Validate resource requirements on edge hardware

Validate adaptation-state memory, per-update latency, and energy consumption for the evaluated online time-series forecasting strategies on Jetson-, Raspberry-Pi-, or meter-class edge hardware before claiming end-to-end deployability.

Background

All resource measurements in the paper are obtained on a datacenter A100 GPU, and the energy estimates rely on an illustrative 5 W assumption. Batch-one GPU timing is launch-bound and may not represent the behavior of embedded devices or smart meters.

The authors therefore leave unresolved whether the measured memory, latency, and energy trade-offs transfer to actual deployment hardware. This validation is necessary to determine whether the proposed commissioning and adaptation recipe is deployable end to end.

References

Validating memory and latency on Jetson- or Raspberry-Pi-class hardware, and ultimately on a meter, is required before the recipe can be called deployable end-to-end.

— 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  (2609.01126 - Fujimoto et al., 1 Sep 2026) in Discussion and Limitations, Section 5; Conclusion

On-device continual learning is an open research problem in its own right.

— Catastrophic Forgetting in Sequential Thermal Anti-UAV Detection: The Role of Scale-Conditioned Gradient Imbalance  (2610.08315 - Nguyen et al., 6 Oct 2026) in Section 2, subsection “Compute Infrastructure”