Researchers have developed a new framework for understanding relevance propagation in neural networks, framing it as a form of entropic optimal transport. This approach unifies various existing LRP rules under a single principle and introduces a novel method called Sinkhorn Relevance Propagation. The new method has been tested on image classifiers like VGG-16 and ConvNeXt-Tiny, demonstrating its ability to accurately account for evidence flow and identify influential hidden units. AI
IMPACT Introduces a novel theoretical framework for understanding and potentially improving the interpretability of neural networks.
RANK_REASON The item is an academic paper detailing a new theoretical framework and method for analyzing neural networks. [lever_c_demoted from research: ic=1 ai=1.0]
- ConvNeXt-Tiny
- Entropic Optimal Transport
- Filippov Neural Networks
- image classifiers
- Layer-Wise Relevance Propagation
- LRP$_0$
- LRP$_{\alpha,\beta}$
- LRPN$_{\alpha,\beta}$
- LRP$_{z^+}$
- Relevance Propagation
- Schrödinger Bridge
- Sinkhorn Propagation
- Sinkhorn Relevance Propagation
- VGG-16
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
