SUB: Benchmarking CBM Generalization via Synthetic Attribute Substitutions

TL;DR

This paper introduces SUB, a synthetic benchmark dataset with 38,400 images to evaluate the robustness of concept bottleneck models (CBMs) under distribution shifts, and proposes a novel image generation method called Tied Diffusion Guidance (TDG).

Contribution

The paper presents a new benchmark dataset, SUB, for testing CBMs' generalization with synthetic concept substitutions, and introduces TDG for precise control of generated images.

Findings

CBMs struggle with concept variations under distribution shifts.

SUB enables rigorous evaluation of CBMs' robustness.

TDG improves control over synthetic image generation.

Abstract

Concept Bottleneck Models (CBMs) and other concept-based interpretable models show great promise for making AI applications more transparent, which is essential in fields like medicine. Despite their success, we demonstrate that CBMs struggle to reliably identify the correct concepts under distribution shifts. To assess the robustness of CBMs to concept variations, we introduce SUB: a fine-grained image and concept benchmark containing 38,400 synthetic images based on the CUB dataset. To create SUB, we select a CUB subset of 33 bird classes and 45 concepts to generate images which substitute a specific concept, such as wing color or belly pattern. We introduce a novel Tied Diffusion Guidance (TDG) method to precisely control generated images, where noise sharing for two parallel denoising processes ensures that both the correct bird class and the correct attribute are generated. This novel benchmark enables rigorous evaluation of CBMs and similar interpretable models, contributing to the development of more robust methods. Our code is available at https://github.com/ExplainableML/sub and the dataset at http://huggingface.co/datasets/Jessica-bader/SUB.