R&D Intern - Fairness in Deep Learning - Paris

Context:

Master's research internship offer for a duration of 4 to 6 months, starting in winter 2026-2027. This document lists the possible directions for the internship, which will be adjusted throughout the internship. The internship will take place within Artefact's research center in Paris.

The objective is to study neural networks in a framework where we want to guarantee the independence of predictions from a continuous sensitive variable (such as age, income, skin color, etc.).

About Artefact:

Artefact is an international data services company, specializing in data transformation consulting, whose mission is to transform data by delivering tangible results across the entire business value chain. Artefact's unique approach, which bridges data and business, allows our clients to achieve their business objectives in a dedicated and effective way. Our 2500 employees combine their multidisciplinary skills for the benefit of companies' business innovation. Our cutting-edge Artificial Intelligence technologies and agile methods ensure the success of our clients' AI projects, from design to deployment, including training and change management.

Subject:

Neural networks are at the heart of many industrial applications today. Their ability to extract rich representations from data makes them powerful tools, but it also makes them susceptible to encoding and amplifying biases present in training data, thus reinforcing discrimination among populations processed by these models [1]. This issue is a research area called fairness, and it is particularly dynamic, with the regulation of AI models evolving, as for example with the European Union's AI Act.

To measure and mitigate discriminatory biases in models, the literature has mostly focused on binary or categorical sensitive variables (gender, ethnicity, etc.), for which groups are well-defined [2-5]. However, many sensitive variables are intrinsically continuous: age, income, or even skin color. Discretizing these variables into groups is a common but problematic practice: the choice of thresholds is arbitrary, information is lost, and a model may appear fair for a given breakdown while remaining strongly biased for another [6]. Controlling the dependence between the outputs of a model and a continuous variable remains an open problem, especially for neural networks, where this dependence must be both measured and penalized in a differentiable way during training.

During this internship, the intern will explore methods for bias reduction with respect to a continuous sensitive variable for neural networks. The work will begin with the implementation of existing methods (regularizations based on dependence measures such as maximum Hirschfeld-Gebelein-Rényi correlation or “Generalized Disparate Impacte”, etc.), listed in references [7-8], and the establishment of a robust protocol for comparing them, taking into account the trade-off between performance and fairness. In a second phase, the internship will focus on the design and experimentation of new bias reduction methods adapted to continuous variables. To do this, the intern will be able to leverage open-source libraries (PyTorch, scikit-learn), in order to maintain a reproducible and easy-to-adopt code framework.

Prerequisites

The profile sought for this position is as follows:
BAC+4/BAC+5: Training in applied mathematics, statistics, or machine learning at a university or engineering school
Gap year internship or final studies internship
Start date: Winter 2026-2027
Duration: 4 to 6 months

References

1. 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.
2. Buolamwini, J., & Gebru, T. (2018, January). Gender shades: Intersectional accuracy disparities in commercial gender classification. In Conference on fairness, accountability and transparency (pp. 77-91). PMLR.
3. Cho, J., Zala, A., & Bansal, M. (2023, October). Dall-eval: Probing the reasoning skills and social biases of text-to-image generation models. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV) (pp. 3020-3031). IEEE.
4. 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.
5. Thong, W., Joniak, P., & Xiang, A. (2023, October). Beyond skin tone: A multidimensional measure of apparent skin color. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV) (pp. 4880-4890). IEEE.
6. 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. In ECAI 2025: 28th European Conference on Artificial Intelligence 25-30 October 2025, Bologna, Italy-Including 14th Conference on Prestigious Applications of Intelligent Systems (PAIS 2025) Proceedings (pp. 5344-5351).
7. Mary, J., Calauzenes, C., & El Karoui, N. (2019, May). Fairness-aware learning for continuous attributes and treatments. In International conference on machine learning(pp. 4382-4391). PMLR.
8. Giuliani, L., Misino, E., & Lombardi, M. (2023, July). Generalized disparate impact for configurable fairness solutions in ML. In International Conference on Machine Learning (pp. 11443-11458). PMLR.

Find Jobs in France on Arbeitnow