Towards Synthetic Concept Activation Vectors via Generative Models

In recent years, the field of Explainable Artificial Intelligence (XAI) has developed a new paradigm called concept-based XAI, which fosters using human-understandable concepts to verify hypotheses or test models against biases. A Concept Activation Vector (CAV) is the representation of a concept in a vision model's embedding space. However, training a CAV requires a properly sized dataset containing images of the selected concept. This may represent a limitation, as acquiring such a dataset can be difficult and time-consuming. In this context, a text-to-image generation system may help analysts build CAVs directly from texts, reducing time requirements while maintaining faithfulness. This work lays the foundations of synthetic CAV generation using pre-trained text-to-image generative models. Our approach consists of producing synthetic concept images from a descriptive prompt and using them to train a CAV in the space of a pre-trained vision model. The methodology also includes a quality control step with a multi-modal embedding model to discard images containing errors or artifacts. We evaluate the quality of our proposal by running experiments on popular ImageNet CNNs using a set of randomly chosen concepts and then comparing the synthetic CAVs with the ones from real images. Our results show that it is possible to train faithful CAVs by generating concept images, particularly for simpler concepts such as textures. Fine-tuning generative models with a few real images also yields promising results.