Layer-Wise Relevance Propagation: A Visual Introduction | XAI Methods Explained

Layer-Wise Relevance Propagation (LRP) is a foundational method in Explainable Artificial Intelligence (XAI). Originally introduced in 2015 by researchers at Fraunhofer HHI in collaboration with TU Berlin, LRP explains neural network predictions by propagating relevance scores backward through the model’s layers. This video provides a clear visual introduction to the core principles of LRP, illustrating how relevance is redistributed across layers to highlight the input features most responsible for a given prediction. It is the first episode in our series on XAI methods and neural network interpretability, with upcoming videos covering related techniques such as Concept Relevance Propagation (CRP). Timestamps 00:00 – Introduction 00:31 – A cats and dogs classifier 01:19 – Black box model 03:32 – Fully-connected layers 05:20 – Conservation property 05:57 – Non linearities 07:08 – Max pooling 07:55 – Convolution 08:23 – Explanation 09:29 – Recap References & Resources More information about the XAI Group: https://www.hhi.fraunhofer.de/en/departments/ai/research-groups/explainable-artificial-intelligence.html Original LRP paper (Bach et al., 2015): https://doi.org/10.1371/journal.pone.0130140 LRP overview paper (Montavon et al., 2019): https://link.springer.com/chapter/10.1007/978-3-030-28954-6_10 Interactive LRP demo: https://lrpserver.hhi.fraunhofer.de/ LXT (LRP for Transformers): https://github.com/rachtibat/LRP-eXplains-Transformers Zennit (PyTorch LRP framework): https://github.com/chr5tphr/zennit iNNvestigate (Keras/TensorFlow): https://github.com/albermax/innvestigate Find more information here: https://www.hhi.fraunhofer.de/en/departments/ai/technologies-and-solutions/layer-wise-relevance-propagation.html Subscribe for more content like this! Follow us on social media: LinkedIn: https://de.linkedin.com/company/fraunhofer-hhi Instagram: https://www.instagram.com/fraunhofer_hhi/ Buesky: https://bsky.app/profile/fraunhoferhhi.bsky.social Mastodon: https://wisskomm.social/@FraunhoferHHI