Bart Cox

Nomad is the first dynamic client transfer framework for multi-server FL, reallocating clients based on network conditions and data alignment to reduce latency and improve learning. Unlike static assignments, Nomad enables flexible migration during training. Experiments show accuracy improvements of up to 31.8 points in join-only settings and 18.8 points under churn, consistently surpassing strong baselines and scaling well across geographic deployments.

GWTF is the first practical, crash-tolerant decentralized framework for collaboratively training LLMs on heterogeneous volunteer clients. It handles node churn and unstable networks through a novel decentralized flow algorithm that optimizes microbatch routing. Evaluations on GPT- and LLaMa-like models show that GWTF reduces training time by up to 45% in challenging, geographically distributed settings.

We propose an asynchronous, Byzantine-resilient FL algorithm that avoids straggler delays and requires no server dataset. By updating after a safe number of client contributions, it outperforms state-of-the-art methods, achieving faster training and higher accuracy under multiple attack types.

Morph is a decentralized learning topology optimizer that adapts peer selection based on model dissimilarity to overcome non-IID data and static communication limits. By reshaping the graph through gossip-based discovery, it boosts robustness and performance. Experiments on CIFAR-10 and FEMNIST show Morph outperforming static and epidemic baselines, achieving higher accuracy, faster convergence, and more stable learning with fewer communication rounds.

This work extends two classical reliable communication protocols to combine authenticated links and processes, introducing DualRC. It leverages trusted nodes (e.g., gateways) and components (e.g., Intel SGX) to improve communication reliability, with methods to validate network implementation.

Spyker is the first fully asynchronous multi-server FL system, eliminating server idle time and single-server bottlenecks. Clients communicate only with their nearest server, while servers also update each other asynchronously. This continuously active design improves scalability and performance across MNIST, CIFAR-10, and WikiText-2.

We introduce a federated diffusion framework that allows independent, privacy-preserving training of DDPMs without exposing local data. By adapting FedAvg and leveraging the UNet backbone efficiently, our method cuts parameter exchange by up to 74% compared to naive FedAvg, while preserving image quality close to centralized training, as measured by FID.

We present the first fully configurable framework for Federated Continual Learning, designed to reproduce complex, evolving learning scenarios. It supports large-scale deployments via containerization and Kubernetes, enabling precise experimentation. Demonstrations on CIFAR-100 and heterogeneous task sequences show Freddie’s effectiveness and uncover persistent performance challenges in real FCL settings.

To speed up the Federated Learning process, learning tasks can be offloaded to other clients. Using similarity metrics and a resource aware scheduler, we are able to speed up the training process for Federated Learning.

Masa is a memory-aware multi-DNN scheduling framework for edge devices that ensures low response times without modifying models. It leverages inter/intra-network dependencies and context to cut latency by up to 90% on low-memory devices.

The paper introduces EdgeCaffe, a framework for exploring scheduling policies in multi-inference DNN jobs on resource-constrained edge devices. It proposes MemA, a memory-aware policy that improves execution time by up to 5x without additional resources, based on layer-specific memory demands.

Masa, a responsive memory-aware multi-DNN execution framework, an on-device middleware featuring on modeling inter- and intra-network dependency and leveraging complimentary memory usage of each layer.

Force processes to swap memory (Ubuntu 18.04) When benchmarking processes, looking how a process perform when swapping memory can be useful. Since the operating system determines the memory allocation policy, testing with limited memory can be cumbersome. One could try to fill the virtual memory with other processes to limit the available memory but since the operating system tries to swap inactive processes, our benchmarked process will probably remain in virtual memory. In any case, we cannot know this for certain. ...

The GDelt database 1, on 18-11-2019, consists of 492,618 segments. Processing the top 10 most mentioned topics for each date on the whole data set would take a long time. Using clusters in the cloud, like AWS EMR, significantly decreases up the needed computation time but might be costly. To make the best use of the clusters on AWS EMR a minimization of the idle time of the machines is desired. ...