An Overview of Prototype Formulations for Interpretable Deep Learning

TL;DR

This paper reviews various prototype formulations for interpretable deep learning, introduces a new probabilistic hyperspherical prototype method, and demonstrates its superior performance and training efficiency across multiple datasets.

Contribution

It introduces HyperPG, a novel probabilistic hyperspherical prototype approach, and provides a comprehensive comparison of point-based and probabilistic formulations in different latent spaces.

Findings

Hyperspherical prototypes outperform Euclidean ones in accuracy.

HyperPG maintains performance with simplified training schemes.

Euclidean prototypes need extensive hyperparameter tuning.

Abstract

Prototypical part networks offer interpretable alternatives to black-box deep learning models by learning visual prototypes for classification. This work provides a comprehensive analysis of prototype formulations, comparing point-based and probabilistic approaches in both Euclidean and hyperspherical latent spaces. We introduce HyperPG, a probabilistic prototype representation using Gaussian distributions on hyperspheres. Experiments on CUB-200-2011, Stanford Cars, and Oxford Flowers datasets show that hyperspherical prototypes outperform standard Euclidean formulations. Critically, hyperspherical prototypes maintain competitive performance under simplified training schemes, while Euclidean prototypes require extensive hyperparameter tuning.