State-of-the-Art Edge Artificial Intelligence (AI) is currently mostly targeted at a train-then-deploy paradigm: edge devices are exclusively responsible for inference, whereas training is delegated to data centers, leading to high energy and CO2 impact. On-Device Continual Learning could help in making Edge AI more sustainable by specializing AI models directly on-field. We deploy a continual image recognition model on a Jetson Xavier NX embedded system, and experimentally investigate how Attention influences performance and its viability as a Continual Learning backbone, analyzing the redundancy of its components to prune and further improve our solution efficiency. We achieve up to 83.81% accuracy on the Core50's new instances and classes scenario, starting from a pre-trained tiny Vision Transformer, surpassing AR1*free with Latent Replay, and reach performance comparable and superior to the SoA without relying on growing Replay Examples.
ViT-LR: Pushing the Envelope for Transformer-Based On-Device Embedded Continual Learning / Dequino, A; Conti, Francesco; Benini, Luca. - (2022), pp. 1-6. (Intervento presentato al convegno 2022 IEEE 13th International Green and Sustainable Computing Conference (IGSC) tenutosi a Pittsburgh, PA (USA) nel 24-25 October 2022) [10.1109/IGSC55832.2022.9969361].
ViT-LR: Pushing the Envelope for Transformer-Based On-Device Embedded Continual Learning
Dequino, A;
2022
Abstract
State-of-the-Art Edge Artificial Intelligence (AI) is currently mostly targeted at a train-then-deploy paradigm: edge devices are exclusively responsible for inference, whereas training is delegated to data centers, leading to high energy and CO2 impact. On-Device Continual Learning could help in making Edge AI more sustainable by specializing AI models directly on-field. We deploy a continual image recognition model on a Jetson Xavier NX embedded system, and experimentally investigate how Attention influences performance and its viability as a Continual Learning backbone, analyzing the redundancy of its components to prune and further improve our solution efficiency. We achieve up to 83.81% accuracy on the Core50's new instances and classes scenario, starting from a pre-trained tiny Vision Transformer, surpassing AR1*free with Latent Replay, and reach performance comparable and superior to the SoA without relying on growing Replay Examples.File | Dimensione | Formato | |
---|---|---|---|
ViT-LR_Pushing_the_Envelope_for_Transformer-Based_On-Device_Embedded_Continual_Learning.pdf
non disponibili
Tipologia:
2a Post-print versione editoriale / Version of Record
Licenza:
Non Pubblico - Accesso privato/ristretto
Dimensione
222.94 kB
Formato
Adobe PDF
|
222.94 kB | Adobe PDF | Visualizza/Apri Richiedi una copia |
Pubblicazioni consigliate
I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/11583/2982851