Machine Learning in the Age of IoT
summary
PervasiveFL proposes a lightweight cloud modulate that pairs with diverse local nets, uses an entropy-based decision gating and ensemble mutual learning, uploads modulate gradients from 10 emulated boards and shows ImageNet10 accuracy gains (~+3.17%) while reducing communication: modulate gradients 8.71 MB vs large model 162 MB and 1.65s vs 13.96s per round. Random-forest work implements three DT kernels on a RISC‑V MCU (struct loop, packed struct, DTE array), achieves 4–4.8× speedups, 4.5 µs latency, 220k inferences/s and large energy savings. Temporal action localization streams video on-device with a contrastive, self-adaptive merging and interest-based sampling to limit oracle labels and improve weakly supervised localization from long untrimmed streams.
AI-written summary · about
chapters (8)
- 0:00Start
- 5:29Step 6: Modellet Synchronization
- 28:29Detecting actions in videos on edge devices
- 31:18Revisiting Video Learning
- 34:31Self-adaptive Video Dividing Approach
- 35:55Interests-based Sampling Strategy
- 36:39Experimental Setup
- 37:47Different Merging and Sampling Strategies
related talks
59mML and the IoT: Living on the Edge
01h 01mML and the IoT: Living on the Edge
01h 01mMachine Learning, Networks, and IoT
54mSmarter Everything with ML & the IoT
52mMachine Learning in IoT Applications
49mEnable IoT with Edge Computing and Machine Learning
30mDeep Learning in Energy Management: Non-Intrusive Load Monitoring for IoT Devices
06mIntermittent Learning: On-Device Machine Learning on Intermittently Powered System
