Explainable Concept Generation through Vision-Language Preference Learning for Understanding Neural Networks' Internal Representations

Aditya Taparia, Som Sagar, Ransalu Senanayake

International Conference on Machine Learning (ICML), 2025

picture_as_pdf PDFplay_circle Videocode Code
Overview figure for RLPO, showing a diffusion model fine-tuned by reinforcement learning preference optimization to generate concept images that explain a classifier.

Summary

We propose RLPO, a reinforcement learning preference optimization method that frames concept-based explanation as an image generation problem, fine-tuning a vision-language generative model to automatically discover interpretable concepts a network has learned — including concepts difficult or impossible for humans to specify in advance.

BibTeX

@inproceedings{taparia2025explainable,
    title={Explainable Concept Generation through Vision-Language Preference Learning for Understanding Neural Networks' Internal Representations},
    author={Taparia, Aditya and Sagar, Som and Senanayake, Ransalu},
    booktitle={Proceedings of the 42nd International Conference on Machine Learning},
    year={2025}
}