arXiv preprint
Implicit Regularization of Mini-Batch Training in Graph Neural Networks
Clement Wang, Antoine Vialle, Robin Vaysse, Thomas Bonald
- May 2026
- Under review
- First author
Training a GNN on random subgraphs is usually seen as an approximation of full-graph training. This paper shows it can be better: the choice of graph sampler acts as an implicit regularizer.
- Random Node Sampling matches or beats full-graph training on 8 of 10 datasets.
- It runs 2 to 12 times faster, with up to 3 times less GPU memory.
- Backward error analysis: mini-batch SGD implicitly minimizes the sampled loss plus a penalty on the mini-batch gradient variance.
BiDS 2023
Leveraging Multi-Temporal Sentinel 1 and 2 Satellite Data for Leaf Area Index Estimation With Deep Learning
Clement Wang, Antoine Debouchage, Valentin Goldité, Aurélien Wery, Jules Salzinger
- Nov 2023
- Published, peer-reviewed
- First author
Predicting Leaf Area Index, a key indicator of vegetation health, for every pixel of a satellite image, by combining Sentinel-1 radar and Sentinel-2 multispectral images taken at several dates.
- Several U-Nets are pretrained to map each input into a shared latent space, then fine-tuned end to end with a common decoder that accounts for seasonality.
- RMSE of 0.06 and R² of 0.93 on public data.
