CENDRe: New Method Extracts Concepts from CNNs for Time-Series
A new method for concept extraction, named CENDRe, has been developed by researchers for convolutional neural networks (CNNs) aimed at time-series classification. This innovative approach tackles three key challenges faced by current methods: functioning solely within the time domain, needing a set number of concepts beforehand, and generating misaligned localizations. CENDRe identifies concepts by clustering latent representations at each time step in a two-phase process, employing silhouette-guided aggregation to determine the number of concepts automatically. It then localizes concepts using gradients from a presence score that contrasts with latent representations. Detailed in a cross-type submission on arXiv (ID 2607.29621), this work seeks to enhance the interpretability of CNNs in crucial areas by uncovering the temporal and spectral patterns that influence predictions.
Key facts
- CENDRe is a concept extraction method for CNNs.
- It addresses three limitations: time-domain-only, predefined number of concepts, and misaligned localizations.
- It uses two-stage clustering of per-timestep latent representations.
- Silhouette-guided aggregation automatically selects the number of concepts.
- Localization uses gradients of a presence score contrasting latent representations.
- The paper is available on arXiv with ID 2607.29621.
- The announcement type is cross.
- The method is designed for time-series classification.
Entities
Institutions
- arXiv