Deep Graph Representation Learning on non-uniform 3D objects

25/04/2024 all-day

Offre en lien avec l’Action/le Réseau : – — –/Doctorants

Laboratoire/Entreprise : CRIL
Durée : 3
Contact :
Date limite de publication : 2024-04-25

Contexte :
Machine learning involves leveraging data to extract mathematical models capable
of generalizing or describing this data according to predefined objectives. This data
comes in various forms, ranging from well-defined structures like images and
matrices to semi-structured formats such as text and graphs. However, dealing with
entirely unstructured data, such as non-uniform 3D objects, poses a challenge for
traditional methods that primarily focus on geometric analysis.
The development of artificial intelligence, particularly deep learning, has greatly
improved performance compared to conventional learning methods, especially when
it comes to textual data, images, graphs, sequences, and more. However, learning on
non-uniform 3D objects remains a significant challenge. This field is garnering
increasing interest in various applications such as predicting molecular properties
based on their 3D structures rather than textual features [2].
In the field of bioinformatics, protein annotation based on their 3D interactions is an
example [1], as is the use of 3D structures in physics to simulate objects [3] or body
parts to analyze their behavior. These applications demonstrate the utility of
analyzing or learning on non-uniform 3D objects, thus sparking considerable interest
within the scientific community.

Sujet :
This thesis focuses on deep learning, with an emphasis on learning graph
representations. Graphs are widely used in many applications, providing a versatile
representation for non-regular objects, including 3D meshes, as an alternative to
traditional methods such as CNNs or image segmentation models like U-net. This
thesis explores graph neural networks (GNNs) for modeling non-regular 3D objects,
such as 3D meshes. Unlike CNNs, GNNs are designed to handle graph-type data,
making them more suitable for representing 3D meshes. They have demonstrated
superior performance in modeling such data, offering a promising alternative to
existing methods. However, despite their effectiveness, GNNs face scalability
challenges, especially with complex meshes. This thesis proposes solutions to
overcome these challenges by exploring mesh-specific pooling methods and other
strategies to simplify learning. It also considers approaches for constructing graphs
from 3D meshes to enhance learning efficiency. In addition to the static aspect of
data, this thesis addresses the application of GNNs to data with temporal patterns or features. It explores their uses in domains such as fluid simulation, weather
modeling, and 3D medical imaging, as well as in physical simulation of 3D meshes.
This highlights the temporal evolution of meshes in both space and time.

[1] Laveglia, V., Giachetti, A., Sala, D., Andreini, C., & Rosato, A. (2022). Learning to Identify
Physiological and Adventitious Metal-Binding Sites in the Three-Dimensional Structures of
Proteins by Following the Hints of a Deep Neural Network. Journal of Chemical Information
and Modeling, 62(12), 2951-2960.
[2] Yang, Y., Yao, K., Repasky, M. P., Leswing, K., Abel, R., Shoichet, B. K., & Jerome, S. V.
(2021). Efficient exploration of chemical space with docking and deep learning. Journal of
Chemical Theory and Computation, 17(11), 7106-7119.
[3] Atz, K., Grisoni, F., & Schneider, G. (2021). Geometric deep learning on molecular
representations. Nature Machine Intelligence, 3(12), 1023-1032.
[4] Cao, Y., Chai, M., Li, M., & Jiang, C. (2023, July). Efficient learning of mesh-based physical
simulation with bi-stride multi-scale graph neural network. In International Conference on
Machine Learning (pp. 3541-3558). PMLR.
[5] Fahim, G., Amin, K., & Zarif, S. (2022). Enhancing single-view 3D mesh reconstruction with
the aid of implicit surface learning. Image and Vision Computing, 119, 104377.

Profil du candidat :
Ideally, the recruited person will hold a Master’s degree in computer science and
have theoretical and practical knowledge in deep learning. Experience of machine
learning on graphs is also desirable but not essential The candidate must
● Programming skills, such as proficiency in Python, for example
● Experience in Deep Learning, data mining
● Synthesis and writing skills allowing for clear and effective reporting of work

Formation et compétences requises :

Adresse d’emploi :
Computer science Research Institute of Lens (CRIL), Lens, France

Document attaché : 202404091446_Deep Graph Representation Learning on non-uniform 3D objects.pdf