Context: Recent advances in multimodal machine learning have demonstrated the potential of integrating diverse data modalities to achieve more comprehensive feature representations. However, multimodal fusion remains fundamentally challenging due to the intrinsic biases and uncertainty variations across different modalities. These arise from heterogeneity in signal characteristics, acquisition conditions, and feature dynamics. For instance, medical imaging modalities such as CT and MRI, or cellular imaging modalities like fluorescence and bright-field microscopy, exhibit distinct sensitivity to physical and biological properties, leading to unequal contributions of information. As a result, naive fusion of modalities can amplify modality- specific errors or lead to inconsistent predictions under changing imaging conditions. Uncertainty is a central factor in this challenge. It stems from two complementary sources: aleatoric uncertainty, which reflects the inherent noise or ambiguity in the data, and epistemic uncertainty, which captures the limitations of the model or data representation. Ignoring these uncertainties leads to unstable learning, over- or under-segmentation, and poor generalization across domains. Therefore, an effective multimodal fusion strategy must be guided by uncertainty estimation to dynamically adjust modality contributions and ensure robust, reliable representation learning.
Problem Statement: The central problem addressed in this internship is the development of an uncertainty-guided multimodal fusion framework capable of producing stable, bias-resilient feature representations for downstream machine learning tasks such as segmentation or classification. Specifically, the task involves learning a joint representation that remains discriminative under variations in modality quality, missing data, and domain shifts.
Main Objective: The aim is to construct a principled, uncertainty-aware fusion mechanism that yields stable feature representations under real-world variability, thereby improving predictive reliability, interpretability, and robustness across data domains.
Expected Outcomes:
By the end of the internship, the student is expected to deliver:
• A functional codebase demonstrating uncertainty-guided multimodal fusion.
• A concise, well-structured report summarizing methods, results, and analysis.
• Visual examples of uncertainty maps and fusion-based reconstructions.
• A discussion on uncertainty’s role in domain adaptation and robustness evaluation.
Application Information:
Applications should include the subject line: “Application – M2 Internship Uncertainty Guided Multimodal Fusion”. Interested students are invited to apply by sending
their CV and a Motivation Letter to the internship supervisors:
• Dawood Al Chanti – MCF, Gipsa-Lab, Grenoble INP-UGA, Phelma dawood.al-chanti@grenoble-inp.fr
• Mauro Dalla Mura – MCF, HDR, Gipsa-Lab, Grenoble INP-UGA, Ense3 mauro.dalla-mura@grenoble-inp.fr
• Diana Mateus – Prof, LS2N, École Centrale Nantes diana.mateus@ec-nantes.fr
Funding and administrative information: This internship is part of the ANR funded project EQUR-XSM, Quantification Evidentielle de l’Incertitude pour la Fiabilité dans les Cartes de Saillance en XAI, registered as ANR-25-CE45-1573. It is aligned with the research objectives of EQUR-XSM, and benefits from the collaborations and resources provided by the project.
Note on graftification costs, the administrative fee for graftification is estimated at approximately 600 euros, payable according to the host institution procedures.
Location: the internship will take place at GIPSA-lab, UMR CNRS 5216, Equip ACTIV | 11 rue des Mathématiques, 38402 Saint Martin d’Hères