HLM-Micro-har-v0 — Real-Data Benchmark Entry
6-class human activity recognition in the HLM-Micro tier, trained on real human IMU data from the UCI HAR dataset (30 volunteers, subject-disjoint train/test split).
Access
Weights + inference code available through Early Access or commercial engagement.
Task
6-class wearable activity classification.
| Class | |
|---|---|
WALKING | |
WALKING_UPSTAIRS | |
WALKING_DOWNSTAIRS | |
SITTING | static posture |
STANDING | static posture |
LAYING | static posture |
Headline
Test accuracy: 89.75% on the UCI HAR held-out subject-disjoint test set. Inside the published TinyML baseline range of 85–95% on the same benchmark.
Per-class: dynamic activities (walking / stairs up / down) all ≥93%; static postures harder — SITTING 87%, STANDING 90%, LAYING 82%. Static-posture confusion is a known HAR failure mode; we reproduce it on real data, which confirms the model learned real IMU structure rather than spurious features.
Why this model matters
This is the first HLM-Micro checkpoint trained on a public real-world dataset, providing a direct side-by-side with established TinyML baselines on a standard benchmark. The other zoo entries (anomaly, gesture, ECG) are synthetic-data proofs of pipeline. This one is the real-data reference.
The key claim: HLM-Micro at this scale is within ~5% of published TinyML baselines on raw single-task accuracy, while carrying the structural advantages that no TinyML entry offers (runtime compute dial, per-inference audit trail, multimodal single-checkpoint design, post-deployment surgery — all of which are described on the HLM-Micro model page).
Positioning
- Wearable activity classification — with per-user calibration
- Research baseline — compare against other polynomial-Hopfield / Hopfield-family approaches on UCI HAR
- Commercial fitness / occupational-monitoring — retrain on your population
Out-of-scope
- Not a medical gait-analysis tool — clinical validation and regulatory workstreams are separate projects
- No per-user calibration — production wearables typically adapt per user on first wear
- UCI HAR demographics — subjects 19–48 years old; broader demographics need broader training data
Dataset citation
Dataset courtesy of UCI ML Repository:
Anguita, D.; Ghio, A.; Oneto, L.; Parra, X.; Reyes-Ortiz, J.L. "A Public Domain Dataset for Human Activity Recognition Using Smartphones." 21st European Symposium on Artificial Neural Networks, ESANN 2013.
How to get started
Early Access or commercial engagement.
License
BSL 1.1 (Apache 2.0 in 2030) for weights + reference code. UCI HAR dataset terms apply if retraining.