Filling the Gap: Fault-Tolerant Updates of On-Satellite Neural Networks Using Vector Quantization
Key: KDSOS-1
Author: Olga Kondrateva, Stefan Dietzel, Maximilian Schambach, Johannes Otterbach, Björn Scheuermann
Date: 2023
Kind: In proceedings
Publisher: IEEE
Book title: Proceedings of the 2023 IFIP Networking Conference
Abstract: The use of small, low-Earth-orbit satellites enables many novel Earth observation use cases due to their cost efficiency. To cope with the challenging communication environment, machine learning algorithms, such as artificial neural networks, can be applied onboard the satellites. They help to prioritize or pre-process sensor measurements and to reduce the amount of data transmitted to Earth. However, transferring and updating machine learning models to suit changing prioritization requirements poses a number of challenges in itself due to short contact times of satellites with ground stations and lossy communication links. We propose a new transmission mechanism for model updates that retains high performance even when these updates have been only partially transmitted. We achieve this by approximating missing model weights using a vector quantization approach. Using a support structure of quantized vector indices, we can approximate the model with a small amount of data, which is transmitted first, while retaining a high performance. The model performance can then be incrementally improved, as more exact model weights are transmitted to the satellites. Our evaluation shows that this approach significantly outperforms existing baselines.

The documents distributed by this server have been provided by the contributing authors as a means to ensure timely dissemination of scholarly and technical work on a non-commercial basis. Copyright and all rights therein are maintained by the authors or by other copyright holders, not withstanding that they have offered their works here electronically. It is understood that all persons copying this information will adhere to the terms and constraints invoked by each author's copyright. These works may not be reposted without the explicit permission of the copyright holder.