Skip to content

Artificial Intelligence

The WILD device supports onboard biomarker detection and behavioral classification for responsive stimulation.

Current WILD device releases are intentionally conservative: AI models are curated into validated release images and reviewed for RAM use, compute timing, sampling schedules, and closed-loop latency.

Current Model Integration

Current AI support uses validated, experiment-specific models rather than arbitrary runtime uploads.

  • Models are distributed as part of validated release images.
  • RAM use is curated before deployment.
  • Inference timing is checked against acquisition, storage, BLE, and DSP tasks.
  • Model output is integrated with device state reporting and closed-loop logic.
  • Release image and model identity are recorded with each experiment.

This approach keeps timing behavior predictable on resource-constrained embedded hardware.

Generic Model Support

Generic model loading is not part of the stable public workflow yet. Current releases prioritize predictable timing, memory use, and closed-loop behavior.

Closed-loop model releases define:

  • Accepted model format and quantization requirements.
  • Maximum model size and tensor arena limits.
  • Input and output tensor conventions.
  • Inference scheduling relative to acquisition and DSP.
  • Validation procedure before animal experiments.
  • How model version, release image, and experiment metadata are recorded.

Deployment discipline

Embedded AI models are part of the validated device release. A model that passes desktop inference tests may still be unsafe for closed-loop use if RAM pressure or inference timing affects acquisition, storage, or stimulation behavior.