Explainable AI for Practitioners

Chapter 4. Explainability for Image Data

Introduced in the 1980s, convolutional neural networks (CNNs), like DNNs, remained unused until the advent of modern ML, at which point they quickly became the backbone of contemporary solutions for computer vision problems. Since then, deep learning models based on CNNs have enabled unprecedented breakthroughs in many computer vision tasks ranging from image classification and semantic segmentation to image captioning and visual question answering, at times achieving near human-level performance. Nowadays, you can find sophisticated computer vision models being used to design smart cities, monitor livestock or crop development, build self-driving cars, or identify eye disease or lung cancer.

As the number of these intelligent systems relying on image models continues to grow, the role of explainability in analyzing and understanding these systems has become more important than ever. Unfortunately, when these highly complex systems fail, they can do so without any warning or explanation and sometimes with unfortunate consequences. Explainability AI (XAI) techniques are essential to build trust not only in the users of these systems but especially for the practitioners putting these models into production.

In this chapter, we’ll focus on explainability methods you can use to build more reliable and transparent computer vision ML solutions. Computer vision tasks differ from other ML tasks in that the base features (i.e., pixels) are rarely ...