Edge intelligence

Research question

What accuracy, latency and energy trade-offs emerge when temporal models run locally?

Proposed method

Compare statistical baselines, compact neural models and quantisation using the same held-out recordings. Measure energy on actual hardware; model size is not a substitute for a power measurement.

Limitations

No sub-watt performance or hardware validation is claimed. Deployment decisions require measurement, reproducibility and a fail-safe design.

Why investigate processing near the sensor?

Processing near acquisition could reduce the amount of raw data transmitted and allow experiments to continue during connectivity interruptions. These are design motivations. Their practical value would need to be measured against the constraints of the selected hardware, model and operating environment.

A proposed comparison plan

Begin with a simple threshold or statistical baseline. Compare a compact temporal model using identical training and test partitions. Record preprocessing cost as well as inference cost; a fast model may still depend on an expensive feature pipeline. Investigate quantisation only after establishing a reproducible baseline.

Measurements to report

• End-to-end latency, including acquisition windows and preprocessing.
• Memory use and storage requirements on the tested device.
• Power measurements under a documented workload.
• Classification or estimation error on unseen recording sessions.
• Behaviour when input is missing, corrupted or outside the training distribution.

Model maintenance

A future deployment design would need versioned models, signed updates, rollback and a record of the configuration used for each result. A model update should trigger regression evaluation. Quietly changing a model would make longitudinal comparisons harder to interpret.