G-Net

This repository contains the code and experiments for our paper "G-Net: A Provably Easy Construction of High-Accuracy Random Binary Neural Networks", which introduces a two-step approach using G-Net and its Embedded High-Dimensional Representation EHDG-Net.

---

Project Structure

• Quick Demo/Step1_GNet_Training.ipynb: Trains the base G-Net model on MNIST.
• Quick Demo/Step2_EHDGNet.ipynb: Loads the trained G-Net and performs inference using its EHD representation.
• requirements.txt: Lists all required Python packages.
• Comparison With Other HDC Methods/: Contains the experiments for comparing EHDG-Net with other Hyperdimensional Computing Methods.
• RASU vs. TASU/: Contains the experiments for comparing Gaussian and Rademacher RASU with Gaussian and Rademacher TASU frameworks in EHDG-Net.
• Robustness Experiments: Contains the experiments for checking the robustness of RASU vs TASU and a comparison with other HDC methods under bit flip perturbations from 0% to 50%.

---

Quick Start

1. Clone the Repository

git clone https://github.com/GNet2025/GNet.git
cd GNet

2. Install Dependencies

pip install -r requirements.txt

3. Training the G-Net

Run the notebook Quick Demo/Step1_GNet_Training.ipynb to train the G-Net model on MNIST. This will save the trained parameters to disk.

4. Inference with EHDG-Net

Run the notebook Quick Demo/Step2_EHDGNet.ipynb to perform the inference using the Embedded High-Dimensional G-Net.