New Metrics and Methods for CAM Evaluation and Refinement
A new synthetic dataset featuring accurate ground-truth attributions has been developed by researchers to assess class attribution maps (CAMs) utilized in convolutional neural networks. They have introduced ARCC, a composite metric designed to more effectively pinpoint reliable explanations, alongside RefineCAM, a technique that generates high-resolution attribution maps by combining CAMs from various layers of the network. This research tackles issues related to the evaluation of CAMs and the constraints posed by low-resolution outputs.
Key facts
- arXiv:2605.14641v1
- Synthetic dataset with ground-truth attributions introduced
- ARCC composite metric proposed
- RefineCAM method produces high-resolution attribution maps
- CAMs provide local explanations for CNN decisions
- Evaluation challenge due to lack of ground-truth explanations
- Most CAM methods produce low-resolution maps
- RefineCAM aggregates CAMs across multiple network layers
Entities
Institutions
- arXiv