Skip to content

About

No description, website, or topics provided.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Repository files navigation

Akida Pico Guides

Draft guides for building, training and fine-tuning models for BrainChip Akida Pico, an ultra-low-power, event-based neural IP for always-on 1-D sensing (audio, vibration, biosignals).

Status: draft for review. Both guides were assembled from internal documents, tickets and the public akida_models 1.14.1 source. The pipeline was not re-run on the Pico FPGA while writing them. Each guide ends with a checklist of open items that need an owner's answer before publication.

Contents

File What it is
Akida_Pico_Model_Development_Guide.md From a trained 1-D model to a quantized network running on Pico: recurrent TENN concepts, the two zoo models, design rules and hardware limits, the convert-quantize-map workflow, validation, on-device measurement, application status, troubleshooting.
Akida_Pico_Model_Training_and_Fine-Tuning_Guide.md How to get trained weights: environment, data preparation, training from scratch, fine-tuning, evaluation and reproducibility, post-training steps, architecture changes, anomaly detection notes, troubleshooting.
customer/ Customer editions, with internal references and reviewer checklists removed: the two guides above, and Customizing_the_Akida_Pico_KWS_Model.md, a short path for adapting the Pico KWS model to your own keywords.
figs/ Slides embedded in the guides: KWS core power (internal edition only) and bearing fault detection results.
resources.md Every source used to write the guides, what was taken from each, what was only seen in search results, and what could not be accessed.

Internal drafts and customer editions. The two guides in the repository root are internal review drafts. They cite internal tickets and documents and end with open-item checklists. Anything intended for customers should come from customer/, which has none of that.

How to read them

  1. Start with the Model Development Guide to learn what Pico runs and what limits a model must meet.
  2. Use the Training & Fine-Tuning Guide when you need weights for your own data. It stops where the first guide's workflow begins and links back to it.
  3. Check resources.md to trace any claim to its source.

Audience

ML engineers at Pico licensees, and BrainChip field engineers supporting them. The guides assume Keras and TensorFlow experience and no prior Akida knowledge.

Scope

  • Covered with runnable models: 12-class keyword spotting (tenn_recurrent_sc12) and 4-label bearing fault classification (tenn_recurrent_uored).
  • Covered as status and design guidance only: 1-D anomaly detection, wake word and health monitoring. These are Pico requirements without a packaged public model yet.
  • Toolchain: MetaTF (akida, cnn2snn, quantizeml, akida_models). Examples target the Akida Pico FPGA cloud platform and its software model.

Known limits of this draft

  • The fine-tuning recipe is proposed, not validated. The public training scripts have no tuning mode; learning rates, epoch counts and the freezing schedule are starting points to replace with measured values.
  • The power figures (about 64 µW and 1 µJ for KWS) are estimates from hardware characterization, not FPGA measurements.
  • The hardware-limits table mixes current mapping checks with the 2024 hardware specification. The mapping error is the final authority.
  • Before external release, remove the internal Jira references, confirm approval to describe the recurrent TENN layer externally, and keep any client-data models out entirely.

Feedback

Open an issue or comment on a pull request. Useful feedback names the section, the claim, and the source or measurement that supports a correction.

© 2026 BrainChip Holdings Ltd. All rights reserved.

About

No description, website, or topics provided.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors