3IA PhD/Postdoc Seminar #45

  • Research
Published on December 10, 2024 Updated on May 14, 2025
Dates

on the May 16, 2025

from 10:30 am to 12:00 pm
Location
Inria Sophia Antipolis

Monthly PhD and Postdoc seminar

Program

 

10:30
Madina Bekbergenova  (Université Côte d'Azur - ICN, PhD)

Flash presentation: MetaboT: AI-based agent for natural language-based interaction with metabolomics knowledge graphs


Abstract: MetaboT is an AI system that leverages large language models to translate natural language questions into SPARQL queries for metabolomics knowledge graphs (KGs). Built with LangChain and LangGraph, it uses specialized agents to extract entities and chemical taxonomies, construct ontology-informed queries, and retrieve structured results from a public plant metabolomics KG. Evaluated on 50 questions, MetaboT achieved 84% accuracy, vastly outperforming a GPT-4o baseline (8%), highlighting the advantage of its multi-agent design. By automating SPARQL query generation and execution, MetaboT lowers technical barriers, enabling researchers to easily access and explore complex metabolomics data.
Access it.

10:30 - 11:00
Xufeng Zhang (Inria - Team Neo, PhD)

Memory-efficient online caching Policies with regret guarantees

Abstract: Online learning algorithms provide robust performance in caching problems but require substantial memory to store per-file historical data, limiting their scalability to large-catalog systems. To overcome this challenge, we propose a dimensionality reduction algorithm based on the Follow-the-Perturbed-Leader framework and the Johnson-Lindenstrauss lemma. Our method significantly reduces memory consumption while preserving sublinear regret, making it well-suited for caching under resource constraints. Experiments on both synthetic and real-world traces demonstrate its advantages over other memory-efficient approaches.

11:00 - 11:30
Cloé Mahé (Université Côte d'Azur, PhD)

First robust census of binary supermassive black holes : Deep Learning techniques applied to the Euclid space mission.

Abstract: The formation of dual and binary supermassive black holes (SMBH) is a solid prediction of the current hierarchical model of galaxy formation [1]. By studying these objects, several astrophysical questions can be addressed, related to the galaxy mass build-up, the feedback of active galaxy nuclei (AGN) on the star formation history of galaxies, or the expected amount of gravitational waves due to the final merging process of the SMBH pair [2]. Detecting such rare systems require simultaneously high angular resolution capabilities and a large surveyed sky area. The ongoing ESA space mission Euclid is groundbreaking in both regards, and offers therefore an unprecedented opportunity to detect and study numerous dual systems. In this talk, I will present the first part of a methodology designed to automatically identify dual systems in the Euclid surveys, based on Machine Learning (ML) and Deep Learning (DL) techniques. The identification of the galaxies harbouring double or dual SMBH in the Euclid images will be done using a Convolutional Neural Network (CNN). As observations of genuine dual objects are scarce, we rely on the objects tagged as dual AGN in the Horizon-AGN cosmological simulation [3] to construct the CNN training database. However, since there are still very few identified systems, it is first necessary to increase the number of such images from the Horizon-AGN dataset.
To do this data augmentation process, I constructed a Variational Autoencoder (VAE) [4] that I combined to a normalizing flow [5], allowing one to generate synthetic images of dual AGN with similar properties (total flux, Gini shape parameter) and visual aspects as the initial images extracted from the simulation. I will present the technical properties of this model, including its structuration, the definition of losses, and the diagnostics used to inquire its capabilities. Using this flow-VAE model, producing a large set of images of dual AGN fulfilling a specified distribution of physical properties is feasible in a fragment of the time necessary to generate such images from the Horizon-AGN simulation.
References

11:30 - 12:00

Open discussion about the two contributions

More information


Event reserved for 3IA Côte d'Azur PhD students and post-docs. ID check at the entrance of the site with visual bag inspection.