Neural Network Interpretability With Layer-Wise Relevance Propagation: Novel Techniques for Neuron Selection and Visualization

Deepshikha Bhati 1 , Fnu Neha 1 , Md Amiruzzaman 2 , Angela Guercio 1 , Deepak Kumar Shukla 3 , Ben Ward 1 · 2024

Read Neural Network Interpretability With Layer-Wise Relevance Propagation: Novel Techniques for Neuron Selection and Visualization by Deepshikha Bhati 1 , Fnu Neha 1 , Md Amiruzzaman 2 , Angela Guercio 1 , Deepak Kumar Shukla 3 , Ben Ward 1 (2024) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2412-05686.txt.

Description

Deepshikha Bhati and five co-authors present a 2024 method for seeing why a neural network answers as it does. The catalog record pads the author names with affiliation numbers, so the byline follows the head of the text. Their base tool passes relevance backward through the network, crediting inputs for the verdict. Older versions judge single neurons poorly.

The new method parses chosen neurons more sharply as the credit flows back, using a known vision network as its test bed. It draws graphs of the key paths and paints them as heat maps, picking neurons by two error scores. A reverse filter rebuilds feature maps for a fuller view. Tests show richer, clearer readings of the network's inner steps, the authors say.

Catalog record

CREATORDeepshikha Bhati 1 , Fnu Neha 1 , Md Amiruzzaman 2 , Angela Guercio 1 , Deepak Kumar Shukla 3 , Ben Ward 1
YEAR2024
COLLECTIONPAPERS
FORMATTXT
FILE SIZE21 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2412-05686

Browse the full library · About the Tower of Babel