Can we Constrain Concept Bottleneck Models to Learn Semantically Meaningful Input Features?

• DOI: 10.48550/arxiv.2402.00912
• Publication Date: 2024-02-01
• Authors: Jack Furby; Daniel Cunnington; Dave Braines; Alun Preece
• Keywords: FOS: Computer and information sciences; Computer Science - Machine Learning; Artificial Intelligence (cs.AI); Computer Science - Artificial Intelligence; Computer Vision and Pattern Recognition (cs.CV); Computer Science - Computer Vision and Pattern Recognition; Machine Learning (cs.LG)

Concept Bottleneck Models (CBMs) are considered inherently interpretable because they first predict a set of human-defined concepts before using these to the output downstream task. For inherent interpretability be fully realised, and ensure trust in model's output, we need guarantee predicted based on semantically mapped input features. example, one might expect pixels representing broken bone an image used for prediction fracture. However, current literature indicates this is not case, as concept predictions often irrelevant We hypothesise that occurs when annotations inaccurate or how features should relate unclear. In general, effect dataset labelling representations CBMs remains understudied area. Therefore, paper, examine learn from datasets with fine-grained annotations. demonstrate can semantic mapping by removing problematic correlations, such two always appearing together. To support our evaluation, introduce new synthetic playing cards domain, which hope will serve benchmark future CBM research. validation, provide empirical evidence real-world chest X-rays, meaningful learned applications.