MAMMOth "Multi-Attribute, Multimodal Bias Mitigation in AI Systems"
- Thessaloniki, Greece
- November 2022
- ETHNIKO KENTRO EREVNAS KAI TECHNOLOGIKIS ANAPTYXIS (CERTH)
The EU-funded MAMMOth project tackles AI bias by focusing on multi-discrimination mitigation for tabular, network and multimodal data. Working with computer science and AI experts, the project will create tools for fairness-aware AI which ensure accountability with respect to protected attributes like gender, race and age. The project will also engage with communities of vulnerable and/or underrepresented groups in AI research to ensure that user needs and pains are truly at the centre of the agenda. The end goal is to develop pilot projects for finance/loan applications, identity verification and academic evaluation.
- Open source
- 48/51/0 (% m/f/d)
- Public
Project stage (in ):
Research/planning
Implemented by:
Inhouse
Industrial Sectors:
Information and communication, Professional, scientific and technical activities, Other service activities
Usage of AI:
Natural Language Processing, Computer Vision, Data Management and Analysis, Human-Computer Interaction, Information Retrieval
Generation of AI:
Deep Learning (CNN, Transformers, etc.), Graph mining
Model training:
Supervised Learning, Semi-supervised Learning, Unsupervised Learning, Reinforcement Learning, Transfer Learning
Contact
Responsible Person
Coordinator: Dr Symeon Papadopoulos
Motivation and values
What in your view defines the public interest and how does your project meet this purpose?
In general, “public interest”, with definitions as the “common good”, is considered a general term, due to its different interpretations. MAMMOth promotes a human centred approach to attribute selection and multi-discrimination to evaluate datasets and develop AI solutions, in the form of a co-creation process that engages both AI developers and affected stakeholders. Human centric approaches have a universal impact and interest; reducing algorithmic bias, increases commercial viability and trust in technology, reduces social and financial exclusion and even affects political stability.
How did the idea of your project come about?
Internal in our organization, From an outside request
What is the goal of your project in relation to the public interest?
MAMMOth is expected, inter alia, to raise awareness of AI developers on prospective types of bias they need to acknowledge and to create the basis for the next generation of socially aware AI developers, in order to develop socially responsible AI solutions and be equipped with better tools that simplify the detection and mitigation of bias in AI solutions. The goal of MAMMOth, in relation to the public interest, is to provide services to the public that are unbiased and eventually contribute to the demand for a fair and inclusive European society.
Did you follow one or more guidelines for ethical AI, and if yes, which one?
European Union - Ethics Guidelines for Trustworthy AI, Council of Europe - Feasibility Study (CAHAI), OECD - Principles on AI, IBM’s Principles for Trust, OpenAI Charter
Guiding Values
What are your top 5 guiding-values for the project? Top 1
Redefine bias based on multiple (protected) characteristics instead of a single attribute.
What measures do you use to implement this value? Top 1
Operationalize definitions of multi-discrimination characteristics.
What are your top 5 guiding-values for the project? Top 2
Create standardised AI solutions to address bias across all phases of development of AI systems.
What measures do you use to implement this value? Top 2
Design step by step processes for the evaluation of bias in datasets and support mitigation strategies and decision making.
What are your top 5 guiding-values for the project? Top 3
Develop and advance new technologies to evaluate and mitigate AI bias
What measures do you use to implement this value? Top 3
Design multi-criteria, multimodal bias mitigation algorithms
What are your top 5 guiding-values for the project? Top 4
Ensure reliability, traceability and explainability of AI solutions
What measures do you use to implement this value? Top 4
Design and advance multi-discrimination accountability / explanations to address different types of bias across all stages of AI development.
What are your top 5 guiding-values for the project? Top 5
Increase availability and deployment of unbiased and bias-preventing AI solutions
What measures do you use to implement this value? Top 5
Integrate and prototype the proposed solutions into the MAMMOth open source suite
Design & Safeguards
Which stakeholders were involved in the process of development and implementation of the project?
Developer, Civil society, Academic researcher, Industry partners
How did you engage relevant stakeholders?
Workshops, Oversight board, Survey, Interviews, Co-creation process
Did you apply specific methods of participatory design, and if yes, which ones?
Νο
Have the project results been validated by third parties?
No
Have the design and the results of the project been made transparent to the public?
No
Involving the people who will be affected by the project is a necessary part of the project design
5
Which direct channels for feedback exist?
In general, a crucial element of the project is the setup and operation of an Expert Advisory Board.
In addition, scheduled workshops for the stakeholders and user engagement.
How did you verify the accuracy and robustness of your system?
The developed methods, tools and processes will be continuously evaluated and feedback by the consortium and extended community of collaborators and stakeholders will be collected. This will take place within the scheduled workshops, employing co-creation principles but also across the duration of the development phase through on-demand self-service crowdsourcing approaches. Quantitative objectives that capture accuracy, robustness, and fairness will be created by the project; adherence to these will be asserted by exploring the behavior produced AI on benchmark datasets or test data partitions withheld from training procedures. Such a process will be validated through the peer-reviewed publication of scientific principles, experimental methodology, and results.
Which transparency measures do you use to document your use of data and possible biases or limitations of the data sets?
Data sheets, Model cards, This project may also produce new transparency measures to report biases or limitations of datasets.
What technical and organizational safeguards have you implemented to protect personal data and mitigate possible harms? Choose all that apply
Data minimization (incl. not gathering personal data), Encryption, Authentication and access management, User control (consent, update, retract), Maintaining an up to date privacy policy
Do you take measures in regards to ecological sustainability of your project? Please elaborate.
Ν/Α
Is your project directly or indirectly contributing to solutions for the climate crisis or other UN sustainability goals?
Directly
Please elaborate
MAMMOth goals are contributing directly to all the 17 UN sustainability goals, especially to gender equality, reduction of inequality, sustained, inclusive and sustainable economic growth, and fostering innovation.
Does your project rely partly or fully on the use of open data and/or publish results in an open data set? Please elaborate.
MAMMOth is going to provide, to the extent possible, open access to its data and publications on the project’s website and to open data repositories.