Department of Computer Engineering

Concept Bottleneck Models (CBMs) have shown promise in improving the interpretability of neural networks by generating intermediate, human-understandable concepts before making predictions. Existing approaches focus on training downstream classifiers jointly with CBMs to produce interpretable classification results, but they often struggle to generalize to unseen classes. In this work, we propose Zero-shot Concept Bottleneck Models (ZS-CBMs), which enable interpretable open-vocabulary recognition. Our methods leverage pre-trained vision-language models to identify attributes related to given image and text inputs and project these inputs into a shared concept space. This allows for zero-shot classification of unseen classes while providing interpretable justifications for model decisions. The accuracy and interpretability of the proposed approaches are evaluated on widely used datasets, including CIFAR-100, CUB-200-2011, CC3M, and ImageNet-100.