Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

On-board Federated Learning for Satellite Clusters with Inter-Satellite Links

Nasrin Razmi, Bho Matthiesen, Armin Dekorsy, Petar Popovski

TL;DR

This paper studies the problem of running a federated learning (FL) algorithm within low Earth orbit satellite constellations connected with intra-orbit inter-satellite links (ISL), aiming to efficiently process collected data in situ.

Abstract

The emergence of mega-constellations of interconnected satellites has a major impact on the integration of cellular wireless and non-terrestrial networks, while simultaneously offering previously inconceivable data gathering capabilities. This paper studies the problem of running a federated learning (FL) algorithm within low Earth orbit satellite constellations connected with intra-orbit inter-satellite links (ISL), aiming to efficiently process collected data in situ. Satellites apply on-board machine learning and transmit local parameters to the parameter server (PS). The main contribution is a novel approach to enhance FL in satellite constellations using intra-orbit ISLs. The key idea is to rely on predictability of satellite visits to create a system design in which ISLs mitigate the impact of intermittent connectivity and transmit aggregated parameters to the PS. We first devise a synchronous FL, which is extended towards an asynchronous FL for the case of sparse satellite visits to the PS. An efficient use of the satellite resources is attained by sparsification-based compression the aggregated parameters within each orbit. Performance is evaluated in terms of accuracy and required data transmission size. We observe a sevenfold increase in convergence speed over the state-of-the-art using ISLs, and $10\times$ reduction in communication load through the proposed in-network aggregation strategy.

On-board Federated Learning for Satellite Clusters with Inter-Satellite Links

TL;DR

This paper studies the problem of running a federated learning (FL) algorithm within low Earth orbit satellite constellations connected with intra-orbit inter-satellite links (ISL), aiming to efficiently process collected data in situ.

Abstract

The emergence of mega-constellations of interconnected satellites has a major impact on the integration of cellular wireless and non-terrestrial networks, while simultaneously offering previously inconceivable data gathering capabilities. This paper studies the problem of running a federated learning (FL) algorithm within low Earth orbit satellite constellations connected with intra-orbit inter-satellite links (ISL), aiming to efficiently process collected data in situ. Satellites apply on-board machine learning and transmit local parameters to the parameter server (PS). The main contribution is a novel approach to enhance FL in satellite constellations using intra-orbit ISLs. The key idea is to rely on predictability of satellite visits to create a system design in which ISLs mitigate the impact of intermittent connectivity and transmit aggregated parameters to the PS. We first devise a synchronous FL, which is extended towards an asynchronous FL for the case of sparse satellite visits to the PS. An efficient use of the satellite resources is attained by sparsification-based compression the aggregated parameters within each orbit. Performance is evaluated in terms of accuracy and required data transmission size. We observe a sevenfold increase in convergence speed over the state-of-the-art using ISLs, and reduction in communication load through the proposed in-network aggregation strategy.
Paper Structure (26 sections, 2 theorems, 9 equations, 7 figures, 4 algorithms)

This paper contains 26 sections, 2 theorems, 9 equations, 7 figures, 4 algorithms.

Table of Contents

  1. Introduction
  2. Related Work
  3. Organization
  4. Notation
  5. System Model
  6. Communication Model
  7. Computation Model
  8. Client Operation
  9. Orchestration of Satellite Federated Learning
  10. Synchronous Orchestration
  11. Asynchronous Orchestration
  12. Intra-Orbit Aggregation for Satellite FL
  13. Incremental Aggregation
  14. Parameter Vector Distribution
  15. Predictive Routing
  16. ...and 11 more sections

Key Result

Lemma 1

Consider $L$iid random vectors $\bm X_1, \dots, \bm X_L$ of dimension $n_d$. Let $\tilde{\bm X}_l = \mathsf{Top}_q(\bm X_l)$ for all $l = 1, \dots, L$. Then, the expected number of nonzero elements in $\bm X_\Sigma = \sum_{l=1}^L \tilde{\bm X}_l$ is $n_d - n_d \left( 1 - \frac{n_a}{n_d} \right)^L$,

Figures (7)

  • Figure 1: Global model distribution (left) and collection of local updates (right) using intra-orbit inter-satellite communication.
  • Figure 2: Connectivity towards the ps from within a 60°: 40/5/1 Walker delta constellation. That is, a constellation of 40 satellites having 60° inclined circular orbits and altitude 2000. The satellites are distributed evenly among five orbital planes, which are spaced equidistantly around Earth. Clusters $\mathcal{C}_p$ are defined as either a single satellite per orbital plane or all satellites within an orbital plane. In the second case, a cluster is considered having a connection to the ps if at least one satellite of the cluster can communicate with the ps. Per-satellite connectivity towards the ps is displayed in gray below the cluster connectivity.
  • Figure 3: Routing tree for incremental aggregation in orbital plane $p$. Satellite $k_{p,2}$ acts as sink node. Satellite $k_{p,6}$ has two shortest-path routes to sink. This is resolved unambiguously by the routing algorithm.
  • Figure 4: Test accuracy with respect to wall-clock time. Synchronous orchestration for the MNIST and CIFAR-10 datasets with non-iid distributions, considering both terrestrial and non-terrestrial PS. i.e., the GS in Bremen and an LEO satellite. Note that isl and non-isl stand for the FedAvg with and without isl algorithms respectively.
  • Figure 5: Comparison of synchronous and asynchronous orchestration. The figure displays test accuracy with respect to wall-clock time for a $\text{W-}\Delta$ constellation with CIFAR-10 dataset, distributed iid and non-iid, and PS located in Bremen.
  • ...and 2 more figures

Theorems & Definitions (2)

  • Lemma 1
  • Proposition 1