Layer-Wise Relevance Propagation Heatmaps for a computationally-aligned CNN

This dataset contains Layer-wise Relevance Propagation (LRP) heatmaps for an extended VGG-Face model (Parkhi et al., 2015) aligned to computational similarity scores. The model was built with an architecture developed by (Hofmann et al., 2024) and fine-tuned on similarity metrics derived from activations of the VGG-Face maxpool5-3 layer to predict the odd-one-out in face triplets. The resulting heatmaps highlight facial regions (e.g., eyes, nose, mouth) that contributed to predictions matching the computational similarity structure. This dataset serves as a baseline for comparison with a human-aligned model (Hofmann et al., 2024) trained to reproduce human similarity judgments. The dataset includes: Relevance maps as .npy files for each odd-one-out prediction of a face triplet. Visual previews of each relevance map in .png format. Metadata CSV mapping each heatmap to triplet indices, computational ground truth, and model predictions. Heatmap ComputationHeatmaps were computed using Layer-wise Relevance Propagation (Bach et al., 2015) implemented with the Zennit library (Anders et al., 2021), using an Epsilon–Alpha2Beta1–Flat composite. Relevance was propagated from the model’s predicted odd-one-out class, and relevance values were summed across RGB channels to obtain 2D maps. Visualization PNG previews use a symmetric seismic colormap, normalized to the maximum absolute relevance value of each individual map ($v_{min} = -|max|$, $v_{max} = |max|$), where red indicates positive relevance (supporting the decision) and blue indicates negative relevance. Please read the README_MAXP.txt for more information about how to use the data. Source code for model training and heatmap generation is available via the linked GitHub repository FaceSimXAI.