R&D Intern – Fairness in Deep Learning

Job description

Context

Master’s research internship lasting 4–6 months, starting in winter 2026–2027. This document outlines possible research directions, which will be refined during the internship. The internship is based at Artefact’s research center in Paris. The goal is to study neural networks in a setting where predictions should be independent of a continuous sensitive variable, such as age, income, or skin color.

About Artefact

Artefact is an international data-services company specializing in data-transformation consulting. Our mission is to turn data into tangible results throughout companies’ value chains. Our distinctive approach bridges data and business, helping clients achieve their business goals effectively. Our 2,500 employees combine multidisciplinary skills to drive business innovation. Our advanced AI technologies and agile methods support client AI projects from design and deployment through training and change management.

Research topic

Neural networks are central to many industrial applications. Their ability to extract rich representations from data makes them powerful tools, but also makes them susceptible to encoding and amplifying biases in training data, thereby reinforcing discrimination against populations affected by these models [1]. This is studied as fairness, a particularly active research area as AI regulation evolves, for example through the European Union AI Act.

Research on measuring and mitigating discriminatory bias has focused mainly on binary or categorical sensitive variables (gender, ethnicity, etc.), where groups are clearly defined [2–5]. However, many sensitive variables are inherently continuous, including age, income, and skin color. Discretizing them into groups is common but problematic: thresholds are arbitrary, information is lost, and a model may appear fair under one grouping while remaining highly biased under another [6]. Controlling dependence between model outputs and a continuous variable remains an open problem, particularly for neural networks, where this dependence must both be measured and penalized differentiably during training.

During this internship, you will explore methods for reducing neural-network bias with respect to a continuous sensitive variable. First, implement existing methods—regularization based on dependence measures such as Hirschfeld–Gebelein–Rényi maximal correlation or “Generalized Disparate Impact,” among those in references [7–8]—and establish a robust comparison protocol that considers the trade-off between performance and fairness. Then design and test new bias-reduction methods suited to continuous variables. You may use open-source libraries such as PyTorch and scikit-learn to keep the code reproducible and easy to adopt.

Requirements

Bac+4/Bac+5 education in applied mathematics, statistics, or machine learning at a university or engineering school. Gap-year or final-year internship. Start: winter 2026–2027. Duration: 4–6 months.

References

Risser, L., Sanz, A. G., Vincenot, Q., & Loubes, J. M. (2022). Tackling algorithmic bias in neural-network classifiers using Wasserstein-2 regularization. Journal of Mathematical Imaging and Vision, 64(6), 672–689.

Buolamwini, J., & Gebru, T. (2018, January). Gender shades: Intersectional accuracy disparities in commercial gender classification. Conference on Fairness, Accountability and Transparency, 77–91. PMLR.

Cho, J., Zala, A., & Bansal, M. (2023, October). Dall-Eval: Probing the reasoning skills and social biases of text-to-image generation models. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), 3020–3031. IEEE.

Li, X., Chen, Z., Zhang, J. M., Sarro, F., Zhang, Y., & Liu, X. (2023). Dark-skin individuals are at more risk on the street: Unmasking fairness issues of autonomous driving systems. CoRR abs/2308.02935.

Thong, W., Joniak, P., & Xiang, A. (2023, October). Beyond skin tone: A multidimensional measure of apparent skin color. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), 4880–4890. IEEE.

Shilova, V., Malherbe, E., Palma, G., Risser, L., & Loubes, J. M. (2025). Fairness-aware grouping for continuous sensitive variables: Application for debiasing face analysis with respect to skin tone. ECAI 2025 Proceedings, 5344–5351.

Mary, J., Calauzenes, C., & El Karoui, N. (2019, May). Fairness-aware learning for continuous attributes and treatments. International Conference on Machine Learning, 4382–4391. PMLR.

Giuliani, L., Misino, E., & Lombardi, M. (2023, July). Generalized disparate impact for configurable fairness solutions in ML. International Conference on Machine Learning, 11443–11458. PMLR.