Abstract
Neural networks are widely adopted to solve complex and challenging tasks.
Especially in high-stakes decision-making, understanding their reasoning
process is crucial, yet proves challenging for modern deep networks. Feature
visualization (FV) is a powerful tool to decode what information neurons are
responding to and hence to better understand the reasoning behind such
networks. In particular, in FV we generate human-understandable images that
reflect the information detected by neurons of interest. However, current
methods often yield unrecognizable visualizations, exhibiting repetitive
patterns and visual artifacts that are hard to understand for a human. To
address these problems, we propose to guide FV through statistics of real image
features combined with measures of relevant network flow to generate
prototypical images. Our approach yields human-understandable visualizations
that both qualitatively and quantitatively improve over state-of-the-art FVs
across various architectures. As such, it can be used to decode which
information the network uses, complementing mechanistic circuits that identify
where it is encoded. Code is available at: github.com/adagorgun/VITAL
BibTeX
@online{Goerguen_2503.22399,
TITLE = {{VITAL}: More Understandable Feature Visualization through Distribution Alignment and Relevant Information Flow},
AUTHOR = {G{\"o}rg{\"u}n, Ada and Schiele, Bernt and Fischer, Jonas},
LANGUAGE = {eng},
URL = {https://arxiv.org/abs/2503.22399},
EPRINT = {2503.22399},
EPRINTTYPE = {arXiv},
YEAR = {2025},
MARGINALMARK = {$\bullet$},
ABSTRACT = {Neural networks are widely adopted to solve complex and challenging tasks.<br>Especially in high-stakes decision-making, understanding their reasoning<br>process is crucial, yet proves challenging for modern deep networks. Feature<br>visualization (FV) is a powerful tool to decode what information neurons are<br>responding to and hence to better understand the reasoning behind such<br>networks. In particular, in FV we generate human-understandable images that<br>reflect the information detected by neurons of interest. However, current<br>methods often yield unrecognizable visualizations, exhibiting repetitive<br>patterns and visual artifacts that are hard to understand for a human. To<br>address these problems, we propose to guide FV through statistics of real image<br>features combined with measures of relevant network flow to generate<br>prototypical images. Our approach yields human-understandable visualizations<br>that both qualitatively and quantitatively improve over state-of-the-art FVs<br>across various architectures. As such, it can be used to decode which<br>information the network uses, complementing mechanistic circuits that identify<br>where it is encoded. Code is available at: https://github.com/adagorgun/VITAL<br>},
}Endnote
%0 Report %A Görgün, Ada %A Schiele, Bernt %A Fischer, Jonas %+ Computer Vision and Machine Learning, MPI for Informatics, Max Planck Society Computer Vision and Machine Learning, MPI for Informatics, Max Planck Society Computer Vision and Machine Learning, MPI for Informatics, Max Planck Society %T VITAL: More Understandable Feature Visualization through Distribution Alignment and Relevant Information Flow : %G eng %U http://hdl.handle.net/21.11116/0000-0010-FC8B-6 %U https://arxiv.org/abs/2503.22399 %D 2025 %X Neural networks are widely adopted to solve complex and challenging tasks.<br>Especially in high-stakes decision-making, understanding their reasoning<br>process is crucial, yet proves challenging for modern deep networks. Feature<br>visualization (FV) is a powerful tool to decode what information neurons are<br>responding to and hence to better understand the reasoning behind such<br>networks. In particular, in FV we generate human-understandable images that<br>reflect the information detected by neurons of interest. However, current<br>methods often yield unrecognizable visualizations, exhibiting repetitive<br>patterns and visual artifacts that are hard to understand for a human. To<br>address these problems, we propose to guide FV through statistics of real image<br>features combined with measures of relevant network flow to generate<br>prototypical images. Our approach yields human-understandable visualizations<br>that both qualitatively and quantitatively improve over state-of-the-art FVs<br>across various architectures. As such, it can be used to decode which<br>information the network uses, complementing mechanistic circuits that identify<br>where it is encoded. Code is available at: https://github.com/adagorgun/VITAL<br> %K Computer Science, Computer Vision and Pattern Recognition, cs.CV