Er-HDC

Error Resilient Hyperdimensional Computing Using Hypervector Encoding and Cross-Clustering

Description

The paper provided two algorithms a checksum hypervector encoding (CHE) framework and cross-hypervector clustering (CHC) framework to make the hyperdimensional computing resilient against soft and timing errors that may occur in the associative memory.

Installation and Setup

Step 1: Install Requirements

First, install the necessary dependencies listed in requirements.txt. Run the following command: pip install -r requirements.txt

Step 2: Create Data Directory

Create a folder named data to store your datasets. Use the following command:

mkdir ./data

Step 3: Download Data

Download the required data they are stored as .npy files. The folder includes ISOLET and UCIHAR datasets.

link: https://drive.google.com/drive/folders/1EoCzkYAKJrCbxAzH5aAgvYHVAjJeZBLF?usp=sharing

Usage

Run the algorithm using the command below. You can adjust the parameters as needed for different configurations.

python3 main.py
--x_train_filepath "../data/isolet/X_train.npy"
--y_train_filepath "../data/isolet/y_train.npy"
--x_test_filepath "../data/isolet/X_test.npy"
--y_test_filepath "../data/isolet/y_test.npy"
--dim 10000
--epochs 200
--lr 0.035
--init_er 8
--final_er 1
--Thresh 10
--encoding_system "kernel"
--nbr_cluster 1500
--bootstrap 1.0
--eps 1e-5
--itr 2
--weight_cluster 1.1

Custom Configurations

You can modify the parameters in the command to fit your specific needs. Here are the parameters you can adjust:

• dim: Dimension size (e.g., 10000)
• epochs: Number of training epochs (e.g., 200)
• lr: Learning rate (e.g., 0.035)
• init_er: Initial error rate (e.g., 8)
• final_er: Final error rate (e.g., 1)
• Thresh: Threshold value (e.g., 10)
• encoding_system: Encoding system used (e.g., "kernel")
• nbr_cluster: Number of clusters (e.g., 1500)
• bootstrap: Bootstrap value (e.g., 1.0)
• eps: Epsilon value for convergence (e.g., 1e-5)
• itr: Number of iterations (e.g., 2)
• weight_cluster: Weight for clustering (e.g., 1.1)

Reproducing Paper Results

To reproduce the results of the paper, particularly for the Clustering method, use the following settings:

Dataset	HDC Dimension	Number of Clusters	Threshold
Isolet	10000	1500	10
Ucihar	10000	1500	10

For academic use, please cite this work as follows:

Citation

If you find this repository useful in your research, please consider citing the following papers:

@article{Er-HDC2024,
  author        = {Mohamed mejri and Chandramouli Amarnath and Abhijit Chatterjee},
  title         = {Error Resilient Hyperdimensional Computing Using Hypervector Encoding and Cross-Clustering},
  conference    = {IEEE VLSI Test Symposium},
  year          = {2024},
}