Title: « Multi-objective optimization and deep learning for healthcare applications »
Mots-clés : Deep learning, Multi-objective optimization, Transformers, EHR.
Stochastic gradient descent methods are widely used in the field of machine learning and more particularly, to optimize the parameters of deep neural networks (deep learning).
Most gradient descent methods aim to optimize a single function representing a single objective or a linear combination of multiple objectives. Recently, variants of gradient descent for multiobjective optimization have been proposed in the literature . These methods have been used for multi-task learning .
We developed, in the AROB@S team, a multi-objective approach that aims to optimize the architecture of a neural network by removing neurons and connections between neurons during the learning phase to improve efficiency and model interpretability. The goal of this internship is to adapt this approach to multi-task learning using real data from electronic health records of patients admitted to intensive care units (MIMIC-IV) . Hypernetworks  or multi-objective gradient descent  will be used for multi-objective optimization.
Steps of the project
- Study of the multi-objective approach proposed in our team as well as state-of-the-art deep learning architectures based on attention mechanisms like Transformers .
- Adapt and implement the proposed approach to process time series and generate a Pareto front.
- Apply the implementation to real-world time series data of patients while considering multiple tasks, including early predictions of sepsis.
Duration of internship
IBISC Laboratory, IBGBI, University of Évry, 23 Boulevard de France, 91000 Évry-Courcouronnes.
 Liu, L.N. Vicente, The stochastic multi-gradient algorithm for multi-objective optimization and its application to supervised machine learning, arXiv:1907.04472, 2021.
 Sener, O., & Koltun, V. Multi-task learning as multi-objective optimization. NeurIPS,
 Rasmy, L., Xiang, Y., Xie, Z., Tao, C., & Zhi, D. (2021). Med-BERT: pretrained contextualized embeddings on largescale structured electronic health records for disease prediction. NPJ digital medicine, 4(1), 1-13.
 Johnson, A.E.W., Bulgarelli, L., Shen, L. et al. MIMIC-IV, a freely accessible electronic health record dataset. Sci Data 10, 1 (2023). https://doi.org/10.1038/s41597-022-01899-x
 Navon, Aviv & Shamsian, Aviv & Chechik, Gal & Fetaya, Ethan. (2020). Learning the Pareto Front with Hypernetworks.
- Date de l’appel : 03/11/2023
- Statut de l’appel : Non pourvu
- Contacts cotés IBISC : Farida ZEHRAOUI (MCF Univ. Évry, IBISC équipe AROB@S), Éric ANGEL (PR Univ. Évry, IBISC équipe AROB@S, Massinissa HAMIDI (MCF Univ. Évry, IBISC équipe AROB@S),
- Sujets de stage niveau Master 2 (format PDF)
- Web équipe AROB@S