Interpreting Linear Regression

2.2.2 Limitations of Linear Regression

2.3 Decision Trees

2.3.1 Interpreting Decision Trees

2.3.2 Limitations of Decision Trees

2.4 Generalized Additive Models (GAMs)

2.4.1 Regression Splines

2.4.2 GAM for Diagnostics+ Diabetes

2.4.3 Interpreting GAMs

2.4.4 Limitations of GAMs

2.5 Looking Ahead to Black-Box Models

2.6 Summary

Part 2: Interpreting Model Processing

3 Model Agnostic Methods - Global Interpretability

3.1 High School Student Performance Predictor

3.1.1 Exploratory Data Analysis

3.2 Tree Ensembles

3.2.1 Training a Random Forest

3.3 Interpreting a Random Forest

3.4 Model Agnostic Methods – Global Interpretability

3.4.1 Partial Dependence Plots

3.4.2 Feature Interactions

3.5 Summary

4 Model Agnostic Methods – Local Interpretability

4.1 Diagnostics+ AI – Breast Cancer Diagnosis

4.2 Exploratory Data Analysis

4.3 Deep Neural Networks

4.3.1 Data Preparation

4.3.2 Training and Evaluating DNNs

4.4 Interpreting DNNs

4.5 LIME

4.6 SHAP

4.7 Anchors

4.8 Summary

5 Saliency Mapping

5.1 Diagnostics+ AI – Invasive Ductal Carcinoma Detection

5.2 Exploratory Data Analysis

5.3 Convolutional Neural Networks

5.3.1 Data Preparation

5.3.2 Training and Evaluating CNNs

5.4 Interpreting CNNs

5.4.1 Probability Landscape

5.4.2 LIME

5.4.3 Visual Attribution Methods

5.5 Vanilla Backpropagation

5.6 Guided Backpropagation

5.7 Other Gradient-based Methods

5.8 Grad-CAM and Guided Grad-CAM

5.9 Which attribution method should I use?

5.10 Summary

Part 3: Interpreting Model Representations

6 Understanding Layers and Units

7 Understanding Semantic Similarity

Part 4: Fairness and Bias

8 Fairness and Mitigating Bias

9 Conclusion

Appendixes

Appendix A: PyTorch

A.1 What is PyTorch?

A.2 Installing PyTorch

A.3 Tensors

A.3.1 Data Types

A.3.2 CPU and GPU Tensors

A.3.3 Operations

A.4 Dataset and DataLoader

A.5 Modeling

A.5.1 Automatic Differentiation

A.5.2 Model Definition

A.5.3 Training

Appendix B: Resources

What's inside

  • Why AI models are hard to interpret
  • Interpreting white box models such as linear regression, decision trees and generalized additive models
  • Partial dependence plots, LIME, SHAP and Anchors, and various other techniques such as saliency mapping, network dissection and representational learning
  • What is fairness and how to mitigate bias in AI systems
  • Implement robust AI systems that are GDPR-compliant

About the reader

For data scientists and engineers familiar with Python and machine learning.

About the author

Ajay Thampi is a machine learning engineer at a large tech company primarily focused on responsible AI and fairness. He holds a PhD and his research was focused on signal processing and machine learning. He has published papers at leading conferences and journals on reinforcement learning, convex optimization, and classical machine learning techniques applied to 5G cellular networks.

placing your order...

Don't refresh or navigate away from the page.
Manning Early Access Program (MEAP) Read chapters as they are written, get the finished eBook as soon as it’s ready, and receive the pBook long before it's in bookstores.
print book $34.99 $49.99 pBook + eBook + liveBook
Additional shipping charges may apply
Interpretable AI (print book) added to cart
continue shopping
go to cart

eBook $27.99 $39.99 3 formats + liveBook
Interpretable AI (eBook) added to cart
continue shopping
go to cart

Prices displayed in rupees will be charged in USD when you check out.

FREE domestic shipping on three or more pBooks