DEPICT: Diffusion-Enabled Permutation Importance for Image Classification Tasks

TL;DR

This paper introduces DEPICT, a permutation-based explanation method for image classifiers that assesses feature importance by permuting interpretable concepts across images and measuring the impact on model performance.

Contribution

DEPICT is the first permutation-based explanation method for image models that operates on interpretable concepts using diffusion models, enabling global feature importance analysis.

Findings

Recovers underlying feature importance on synthetic data

Effectively ranks concepts by importance in real-world tasks

Provides global explanations beyond pixel-level maps

Abstract

We propose a permutation-based explanation method for image classifiers. Current image-model explanations like activation maps are limited to instance-based explanations in the pixel space, making it difficult to understand global model behavior. In contrast, permutation based explanations for tabular data classifiers measure feature importance by comparing model performance on data before and after permuting a feature. We propose an explanation method for image-based models that permutes interpretable concepts across dataset images. Given a dataset of images labeled with specific concepts like captions, we permute a concept across examples in the text space and then generate images via a text-conditioned diffusion model. Feature importance is then reflected by the change in model performance relative to unpermuted data. When applied to a set of concepts, the method generates a ranking of feature importance. We show this approach recovers underlying model feature importance on synthetic and real-world image classification tasks.