Believe The HiPe: Hierarchical Perturbation for Fast, Robust, and Model-Agnostic Saliency Mapping

TL;DR

This paper introduces Hierarchical Perturbation, a fast, robust, and model-agnostic saliency mapping technique that improves interpretability of AI models with significantly reduced computation time.

Contribution

The paper presents a novel hierarchical perturbation method that is faster and more robust than existing model-agnostic saliency mapping approaches.

Findings

Saliency maps are of competitive or superior quality.

Method is over 20 times faster to compute.

Effective on standard benchmarks and datasets.

Abstract

Understanding the predictions made by Artificial Intelligence (AI) systems is becoming more and more important as deep learning models are used for increasingly complex and high-stakes tasks. Saliency mapping -- a popular visual attribution method -- is one important tool for this, but existing formulations are limited by either computational cost or architectural constraints. We therefore propose Hierarchical Perturbation, a very fast and completely model-agnostic method for interpreting model predictions with robust saliency maps. Using standard benchmarks and datasets, we show that our saliency maps are of competitive or superior quality to those generated by existing model-agnostic methods -- and are over 20 times faster to compute.