ALERT: Accurate Learning for Energy and Timeliness
summary
ALERT is a runtime scheduler that simultaneously picks a DNN and a power cap per input, using a probabilistic global slowdown factor and a Kalman filter to predict latency, accuracy, and energy. It evaluates on CPU and GPU for image and speech tasks and reports coordinating application and system adaptation reduces energy by 13% and error by 27% versus single-layer approaches, while running within 3% energy and 2% error of an oracle with perfect knowledge.
AI-written summary · about
chapters (14)
- 0:00Intro
- 0:22DNN is Deployed Everywhere
- 3:49Evaluation Highlights
- 5:32Tradeoffs from System Settings
- 6:11Run-time Variability
- 6:50Potential Solutions
- 7:40Outline
- 8:08Maximize Accuracy Task
- 8:49How to estimate the inference latency? . Two key challenges
- 12:23What is an Anytime DNN?
- 14:05How to estimate the energy consumption?
- 15:21Experiment Settings
- 16:31Evaluation: Scheduler Performance
- 17:20How ALERT Works with Traditional DNN
related talks
17mRECL: Responsive Resource-Efficient Continuous Learning for Video Analytics
11mLitePred: Transferable and Scalable Latency Prediction for Hardware-Aware Neural...
25mSymptom-based Alerting for Machine Learning - What I Learned...
06mZygarde: Time-Sensitive On-Device Deep Inference and Adaptationon Intermittently-Powered Systems
06mIntermittent Learning: On-Device Machine Learning on Intermittently Powered System
23mDecentralized Application-Level Adaptive Scheduling for Multi-Instance DNNs on...
40mNext-generation Alerting and Fault Detection
28mAI vs Climate Change: Insights from the Cutting Edge