Random sampling for trustworthy pooling layers in graph neural networks

Informations projet

Rand4TrustPool

N° de convention de subvention: 101148828

DOI

10.3030/101148828

Date de signature de la CE 25 Avril 2024

Date de début 10 Mars 2025

Date de fin 9 Mars 2027

Financé au titre de

• Marie Skłodowska-Curie Actions (MSCA)

Coût total

Aucune donnée

Contribution de l’UE

€ 226 751,04

Coordonné par

UNIVERSITETET I TROMSOE - NORGES ARKTISKE UNIVERSITET
Norway

Objectif

Graph neural networks (GNNs) are deep learning architectures that hold the promise of durably changing our way of, e.g. predicting inter-molecular chemical affinity to accelerate drug discovery or smartly balancing power loads on the electricity grid. My project focuses on pooling layers, a major building block of neural networks, including GNNs. In a nutshell, pooling layers generate local summaries of the data to enable faster computations and generate discriminative multiscale representations. The traditional periodic-sampling-based pooling that has proven so successful for data defined on regular grids such as time series or images becomes ill-defined when the underlying space of the data loses its regularity such as data defined over graphs. Todays existing solutions stem from a trial-and-error approach often driven by two practical and somewhat short-sighted objectives: computation efficiency and empirical predictive performance. The aim of my project is to add trustworthiness as a main objective in the design of novel pooling layers for GNNs. To achieve this, I will bridge the fields of signal processing, point processes, and random numerical linear algebra with graph machine learning to incorporate random methods (mainly random sampling and random projections) into the design of pooling layers for GNNs. Indeed, randomized approaches are provably well adapted to three pillars of trustworthiness: technical robustness, transparency, and privacy. The outcome of my project will be an extensively tested, theoretically-grounded, trustworthy, ready-to-be-deployed, open-source, GNN-based toolbox, along with its user guide for practitioners. It will be the first GNN architecture benefiting from the theoretical guarantees random pooling layers come with. Such trustworthy GNNs will have an impact in the increasingly many fields in which GNNs are pushing the state-of-the-art: including circuit chip design, physics, drug discovery and network neuroscience.

Champ scientifique (EuroSciVoc)

• sciences naturelles › sciences biologiques › neurobiologie
• sciences naturelles › mathématiques › mathématiques pures › algèbre

Mots‑clés

• graph machine learning
• graph neural networks
• trustworthy AI
• pooling
• sampling
• graph signal processing
• point processes
• random numerical linear algebra

Programme(s)

• HORIZON.1.2 - Marie Skłodowska-Curie Actions (MSCA) PROGRAMME PRINCIPAL
  

Thème(s)

• HORIZON-MSCA-2023-PF-01-01 - MSCA Postdoctoral Fellowships 2023
  

Régime de financement

HORIZON-TMA-MSCA-PF-EF - HORIZON TMA MSCA Postdoctoral Fellowships - European Fellowships

Appel à propositions

HORIZON-MSCA-2023-PF-01

Coordinateur

UNIVERSITETET I TROMSOE - NORGES ARKTISKE UNIVERSITET

• Contribution nette de l'UE: € 226 751,04
• Adresse: HANSINE HANSENS VEG 14 ; 9019 Tromso ; Norvège
• Type d’activité: Higher or Secondary Education Establishments
• Coût total: Aucune donnée