Studying the impact of data heterogeneity in Split Learning. Specifically, we conduct systematic analysis with Deep Neural Networks (ResNet, VGG, MobileNet) using various data-sets. As a result, we indetified the existence of catastrophic forgetting (CF) in the training. Finally, we proposed a new ML solution for tackling CF caused by non-IID data.

In this work, we propose SplitPipe, a Machine Learning as a Service (MLaaS) modular and extensible framework for collaborative and distributed training. SplitPipe processes high-level tasks (e.g., with the model’s description that will be trained) and orchestrates the training process based on a novel Split Learning (SL) protocol. Additionally, SplitPipe supports multihop SL-based training that enhances data privacy and relaxes memory demands.

In this work, we consider a parallel SL system with multiple helper nodes. Specifically, we focus on orchestrating the workflow of this system, which is critical in highly heterogeneous systems. In particular, we formulate the joint problem of client-helper assignments and scheduling decisions to minimize the training makespan. We propose a solution method based on the decomposition of the problem by leveraging its inherent symmetry.

Developed a distributed system consisting of a server in the cloud and multiple servers on edge nodes. Each edge-node is located near a group of drones, with direct access to them. Edge-nodes can process the generated data in parallel and independently of each other. The system offers users a shell interface through which one can initiate tasks to specific edge nodes and afterwards combine the results. The communication between the server and the edges is done without any user intervention. Also, created an estimation model using metrics that were extracted from experimental testing.

Build multiple distributed computing systems during Bachelor's and Master's studies. Some indicative examples are: Distributed computing environment with transparent migration and load balancing, distributed system for Uniform Reliable multicast communication with synchronous view.