Believe the HiPe: Hierarchical perturbation for fast, robust, and model-agnostic saliency mapping. (September 2022)

• Record Type: Journal Article
• Authors: Cooper, Jessica; Arandjelović, Ognjen; Harrison, David J
• Is Part Of: Pattern recognition. Volume 129(2022)
• Publication Date: 2022-09
• Subjects: XAI -- AI safety -- Saliency mapping -- Deep learning explanation -- Interpretability -- Prediction attribution; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2022.108743
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: Explainable AI (XAI) is increasingly necessary for AI safety as we build complex models in high-domains and deploy them widely. Saliency mapping is a popular explanation/attribution XAI technique for deep learning. Existing model-agnostic saliency mapping approaches are prohibitively slow. Hierarchical Perturbation (HiPe) is a new model-agnostic method which generates heatmaps of comparable or superior quality to the state-of-the-art. And is 20 × faster than existing model-agnostic saliency methods. 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 × faster to compute.