Embedded AI Portfolio

This repository contains my coursework, lab implementations, and research presentation for the Embedded Artificial Intelligence course at Polytech Nice Sophia. The projects focus on deploying Deep Learning models (CNNs) onto resource-constrained microcontrollers (STM32L4) and exploring unsupervised learning algorithms.

📂 Repository Structure

🧪 Labs: Edge AI Pipeline

A progressive series of labs demonstrating the full lifecycle of Embedded AI, from training to deployment on an STM32L476RGT6 (Cortex-M4).

Figure 1: RFThings-AI Dev Kit board equipped with a STM32L476RGT6 Microcontroller

• Lab 1: Environment Setup

  • Established a Docker-based toolchain with TensorFlow and Arduino IDE.
  • Verified hardware constraints (80 MHz, 128 KB SRAM, 1 MB Flash).

• Lab 2: CNNs & MicroAI Framework

  • Designed and trained Convolutional Neural Networks (CNNs) on standard datasets (MNIST, UCI HAR).
  • Key Achievement: Manually implemented Dense and Convolutional layers in C to understand the low-level arithmetic of inference.
  • Used the MicroAI tool to automatically generate optimized, fixed-point C code from Keras models.

• Lab 3: Human Movement Recognition (PolyHAR)

  • Goal: Real-time Human Activity Recognition (HAR) on-device.
  • Data: Collected a custom accelerometer dataset (Positive/Negative activity classes) using the RFThings board.
  • Modeling: Trained a 1D-CNN in TensorFlow/Keras on the custom time-series data.
  • Deployment: Converted the model to a 16-bit fixed-point C library and integrated it into the microcontroller firmware for real-time inference (LED actuation upon detection).

Figure 2: Training Accuracy & Loss Curves

🧠 Research Presentation: Growing Neural Gas (GNG)

Located in /GNG_presentation

An exploration of Growing Neural Gas, an unsupervised learning algorithm that learns the topology of data inputs without a predefined structure (unlike SOMs).

Figure 3: Training -> Step 0 - 100

Figure 4: Training -> Step 100 - 1,000

Figure 5: Training -> Step 1,000 - 3,000

Figure 6: Training -> Step 3,000 - 8,000

Figure 7: Training -> Step 8,000 - 12,000

• GNG.py: A Python implementation of the GNG algorithm from scratch.
  • Includes a real-time visualization of the network "growing" to fit a non-linear dataset (Two Moons).
• Slides: Self-Growing Neural Networks.pdf - A detailed presentation on the algorithm's theory, math, and comparison to Self-Organizing Maps (SOM).

🛠️ Tech Stack & Skills

• Hardware: STM32L4 Microcontrollers, RFThings-AI Dev Kit, IMU Sensors.
• Languages: C/C++ (Firmware), Python (Training & Simulation).
• Libraries: TensorFlow/Keras, NumPy, MicroAI (C inference engine).
• Tools: Docker, Arduino IDE, Jupyter Notebooks.

🚀 Usage

Running the GNG Simulation:

cd GNG_presentation
python GNG.py

(Ensure you have numpy, matplotlib, and scikit-learn installed)