Lecture Notes: Top 8 TinyML Frameworks and Compatible Hardware Platforms

Introduction to TinyML Frameworks

• TinyML frameworks enable the development and deployment of machine learning models on edge devices.
• They provide tools and resources for implementing machine learning on microcontrollers and embedded systems.
• The focus is on low-power, small-footprint microcontrollers.

Importance of TinyML

• Executes ML models on small microcontrollers like Raspberry Pi, ESP32.
• Consumes less power, generally less than one milliwatt.
• Offers low lag time with integrated ML algorithms.

Understanding TinyML Frameworks

• Specialized software/tools for developing ML models for edge devices.
• Supports data collection, model training, optimization, and deployment.

Top 8 TinyML Frameworks

1. TensorFlow Lite (TFLite)

• Google's open-source ML framework for microcontrollers.
• Advantages: Fast inference, flexibility, ease of integration.
• Limitations: Limited TensorFlow operations, device support.

2. Edge Impulse

• Facilitates the whole edge AI lifecycle from data collection to deployment.
• Advantages: EON Compiler reduces RAM and flash usage.
• Limitations: Limited customization and hardware compatibility.

3. PyTorch Mobile

• Supports model deployment on mobile devices (iOS, Android).
• Key Features: Multi-platform support, API availability, TorchScript support.

4. uTensor

• Lightweight framework for Arm platforms based on TensorFlow.
• Key Features: Secure, user-friendly API, flexible extensibility.

5. STM32Cube.AI

• Generates optimized code for STM32 ARM Cortex M-based boards.
• Key Features: Supports multiple frameworks, easy portability.

6. NanoEdgeAIStudio

• User-friendly tool requiring no data science expertise.
• Features: Supports anomaly detection, offers auto-search engines.

7. NXP eIQ

• Includes libraries and tools for NXP microcontrollers.
• Key Features: Optimization tools, variety of inference engines.

8. Embedded Learning Library (ELL)

• Developed by Microsoft for Raspberry Pi, Arduino, micro:bit.
• Supports image and audio classification.

Compatible Hardware Platforms

• Detailed table of hardware platforms supported by each TinyML framework.

Targeted Applications

• Frameworks support a wide range of applications like image/audio classification, object detection, predictive maintenance.

Conclusion

• TensorFlow offers flexibility across various platforms and applications.
• Edge Impulse provides efficient neural network optimization.
• Other frameworks like PyTorch, uTensor, and platforms by ST, NXP, Microsoft contribute significantly to TinyML advancement.