What problem does it solve? Standard attribution maps show where a model looks but not which latent concepts drive its predictions. This Skill applies Concept Relevance Propagation (CRP) to isolate individual channel concepts in PyTorch networks, producing concept-conditional heatmaps, relevance-ranked concepts, and representative reference images. ## Core Features & Use Cases - Conditional Attributions (CRP): Mask relevance flows during backpropagation to compute heatmaps conditioned on specific channels and output classes via CondAttribution and ChannelConcept. - Relevance/Activation Maximization: Use FeatureVisualization to precompute RelMax or ActMax reference images showing which dataset samples best represent each neuron or channel. - Attribution Graphs & Concept Statistics: Trace relevance through model layers and rank channel-level importance across Conv2d and Linear layers. - Use Case: Given a pretrained VGG16, identify the top-5 most relevant channels in features.40 for a prediction, then retrieve and crop their RelMax reference images for embedding into CLIP or DINOv2 to auto-label each neuron with a semantic concept. ## Quick Start Run the conditional attribution demo script to compute a CRP heatmap for channels 50 and 100 in layer features.28 of a pretrained VGG16 model.
