Search

Browse

• Communities
• Authors
• Year
• Titles
• Subjects

My Space

• Login

Federated learning at TensorFlow Using the geometric approach

Frangias Georgios

Full record

---

Summary

The rapid growth of data generation and internet usage in recent years has created an unprecedented demand for efficient Big Data collection, processing and analysis. The ever-growing privacy concerns of the public opinion and the enactment of regulations on this subject, induce the need for the development of decentralized, distributed and scalable Machine Learning mechanisms, that can assure both personal data security and high accuracy collective training. The scientific field of Federated Learning is dedicated to achieving exactly that; train a global machine learning model without communicating sensitive locally generated data. For the purpose of the current thesis, we have developed a deployable extension to the Distributed Machine Learning library KungFu, to effortlessly execute Federated Learning training jobs on decentralized compute nodes. The implemented algorithms are the three Functional Dynamic Averaging methods, inspired by the Geometric Approach. These algorithms have the ability to approximately monitor a global threshold function, using solely local data and, subsequently, dynamically determine the need for synchronization and model aggregation. We have put our implementation to the test by executing exhaustive experiments on multi-node GPU infrastructure, and compared it to a classic distributed algorithm. The results demonstrate a significant training time reduction, due to reduced communication overhead, without having repercussions on accuracy, especially for non-ideal network topologies.

Available Files

Services