One of the eyes of the Mars smart droid Perseverance — NASA
Artificial intelligence (AI) has the potential to revolutionize space exploration by delegating some spacecraft decisions to on-board AI, rather than relying on ground control or predefined procedures.
The spacecraft will likely have an AI/ML processing unit running an inference engine, and the neural network will be pre-installed with parameters that can be updated on board the spacecraft by uploading parameters obtained during ground training via remote command, although satellite uplink bandwidth is limited and transmissions can be costly.
Furthermore, missions operating with suboptimal neural networks will miss out on valuable science data. Therefore, smaller networks would reduce uplink costs and increase the value of the science data downloaded. In this work, we evaluate and discuss the use of minimalist neural networks with reduced precision to reduce upload times.
As an example of an AI use case, we focus on NASA's Magnetospheric Multiscale (MMS) mission. We show how onboard AI can be used within Earth's magnetosphere to classify data and selectively downlink more valuable data, as well as recognize regions of interest and trigger burst modes to collect data at higher rates. We show how simple filtering schemes and algorithms can be used to detect the start and end of regions of interest in the stream of classifications.
To provide classification, we use an established Convolutional Neural Network (CNN) trained to over 94% accuracy. We also show how to reduce the network to a single linear layer and train it to the same accuracy as the established CNN, reducing the overall size of the model by up to 98.9%.
Furthermore, we show how by using a lower precision format to represent network parameters, we can reduce each network by up to 75% of its original size, with less than 0.6 percentage points change in accuracy.
Jonah Ekelund, Ricardo Vinuesa, Yuri Kotiansev, Pierre Henry, Jean-Luca Delzano, Stefano Marchidis
Subjects: Artificial Intelligence (cs.AI); Measurements and Methods for Astrophysics (astro-ph.IM)
Source: arXiv:2406.14297 [cs.AI] (or arXiv:2406.14297v1 [cs.AI] For this version
https://doi.org/10.48550/arXiv.2406.14297
Focus on learning more
Submission History
Source: Jonah Ekelund
[v1] June 20, 2024 (Thu) 13:24:52 UTC (830 KB)
https://arxiv.org/abs/2406.14297
Astrobiology, AI, artificial intelligence, machine learning, deep learning, neural networks,
