Zero-Shot T2I Models as Synthetic Concept Sources for XAI
A new arXiv preprint (2605.19855) proposes using zero-shot Text-to-Image (T2I) generative models to create synthetic concept datasets for concept-based Explainable AI (XAI). Concept-based XAI interprets deep learning models via human-understandable visual features like textures or object parts, but traditionally requires large labeled image sets per concept, limiting scalability. The study generates concepts from predefined prompts and evaluates their faithfulness to real concepts through four analyses: concept representation similarity, intra-similarity across subsets of increasing size, and two additional analyses not fully described in the abstract. The work aims to bridge low-level image data and high-level semantics without manual labeling.
Key facts
- arXiv:2605.19855v1
- Announce Type: cross
- Concept-based XAI uses visual features like textures or object parts
- Traditional methods require large labeled image sets per concept
- Zero-shot T2I models generate synthetic concept datasets
- Concepts generated via predefined prompts
- Four complementary analyses evaluate faithfulness to real concepts
- First analysis: concept representation similarity between synthetic and real images
- Second analysis: intra-similarity comparing pairs of subsets of same concept with increasing size
Entities
—