Disentangled deep generative models reveal coding principles of the human face processing network

This is the code accompanying the paper "Disentangled deep generative models reveal coding principles of the human face processing network" by Paul Soulos and Leyla Isik.

Data

The data is available on OSF. Please unzip the archive, there is a README in the archive that explains the data structure.

Code

The matlab code requires Fieldtrip for processing the fMRI data.

Encoding Performance by ROI (Figure 3)

Encoding performance values are generated by correlate_betas.py and correlate_betas_vgg.py using the argument --localizer=roi . The results can be visualized using notebooks/data plots.ipynb.

Whole brain encoding performance (Figure 4)

Encoding performance values are generated by correlate_betas.py and correlate_betas_vgg.py using the argument --localizer=score . The resulting correlation mat file can be converted to nifti using convert_whole_brain_correlation_mat_to_nifti.m and viewed using Freesurfer.

ROI preference mapping (Figure 5)

The ROI preference map data is generated using preference_mapping_roi.m. The results can be visualized using notebooks/encoding feature performance.ipynb.

Facial identity decoding (Figure 6)

The identity decoding accuracies are generated by identity_decoding_whole_brain_xhat.m. The results can be visualized using notebooks/identity decoding.ipynb.

Whole brain encoding performance (Figure S1 and S2)

See the section titled "Whole brain encoding performance (Figure 4)".

STS preference mapping (Figure S3)

The ROI preference map data is generated using preference_mapping_roi.m. The results can be visualized using notebooks/encoding feature performance.ipynb.

Whole brain preference mapping (Figure S4)

The nifti files for the whole brain preference mapping are generated by whole_brain_preference_mapping.m. This nifti file can be viewed using Freesurfer.