Do Explanations Expose Bias? How Saliency Maps Affect Judgements of Biased Face-Recognition Models

Accompanying data for the paper "Do Explanations Expose Bias? How Saliency Maps Affect Judgements of Biased Face-Recognition Models".

MODELS

fair.pt: the balanced model.

light_undersampled.pt: the light skin tone undersampled model.

dark_undersampled.pt: the dark skin tone undersampled model.

DATASETS

These are the subset of each dataset used to train the models above.

light_undersampled_cropped.zip: the dataset for the light_undersampled.pt.

dark_undersampled_cropped.zip: the dataset for the dark_undersampled.pt.

diverse_human_faces_cropped.zip: the dataset for the fair.pt.

TRIAL DATA

These relate to the data used during the trials (experiment ran on Prolific).

trial_dataset.csv: overview of the images used for each trial. model_1 and model_2 refers to which models relate to the specific trial.

trial_dataset_long.csv: a deeper overview of trial_dataset.csv, giving specific details about each image.

trial_images.zip: for each trial n, all the images that were generated. Refer to trial_dataset_long.csv to see which images were used in each trial.

METADATA

This is the metadata from the original dataset, Diverse Human Faces Dataset.

dataset_metadata.csv: metadata for all the images in original dataset with labels created in 01_Dataset_Prep.ipynb.