This is the repository for the article: https://distill.pub/2020/attribution-baselines/
Integrated Gradients has become a popular method for interpreting deep neural networks. As a hyper-parameter, the method requires the user to choose a baseline input x' that the explanations are relative to. What does the baseline input mean? And how important is it? In this article, we explore this hyper-parameter, and argue that, although it is often over-looked, it is an important hyper-parameter. We use the example of an image classification network to demonstrate why this hyper-parameter can impact how you interpret your networks. Finally, we discuss how this choice of hyper-parameter relates to a broader understanding of how we interpret machine learning models, and what it means to represent missingness to a model.
Note: the public/
folder in this repository contains a static webpage. Simply clone it, enter the directory and start up a server (something like python3 -m http.server
in the repo directory) to view.