Machine Learning Bias Exploration

Explore how bias can be detected and mitigated in real-world datasets

This German scientific poster (developed as HTML5/React web app) shows the potential implications of racial/social/... Biases in the context of Machine Learning

A face database with a large number of high-quality attribute annotations

Mitigating algorithmic bias using regularisation and reweighing with bi-objective (accuracy and fairness) measure

The code generates simulated research landscapes of any given bias. It can be used to calculate the Machine Learning Gain under different conditions.

Coursework solutions for a 2nd year Computer Science sub-module on Bias in AI @ Durham University. Attempts to minimise gender bias when using ML to predict income.

Source code and notebooks to reproduce experiments and benchmarks on Bias Faces in the Wild (BFW).

An ML competition on CodaLab to estimate the age from images while mitigating the bias

A small and simple prototype designed to alert users of the bias of the news source.

Explainable AI & fashion talk & experiments

Official code of "Discover and Mitigate Unknown Biases with Debiasing Alternate Networks" (ECCV 2022)

Data and Results for WWW paper

A teacher model with minimal bias or calibrate its outputs to reduce bias transfer during distillation. Focus on distilling information from teacher features that are less prone to bias, like semantic representations instead of raw activations.

CIRCLe: Color Invariant Representation Learning for Unbiased Classification of Skin Lesions

CIRCLe: Color Invariant Representation Learning for Unbiased Classification of Skin Lesions. Mirror of https://github.com/arezou-pakzad/CIRCLe

Research POC on the mitigation of bias in large language models (FLAN-T5 and Bloomz) through fine-tuning.

A demonstration of detecting and mitigating bias in AI.

Code for the paper The Other Side of Compression: Measuring Bias in Pruned Transformers (IDA23)

Gender classification model that uses a CNN to classify images of faces as male or female. The notebook includes code for data preprocessing, model architecture, training, and evaluation which will then be used for algorithmic bias detection.