Controversial Face Model Checkpoints
This repository contains model checkpoints used in the study:
Human face perception reflects inverse-generative and naturalistic discriminative objectives
These checkpoints are released to support reproducibility of the model-based stimulus optimization and model-behavioral comparison analyses reported in the paper.
The models were trained for research purposes.
Models
All models use the same VGG-16 backbone architecture without batch normalization. They differ in their training objective and training data.
| Model name | Objective | Training data |
|---|
invRend-BFM | Inverse rendering / prediction of BFM latent parameters | BFM synthetic faces |
faceID-BFM | Face identification | BFM synthetic identities |
faceID-VGGFace2 | Face identification | VGGFace2 |
autoenc-BFM | Autoencoding | BFM synthetic faces |
autoenc-VGGFace2 | Autoencoding | VGGFace2 |
objCat-ImageNet | Object categorization | ImageNet |
These models were used as computational hypotheses for human face-dissimilarity judgments. For each model, we trained three independent instances from different random weight initializations. One instance was used for generating ground-truth responses during stimulus optimization (the data-generating instance), a second was used as a reference instance when computing within-model similarity and comparing model representations to human behavior, and a third was held out for model recovery experiments
The checkpoints support extraction of model representations, computation of model-predicted dissimilarities, and model-driven controversial stimulus optimization.
Intended use
These checkpoints are intended for:
- reproducing the analyses reported in the associated paper;
- extracting model representations for the stimulus sets used in the study;
- running controversial stimulus optimization using the code on Github: https://github.com/kriegeskorte-lab/controversial-face.
- comparing computational objectives for face-perception research.
The associated data repository is available at:
https://osf.io/bzx4e
