Welcome to MAMMOth! This is an EU project that has created ready-to-use solutions for multi-attribute, multimodal bias mitigation of AI systems. Learn more about the project at mammoth-ai.eu.
Software developers, policymakers, and non-experts: come visit us at the MAI-BIAS portal!
Since you are already in GitHub, below are quick links to several related repositories. These can be also be discovered from the portal, alongside more material.
A fairness assessment and mitigation app. It loads datasets and models, and runs fairness/bias analysis through the same simple UI:
- set up a local runner - fast bootstrapping, work with local data
- set up in your organization's server - easy and safe sharing of results (data still don't leave your organization's security)
The solutions below support the MAI-BIAS toolkit but are also usable independently. They are organized per data modality.
Covering multiple modalities:
- FairBench - A comprehensive AI fairness exploration library with >300 standardized measures.
- MMM-fair - A library that supports high-stakes AI decision-making under competing fairness and accuracy demands.
- FairBranch - Implementation of FairBranch pipelines for vision and tabular modalities.
For tabular data:
- MBFPP - Implementation of the work Multi-Fairness-under-Class-Imbalance.
- FB-tiny - Terminal tool for fast tabular data bias audits with minimal memory and energy consumption.
For visual data:
- VB-Mitigator - Implementation and evaluation of existing and new visual bias mitigation methods.
- FLAC - Implementation of the model for Fairness-Aware Representation Learning by Suppressing Attribute-Class Associations.
- SDFD - The Stable Diffusion Face-image Dataset, which captures a broad spectrum of facial diversity encompassing not only demographics and biometrics but also non-permanent traits like make-up, hairstyle, and accessories.
For graphs: