AI at the edge
Sat 17 Oct · 14:00–15:00 · Room A (First floor)
Paper session · AI at the edge · Chair: Ingrid Lund
Papers
Sign in to ask or upvote questions. Sign in
- Full paperPresented
On-device training of TabNet models on Raspberry Pi 5: memory, time and energy
Maya Dimou (Baltic School of Computing) (presenting), Nikolai Sokolov (Alpine AI Institute)
Abstract
Edge training is often dismissed as too slow for tabular deep learning. We profile TabNet training on Raspberry Pi 4 and 5 boards across batch sizes and feature counts. One local epoch on 50,000 rows takes 94 seconds on a Pi 5 using 610 MB of memory and 0.9 Wh of energy. Orchestration overhead from the federated client stays below 0.7% of local training time.
edge AI · on-device training · Raspberry Pi · TabNet
Questions from the audience 0
No questions yet.
- Full paperPresented
Quantisation-aware federated training for microcontrollers
Luca Greco (Aegean Institute of Technology) (presenting), Alma Hansen (Lakeside Technical University), Theo Laurent (Meridian Research Centre)
Abstract
Microcontrollers can run inference but rarely take part in training. We train 8-bit quantised models directly on Cortex-M7 boards inside a federated loop, sending integer updates that are four times smaller. Keyword-spotting accuracy is within 1.5 points of float training while energy per round drops by 63%.
quantisation · microcontrollers · federated learning · TinyML
Questions from the audience 0
No questions yet.