CENDRe: Concept Extraction with Natural Domain Representations
paper-conference
Concept discovery and gradient-based localization in time and frequency for time-series CNNs.
Resources
Publication history
Preprint: 31 July 2026. Accepted at NeurIPS 2026; final proceedings details are forthcoming.
Summary
CENDRe discovers concepts in time-series CNNs through two-stage clustering of latent representations, selecting the number of concepts automatically. Gradients of a prototype-based presence score localize each concept in the time domain and, through a differentiable Fourier mapping, in the frequency domain. Class-specific relevance scores connect the concepts to predictions. Synthetic benchmarks and bearing-fault data evaluate explanation quality and reveal frequency bands used by the model.
Citation
@inproceedings{holzapfel2026cendre,
author = {Holzapfel, Antonia and Posada Moreno, Andres Felipe and Trimpe, Sebastian},
title = {CENDRe: Concept Extraction with Natural Domain Representations},
booktitle = {Advances in Neural Information Processing Systems},
year = {2026},
note = {Accepted at NeurIPS 2026; proceedings forthcoming},
eprint = {2607.29621},
archivePrefix = {arXiv},
url = {https://neurips.cc/virtual/2026/poster/150459}
}