Zygarde: Time-Sensitive On-Device Deep Inference and Adaptationon Intermittently-Powered Systems
summary
Zygarde is an energy- and accuracy-aware scheduler that runs compressed DNNs on batteryless microcontroller-class devices, modeling bursty harvestable energy with an eta predictability metric and treating each DNN layer plus a tiny classifier as an imprecise unit. A lightweight utility test uses the centroid distance gap to decide early termination and a layer-aware loss recovers ~4% accuracy. Evaluated on four datasets including ESC-10 and six real deployments, Zygarde cuts execution time up to 26% and schedules 9–34% more tasks with up to 21% higher inference accuracy versus EDF baselines.
AI-written summary · about
chapters (7)
- 0:00Start
- 0:40Alternative Energy Sources
- 1:08Research Challenge
- 2:07Observation 1
- 4:14Performance of Utility-based Early Termination
- 4:30Step 2: Layer-Aware Loss Function
- 5:19Performance of Intermittent Priority Function
related talks
15mZeus: Understanding and Optimizing GPU Energy Consumption of DNN Training
26mNeuOS: A Latency-Predictable Multi-Dimensional Optimization Framework for DNN-drive
41mBeefy ML: Ultra-Low Power Algorithms on Cattle
41mBeefy ML: Ultra-Low Power Algorithms on Cattle
20mALERT: Accurate Learning for Energy and Timeliness
23mDecentralized Application-Level Adaptive Scheduling for Multi-Instance DNNs on...
05mPlatform-Agnostic Lightweight Deep Learning for Garbage Collection...
13mLightning Talk: Energy-Efficient Deep Learning with PyTorch and Zeus