Concept Bottleneck Models for Explainable Image Classification under Noisy Annotations

Concept Bottleneck Models (CBMs) have emerged as a promising paradigm in concept-based Explainable Artificial Intelligence (C-XAI) by introducing human-interpretable concepts as an intermediate representation between the input and the final prediction. By explicitly modeling semantic concepts, CBMs provide greater transparency, facilitate concept-level interventions, and enable more interpretable decision-making than conventional end-to-end deep learning models. However, their effectiveness critically depends on the quality of concept supervision, the choice of training strategy, and the design of the downstream prediction model. In practical applications, concept annotations are frequently affected by noise, ambiguity, or inconsistency, which can significantly degrade both concept learning and predictive performance. This thesis presents a comprehensive empirical investigation of Concept Bottleneck Models under realistic conditions of imperfect concept supervision. Three adopted CBM training paradigms named as independent, sequential, and joint training and optimization, are systematically compared to analyze their impact on predictive accuracy, concept learning quality, and model interpretability. Furthermore, the influence of the downstream prediction module is investigated by comparing a multilayer perceptron (MLP) with a decision tree classifier, providing insights into the trade-offs between predictive expressiveness and interpretability. To improve the robustness of Concept Bottleneck Models in the presence of noisy concept annotations, this thesis proposes a class-wise concept denoising framework based on clustering. The proposed approach refines concept annotations prior to model training by exploiting similarities among concept representations within each target class. Extensive experiments are conducted on the CUB-200-2011 dataset to evaluate the proposed framework under varying levels of concept denoising. Model performance is assessed using both conventional predictive metrics and concept-level evaluation measures, providing a comprehensive analysis of model behavior. The experimental results demonstrate that the choice of training paradigm and downstream predictor substantially influences both the predictive performance and interpretability of Concept Bottleneck Models. Moreover, concept denoising consistently improves the quality of learned concept representations and enhances model robustness under noisy supervision. At the same time, the results reveal an inherent trade-off between concept alignment and predictive flexibility. These findings contribute to a deeper understanding of the design choices underlying Concept Bottleneck Models and provide practical guidelines for developing more robust, reliable, and interpretable concept-based learning systems for Explainable Artificial Intelligence.