Caption-Driven Explainability: Probing CNNs for Bias via CLIP

Caption-based explanations can reveal hidden biases that pixel-level analysis misses: Traditional saliency maps highlight which pixels matter most, but this approach can be misleading when spurious features overlap with legitimate ones. By generating natural language captions describing what the model "sees," this method identifies the actual concepts driving predictions, making it easier to spot when models rely on problematic shortcuts.

The method helps detect "covariate shift"—when models fail on data that looks different from training data: By understanding which concepts a model prioritizes, researchers can identify whether it's relying on superficial patterns that won't generalize. This is critical for deploying AI systems in real-world conditions where data inevitably differs from training scenarios.

Integrating language models with vision models provides more interpretable explanations: Rather than trying to interpret images alone, this approach uses CLIP (a model trained on images and text together) to translate visual patterns into human-readable concepts. This bridge between vision and language makes it easier for non-experts to understand and audit AI decision-making.

This technique supports the development of more robust and trustworthy ML systems: By making model behavior more transparent and identifying bias sources earlier in development, this XAI method contributes to building AI systems that are less likely to fail or discriminate in deployment.