Revisiting FunnyBirds evaluation framework for prototypical parts networks

TL;DR

This paper critically re-evaluates the FunnyBirds benchmark for prototypical parts networks, revealing that using similarity maps instead of bounding boxes provides more consistent and meaningful explanations aligned with ProtoPNet's core principles.

Contribution

It demonstrates that employing similarity maps for ProtoPNet explanations improves evaluation consistency and suggests a change in visualization practices for explainability benchmarks.

Findings

Similarity maps yield different, more aligned metric scores.

Bounding boxes influence evaluation outcomes significantly.

Similarity maps better reflect ProtoPNet's explanation mechanism.

Abstract

Prototypical parts networks, such as ProtoPNet, became popular due to their potential to produce more genuine explanations than post-hoc methods. However, for a long time, this potential has been strictly theoretical, and no systematic studies have existed to support it. That changed recently with the introduction of the FunnyBirds benchmark, which includes metrics for evaluating different aspects of explanations. However, this benchmark employs attribution maps visualization for all explanation techniques except for the ProtoPNet, for which the bounding boxes are used. This choice significantly influences the metric scores and questions the conclusions stated in FunnyBirds publication. In this study, we comprehensively compare metric scores obtained for two types of ProtoPNet visualizations: bounding boxes and similarity maps. Our analysis indicates that employing similarity maps aligns better with the essence of ProtoPNet, as evidenced by different metric scores obtained from FunnyBirds. Therefore, we advocate using similarity maps as a visualization technique for prototypical parts networks in explainability evaluation benchmarks.