Data and Results - Explainable AI methods for human-aligned face perception

Explainable AI methods for human-aligned face perception This record accompanies the GitHub repository for the project: FaceSimXAI. Original Project Base: FaceSim3D by Hofmann et al. (2024) licensed under MIT License (see LICENSE-MIT). Dataset Description This record contains the data and results for the project "Explainable AI methods for human-aligned face perception". This work extends the FaceSim3D project by applying Explainable AI (XAI) methods, specifically Layer-wise Relevance Propagation (LRP), to investigate the decision-making processes of human-aligned face similarity models. Folder Structure This record contains the zipped data and results from the original project. data.zip Contains the raw and intermediate data used by the models. faces: Head mapping data, reproduced from Hofmann et al. (2024) frontal_view_heads: 3D reconstruction of the 100 faces, reproduced from Hofmann et al. (2024) and originates from the Chicago Face Database (CFD) heatmaps: averaged_heatmaps_humJudgement: Averaged LRP heatmaps for VGG-Hum. averaged_heatmaps_maxp5_3: Averaged LRP heatmaps for VGG-MaxP. single_LRP_HJ_heatmaps_pred_ID: Single instance heatmaps for VGG-Hum. stored on Zenodo with DOI 10.5281/zenodo.18152382 single_LRP_MAXP_heatmaps_pred_ID: Single instance heatmaps for VGG-MaxP. stored on Zenodo with DOI 10.5281/zenodo.18153823 VGGFace_Base_Heatmaps: Heatmaps for VGG-Face Base Model main: Main data, reproduced from Hofmann et al. (2024) models: Model weights and scripts, adapted from Hofmann et al. (2024) VGGFace_MaxPSim_Model: Computed features of the VGG-Face MaxPooling5_3 layer. results.zip Contains the final output of the analyses. heads/VGGface: VGG-Face Feature Maps, reproduced from Hofmann et al. (2024) LRP_LLM_analysis: Results from the Large Language Model comparisons. LRP_PCA_Autoencoder_analysis: Results from PCA and Autoencoder structure analysis. LRP_Region_analysis: Results from the facial region importance analysis. Usage For instructions on how to use the data please read the README.md in the linked GitHub repository: FaceSimXAI Important Note on Missing Data Some source facial images originate from the Chicago Face Database (CFD) and cannot be redistributed due to license restrictions. Consequently, the following folders in this dataset do not contain the used images: data/frontal_view_heads Citation If you use the original FaceSim3D data or models, please additionally cite: Hofmann, S. M., Ciston, A. B., Koushik, A., Klotzsche, F., Hebart, M. N., Müller, K., … Gaebler, M. (2025, August 11). Dynamic presentation in 3D modulates face similarity judgments – A human-aligned encoding model approach. https://doi.org/10.31234/osf.io/f62pw_v4