Yaël Frégier

Astrid Klipfel, Yaël Frégier, Adlane Sayede et al.

One of the greatest challenges facing our society is the discovery of new innovative crystal materials with specific properties. Recently, the problem of generating crystal materials has received increasing attention, however, it remains unclear to what extent, or in what way, we can develop generative models that consider both the periodicity and equivalence geometric of crystal structures. To alleviate this issue, we propose two unified models that act at the same time on crystal lattice and atomic positions using periodic equivariant architectures. Our models are capable to learn any arbitrary crystal lattice deformation by lowering the total energy to reach thermodynamic stability. Code and data are available at https://github.com/aklipf/GemsNet.

Astrid Klipfel, Yaël Frégier, Adlane Sayede et al.

Graph neural networks are widely used in machine learning applied to chemistry, and in particular for material science discovery. For crystalline materials, however, generating graph-based representation from geometrical information for neural networks is not a trivial task. The periodicity of crystalline needs efficient implementations to be processed in real-time under a massively parallel environment. With the aim of training graph-based generative models of new material discovery, we propose an efficient tool to generate cutoff graphs and k-nearest-neighbours graphs of periodic structures within GPU optimization. We provide pyMatGraph a Pytorch-compatible framework to generate graphs in real-time during the training of neural network architecture. Our tool can update a graph of a structure, making generative models able to update the geometry and process the updated graph during the forward propagation on the GPU side. Our code is publicly available at https://github.com/aklipf/mat-graph.

Benoît Brebion, Alban Gallard, Katrin Sippel et al.

Background and objective: Brain activity in premature newborns has traditionally been studied using electroencephalography (EEG), leading to substantial advances in our understanding of early neural development. However, since brain development takes root at the fetal stage, a critical window of this process remains largely unknown. The only technique capable of recording neural activity in the intrauterine environment is fetal magnetoencephalography (fMEG), but this approach presents challenges in terms of data quality and scarcity. Using artificial intelligence, the present research aims to transfer the well-established knowledge from EEG studies to fMEG to improve understanding of prenatal brain development, laying the foundations for better detection and treatment of potential pathologies. Methods: We developed an unpaired diffusion translation method based on dual diffusion bridges, which notably includes numerical integration improvements to obtain more qualitative results at a lower computational cost. Models were trained on our unpaired dataset of bursts of spontaneous activity from 30 high-resolution premature newborns EEG recordings and 44 fMEG recordings. Results: We demonstrate that our method achieves significant improvement upon previous results obtained with Generative Adversarial Networks (GANs), by almost 5% on the mean squared error in the time domain, and completely eliminating the mode collapse problem in the frequency domain, thus achieving near-perfect signal fidelity. Conclusion: We set a new state of the art in the EEG-fMEG unpaired translation problem, as our developed tool completely paves the way for early brain activity analysis. Overall, we also believe that our method could be reused for other unpaired signal translation applications.

Yaël Frégier, Jean-Baptiste Gouray

Training generative adversarial networks (GANs) on high quality (HQ) images involves important computing resources. This requirement represents a bottleneck for the development of applications of GANs. We propose a transfer learning technique for GANs that significantly reduces training time. Our approach consists of freezing the low-level layers of both the critic and generator of the original GAN. We assume an autoencoder constraint in order to ensure the compatibility of the internal representations of the critic and the generator. This assumption explains the gain in training time as it enables us to bypass the low-level layers during the forward and backward passes. We compare our method to baselines and observe a significant acceleration of the training. It can reach two orders of magnitude on HQ datasets when compared with StyleGAN. We prove rigorously, within the framework of optimal transport, a theorem ensuring the convergence of the learning of the transferred GAN. We moreover provide a precise bound for the convergence of the training in terms of the distance between the source and target dataset.