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

Laboratoire/Entreprise : LIFAT EA 6300 (lifat.univ-tours.fr)
Durée : 3 ans
Contact : donatello.conte@univ-tours.fr
Date limite de publication : 2022-02-24

Contexte :
The thesis will be part and funded (gross salary: 2 000 € approximately) under the ANR project CodeGNN (http://www.normastic.fr/projet-anr-codegnn/)
The PhD could start around October 2022

Supervisors and contact
Donatello Conte (University of Tours, France) donatello.conte@univ-tours.fr
Sébastien Bougleux (Université de Caen Normandie, France) bougleux@unicaen.fr
Nicolas Ragot (University of Tours, France) nicolas.ragot@univ-tours.fr

Ph.D. Position (10/2022): Learning Spatio-temporal data by graph representations

Sujet :
In many application domains like action recognition or prediction, video segmentation, traffic forecasting or anomaly detection in brain activity signals, time-varying data are frequently represented by graphs. Two main representations are commonly considered: a temporal sequence of graphs or a spatio-temporal graph connecting graph nodes through time. While there is a solid literature on data analysis based on such representations, the domain has strongly evolved over the last 5 years with the advances in deep learning on Graph Neural Networks.
Such methods have been less investigated for time-varying graphs, particularly when both the graph structure and the data attached to this structure are varying.
We can distinguish two main models: Recurrent Neural Networks (RNN) combined with spatial convolutions rely on the sequential representation [1, 2, 3]; or Graph Convolutional Networks alternating temporal and spatial convolutions [4, 5, 6].
The aim of this thesis is to:
1. Study new representations for spatio-temporal graphs: we want to investigate some new representations in two main directions: representing temporal data as attributes of nodes and edges, and representing temporal data as edge connections between spatial positions represented by nodes at different times.
2. Propose new Neural Network architectures for data represented by this kind of graphs: we want to propose adapted convolutions, decimation and pooling, and study the definition of a recurrent neural network that operates directly in the space of the graphs (for example generating new graphs). One direction of study will also be the Spatial- Temporal Graph Attention Networks (STGAT [7]) and Graph Transformer Networks (GTN [8]).
3. Program these models (in Python), and compare them to the state-of-the-art on standard datasets for different applications, in particular, skeleton-based gesture recognition.

References
[1] A. Jain, A. R. Zamir, S. Savarese, and A. Saxena, “Structural-rnn: Deep learning on spatio- temporal graphs,” CoRR, vol. abs/1511.05298, 2015
[2] C. Si, W. Chen, W. Wang, L. Wang, and T. Tan, “An attention enhanced graph convolutional lstm network for skeleton-based action recognition,” in Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 1227–1236, 2019
[3] Li, M., Chen, S., Zhao, Y., Zhang, Y., Wang, Y., & Tian, Q. (2020). Dynamic multiscale graph neural networks for 3d skeleton based human motion prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 214-223).
[4] B. Yu, H. Yin, and Z. Zhu, “Spatio-temporal graph convolutional neural network: A deep learning framework for traffic forecasting,” CoRR, vol. abs/1709.04875, 2017.
[5] L. Shi, Y. Zhang, J. Cheng, and H. Lu, “Skeleton-based action recognition with directed graph neural networks,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 7912–7921, 2019.
[6] Chen, T., Zhou, D., Wang, J., Wang, S., Guan, Y., He, X., & Ding, E. (2021). Learning Multi-Granular Spatio-Temporal Graph Network for Skeleton-based Action Recognition. arXiv preprint arXiv:2108.04536.
[7] Kong, X., Xing, W., Wei, X., Bao, P., Zhang, J., & Lu, W. (2020). STGAT: Spatial-temporal graph attention networks for traffic flow forecasting. IEEE Access, 8, 134363-134372.
[8] Yun, S., Jeong, M., Kim, R., Kang, J., & Kim, H. J. (2019). Graph transformer networks. Advances in neural information processing systems, 32.

Profil du candidat :
– Master degree in Computer Science, Applied Mathematics, Data Science, or similar.

Formation et compétences requises :
– strong background in computer science and maths
– experiences in neural networks, deep learning, Python programming,
numerical analysis will be privileged
– knowledge in video and image analysis would be appreciated
– good communication skills and reporting, autonomy and curiosity

Adresse d’emploi :
LIFAT, 64 Avenue Jean Portalis, 37200 Tours

Document attaché : 202202231030_PhD_Thesis_Proposal_SpatioTemporalGraphs.pdf