Apprentissage sur des graphes attribués
Event details
| Date | 22.06.2023 |
| Hour | 13:30 › 15:30 |
| Speaker | Odilon Duranthon |
| Category | Conferences - Seminars |
EDIC candidacy exam
Exam president: Prof. Nicolas Macris
Thesis advisor: Prof. Lenka Zdeborová
Co-examiner: Prof. Pascal Frossard
Abstract
In the frame of inference on graphs growing interest has been given to synthetic datasets generated by tractable models. For such models the average performance of simple graph neural networks (GNNs) can be predicted and an optimal algorithm based on belief-propagation (BP) can be derived. We derive the optimal solution for the most commonly used model and we show that state-of-the-art GNNs are far from the optimality. We propose to build an architecture inspired by BP and we hope it will close the gap.
Background papers
1. « Contextual Stochastic Block Models » https://dspace.mit.edu/bitstream/handle/1721.1/138073/8077-contextual-stochastic-block-models.pdf
2. « Statistical Mechanics of Generalization In Graph Convolution Networks »https://arxiv.org/abs/2212.13069
3. « AMP-Inspired Deep Networks for Sparse Linear Inverse Problems » https://arxiv.org/abs/1612.01183
Exam president: Prof. Nicolas Macris
Thesis advisor: Prof. Lenka Zdeborová
Co-examiner: Prof. Pascal Frossard
Abstract
In the frame of inference on graphs growing interest has been given to synthetic datasets generated by tractable models. For such models the average performance of simple graph neural networks (GNNs) can be predicted and an optimal algorithm based on belief-propagation (BP) can be derived. We derive the optimal solution for the most commonly used model and we show that state-of-the-art GNNs are far from the optimality. We propose to build an architecture inspired by BP and we hope it will close the gap.
Background papers
1. « Contextual Stochastic Block Models » https://dspace.mit.edu/bitstream/handle/1721.1/138073/8077-contextual-stochastic-block-models.pdf
2. « Statistical Mechanics of Generalization In Graph Convolution Networks »https://arxiv.org/abs/2212.13069
3. « AMP-Inspired Deep Networks for Sparse Linear Inverse Problems » https://arxiv.org/abs/1612.01183
Practical information
- General public
- Free