TinyOL-HITL

Incremental learning framework for edge devices with human-in-the-loop adaptation.

Overview

Applied on edge devices, TinyOL-HITL enables on-device incremental learning with human feedback. No cloud dependency. Streaming algorithms. Industrial integration.

Key Features:

• Streaming k-means clustering (<100KB model footprint)
• Human-in-the-loop corrections via MQTT
• Flash-based model persistence
• Industrial protocols (supOS-CE, Ignition)
• 1-year battery life target

Baseline: Improves on TinyOL (2021) with streaming updates and HITL integration.

Quick Start

See QUICKSTART.md for hardware setup and first run.

Documentation:

• Setup Guide - Hardware validation
• Technical Spec - System architecture
• Research Report - Paper structure

Architecture

Sensor → Features → Clustering → WiFi → supOS-CE → Ignition
                         ↑                              ↓
                         └──────── Human Labels ────────┘

Hardware

Reference: Raspberry Pi Pico 2 W (RP2350B)
Target: 1-year battery life, <100KB RAM
Extensible: Any Cortex-M33 platform

Components

• core/clustering/ - Streaming algorithms
• core/persistence/ - Model state management
• platforms/rp2350/ - Pico SDK wrapper
• integrations/ - supOS-CE and Ignition connectors

Use Cases

Condition Monitoring:

• Vibration anomaly detection
• Acoustic pattern recognition
• Temperature trend analysis

Adaptive Systems:

• User behavior learning
• Environmental adaptation
• Predictive maintenance

Project Status: Day 2 Complete

Day 1: Toolchain validated ✓
Day 2: Core k-means + RP2350 integration ✓
Day 3: supOS-CE integration (next)

Day 2 Metrics:

• Memory: 4.2 KB model footprint
• Convergence: 150 points, ±0.05 error
• Throughput: 150 points/sec @ 150MHz
• Tests: 9/9 pass

Roadmap:

• Toolchain validation
• Core k-means implementation
• RP2350 platform support
• supOS-CE integration
• Ignition visualization
• Energy profiling

Contributing

See CONTRIBUTING.md for development setup.

We welcome:

• Platform ports
• Algorithm improvements
• Integration modules
• Documentation enhancements

License

Apache-2.0 - See LICENSE for details.

Citation

If you use this framework in research:

@misc{tinyol-hitl,
  author = {Lee Kai Ze},
  title = {TinyOL-HITL: Incremental Learning Framework for Edge Devices with Human-in-the-Loop},
  year = {2025},
  publisher = {GitHub},
  url = {https://github.com/leekaize/tinyol-hitl}
}

References

Baseline: TinyOL (2021) - On-device learning with 256KB model, batch processing.

Our improvements: Streaming updates, <100KB memory, HITL integration, industrial protocols.

Contact

• Issues: Bug reports and feature requests
• Discussions: Architecture questions and use cases
• Email: mail@leekaize.com (research collaboration)

---

Built for: Embedded researchers, industrial IoT developers, edge ML practitioners