DEPICT: Diffusion-Enabled Permutation Importance for Image Classification Tasks.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Current image-model explanation methods, such as activation maps, primarily offer instance-based insights within the pixel space.
• This pixel-centric approach limits the understanding of global image classifier behavior and feature importance.
• Permutation-based methods are effective for tabular data but lack direct application to image models.

Purpose of the Study:

• To develop a novel permutation-based explanation method for image classifiers.
• To enable a more comprehensive understanding of global model behavior by assessing concept importance.
• To provide a ranking of feature importance for image classification models.

Main Methods:

• Propose a method that permutes interpretable concepts (e.g., from image captions) across a dataset.
• Utilize text-conditioned diffusion models to generate new images with permuted concepts.
• Measure feature importance by evaluating changes in classifier performance on permuted versus unpermuted data.

Main Results:

• The proposed method successfully recovers underlying model feature importance.
• Demonstrated effectiveness on both synthetic and real-world image classification tasks.
• Generates a ranked list of concept importance for image classifiers.

Conclusions:

• Permutation-based concept manipulation offers a powerful alternative to pixel-based explanations for image models.
• This method enhances interpretability by assessing the impact of semantic concepts on classification.
• The approach provides valuable global insights into how image classifiers utilize features.