Skip to content

/projects · seeker

Seeker

Coreset-based partial inference on energy-harvesting wearables.

arXiv 2022

The problem

In an energy-harvesting wearable network, radio is the most expensive operation by an order of magnitude. Streaming raw sensor traces back to a host phone or hub burns the energy budget before anything useful gets classified.

What Seeker does

Each sensor node runs a partial DNN locally — just enough to construct a coreset: a compressed, representative subset of features that preserves the decision-theoretic information about the input. The host receives the coreset, not the raw stream, and finishes inference with near-baseline accuracy.

Two co-design pieces make this work on a tiny budget:

  • Lightweight hardware support for coreset construction, with store-and-execute on non-volatile memory so power loss doesn’t lose the in-flight features.
  • Edge-host synergy: the sensor and the host share a learned representation that’s tuned for the specific upstream model the host will run.

Results

Up to 8.9× reduction in communication versus full-trace baselines, with classification accuracy within a few percent of cloud-only inference. Energy savings dominate compute overhead by a wide margin because radio is so much more expensive than local arithmetic.

Why it matters

The bottleneck on energy-harvesting wearables isn’t the model — it’s the bus to the model. Seeker makes the bus quiet enough that intermittent nodes can stay useful long-term.