Keywords:	deep learning, machine learning, trustworthy AI, fairness, bias
Prerequisites:	neural networks, knowledge of deep learning (also self-taught is good)
Difficulty:	variable (depends on the chosen topics)
Group work (only for B.Sc.):	possible

Abstract

Bias/unfairness is a concept with several definitions, the most common one being a condition whereas a Machine Learning model displays systematically different predictions/error rates across different subgroups contained within a dataset. An example might be different error rates in some face recognition models across race and age, as discovered in the GenderShades project. The present project proposal is very open and revolves around the generic theme of assessment of fairness. It is up to the student to define a domain/dataset and to establish which notion(s) of fairness to study. The goal of the project can be either uncovering potential cases of (un)fairness or to investigate the meaning and consequences of adopting given definitions of fairness in a specific domain. Additionally, it may also be possible to supplement the findings with some additional actions, like, for instance, operating bias-correcting measures.

Required work

Notice that it is up to the student (with my supervision) to identify areas where studies on fairness are underrepresented, in order to allow for a certain degree of scientific novelty in the research.

• identify the domain(s) to study
• find a dataset useful for fairness assessment (i.e., with the proper subgroups already defined in the annotations)
• identify the notion(s) of fairness to adopt
• find a model to study (can be pre-trained or trained from scratch)
• experiment & analyze

Python + PyTorch is the preferred language/library, but I’m open to considering also other options

Relevant literature

• Survey on fairness: Mehrabi et al. A Survey on Bias and Fairness in Machine Learning. ACM Computing Surveys. 2021.
• Ways to measure fairness in face recognition: Howard et al. Evaluating Proposed Fairness Models for Face Recognition Algorithms. 2022.
• An example of fairness assessment for a class of computer vision models. Zullich, Marco, and Giovanni Santacatterina. Assessing Fairness in Open-Source Face Mask Detection Algorithms. (2023).
• Recently-published dataset for fairness assessment: Gustafson et al. FACET: Benchmarking fairness of vision models. 2023.
• A short essay on the ambiguity of the concept of fairness in Machine Learning, with a historiographic approach: Möding, Amira, et al. Contexts, Concepts, and Critique. On the limits of the notions of ‘fairness’ and ‘bias’ employed in Computer Vision and Natural Language Processing. 2023 Currently under review.