Gradient compression with variable-rate quantization for federated learning under heterogeneous bandwidth
Research division · CDS
Computational & Data Systems
Machine learning, probabilistic inference, high-performance computing and human-centred interfaces — the analytical substrate running through every division at Veyra.
Overview
The Computational & Data Systems division was established in 2001 from what had been an informal computing services group. Its founding premise — that computational methods are not support infrastructure but core science — has shaped the division's culture ever since. CDS researchers do not provide analysis as a service to other groups; they participate in scientific projects as equal contributors, and they publish the methods they develop independently of any particular application domain.
The division spans three modes of work. The Distributed Learning Systems Group focuses on the engineering of training pipelines that scale reliably across hundreds or thousands of compute nodes, including the heterogeneous, fault-prone clusters typical of federated settings. The Probabilistic Inference Lab builds Bayesian computational tools for inverse problems that arise across MME, ECS and CNS — parameter estimation in physical simulators, uncertainty propagation through multi-stage pipelines, and calibrated decision support. The Human-Centred Computing Group studies how scientists interact with the outputs of complex systems and designs interfaces that make high-dimensional, uncertain data interpretable without flattening it.
CDS is the primary user of the Meridian High-Performance Computing Cluster, which the division co-designed and which now serves all of Veyra. Division researchers also contribute to the Institute's Software & AI Services, including the Veyra Atlas materials-property predictor.
Research themes
- Communication-efficient distributed training — gradient compression, asynchronous protocols and convergence guarantees under partial participation.
- Scientific Bayesian inference — variational methods, sequential Monte Carlo and probabilistic programming for coupled physical-statistical models.
- Federated and privacy-preserving learning — Byzantine-resilient aggregation, differential privacy and secure multi-party computation.
- Explainability and sensemaking — adaptive interfaces, uncertainty visualization and interactive model interrogation for domain scientists.
- AI for materials and climate — ML-driven property prediction, surrogate models for expensive simulations, active learning for experiment design.
Research groups
Three groups, each with a distinct technical focus and its own publication output.
Distributed Learning Systems Group
Training and deploying large models across heterogeneous compute nodes, with a focus on communication-efficient algorithms and fault-tolerant federated learning.
PI Dr. Naila Ravelo
Group page Probabilistic Inference Lab
Bayesian methods for scientific inverse problems, uncertainty quantification and probabilistic programming systems that interoperate with physical simulators.
PI Prof. Tomas Eberhardt
Group page Human-Centred Computing Group
Interaction design for complex scientific interfaces, adaptive explanation systems and the study of how researchers make sense of high-dimensional outputs.
PI Dr. Imara Solveig
Group page Selected publications
Amortized variational inference for coupled physical-statistical models
Legibility under load: adaptive interface density in real-time data exploration
Byzantine-resilient aggregation without full participation: a majority-of-gradients approach
Sequential Monte Carlo for high-dimensional geophysical inversion