Evaluation of Edge AI Co-Processing Methods for Space Applications
Vicenzi, Luca (2023-10-14)
Evaluation of Edge AI Co-Processing Methods for Space Applications
(14.10.2023)
Lataukset:
Julkaisu on tekijänoikeussäännösten alainen. Teosta voi lukea ja tulostaa henkilökohtaista käyttöä varten. Käyttö kaupallisiin tarkoituksiin on kielletty.
avoin
Julkaisun pysyvä osoite on:
https://urn.fi/URN:NBN:fi-fe20231023140903
https://urn.fi/URN:NBN:fi-fe20231023140903
Tiivistelmä
The recent years spread of SmallSats offers several new services and opens to the implementation of new technologies to improve the existent ones. However, the communication link to Earth in order to process data often is a bottleneck, due to the amount of collected data and the limited bandwidth.
A way to face this challenge is edge computing, which supposedly discards useless data and fasten up the transmission, and therefore the research has moved towards the study of COTS architectures to be used in space, often organized in co-processing setups.
This thesis considers AI as application use case and two devices in a controller-accelerator configuration. It proposes to investigate the performances of co-processing methods such as simple parallel, horizontal partitioning and vertical partitioning, for a set of different tasks and taking advantage of different pre-trained models.
The actual experiments regard only simple parallel and horizontal partitioning mode, and they compare latency and accuracy results with single processing runs on both devices.
Evaluating the results task-by-task, image classification has the best performance improvement taking advantage of horizontal partitioning, with a clear accuracy improvement, as well as semantic segmentation, which shows almost stable accuracy and potentially higher throughput with smaller models input sizes. On the other hand, object detection shows a drop in performances, especially accuracy, which could maybe be improved with more specifically developed models for the chosen hardware.
The project clearly shows how co-processing methods are worth of being investigated and can improve system outcomes for some of the analyzed tasks, making future work about it interesting.
A way to face this challenge is edge computing, which supposedly discards useless data and fasten up the transmission, and therefore the research has moved towards the study of COTS architectures to be used in space, often organized in co-processing setups.
This thesis considers AI as application use case and two devices in a controller-accelerator configuration. It proposes to investigate the performances of co-processing methods such as simple parallel, horizontal partitioning and vertical partitioning, for a set of different tasks and taking advantage of different pre-trained models.
The actual experiments regard only simple parallel and horizontal partitioning mode, and they compare latency and accuracy results with single processing runs on both devices.
Evaluating the results task-by-task, image classification has the best performance improvement taking advantage of horizontal partitioning, with a clear accuracy improvement, as well as semantic segmentation, which shows almost stable accuracy and potentially higher throughput with smaller models input sizes. On the other hand, object detection shows a drop in performances, especially accuracy, which could maybe be improved with more specifically developed models for the chosen hardware.
The project clearly shows how co-processing methods are worth of being investigated and can improve system outcomes for some of the analyzed tasks, making future work about it interesting.