When Does Online Adaptation Pay on the Edge? A Leakage-Free Evaluation of Warmup, Learning-Rate Selection, and Resource Trade-offs for Time-Series Forecasting
Adam’s edge-adaptation advantage mostly survives a leakage-free test, but only after tuning the comparison setup.
The paper evaluates online adaptation for edge time-series forecasting across six public multivariate streams using pre-drift validation rather than test leakage. Warmup choices alone shifted the measured benefit by 3.0 to 18.8 percentage points. With validation-selected learning rates, Adam beat SGD with momentum in 310 of 360 cells, though four Adam cases still trailed the static baseline. The authors also report memory and A100 latency trade-offs for full, head-only, and calibration-based adaptation, while noting target-device latency and energy remain unmeasured. ArXiv · AI/CL/LG's note
The paper evaluates online adaptation for edge time-series forecasting across six public multivariate streams using pre-drift validation rather than test leakage. Warmup choices alone shifted the measured benefit by 3.0 to 18.8 percentage points. With validation-selected learning rates, Adam beat SGD with momentum in 310 of 360 cells, though four Adam cases still trailed the static baseline. The authors also report memory and A100 latency trade-offs for full, head-only, and calibration-based adaptation, while noting target-device latency and energy remain unmeasured. ArXiv · AI/CL/LG's note
score 3