HDMNIST

Hyperdimensional Computing implementation with MNIST dataset.

Instructions

Preparing the dataset

• The MNIST dataset can be obtained from http://yann.lecun.com/exdb/mnist/
• Put the extracted data and label files in one folder.
• Modify the mnist_path variable to the folder.

Installing prerequisites.

• This application runs udner Python 3. Please have the newest python version installed.

• This application requires the following packages: pickle, pandas, numpy, scikit-learn and python-mnist. Packages can be installed via:

  pip install pickle pandas numpy scikit-learn python-mnist

Run the application.

• Application can be run by directly executing the python script, e.g. "python3 HDMNIST.py". Give mode argument to main() to select which mode this script operates on. Please note that you must run in "train" mode to train your AM before running in "test" mode.
• You can adjust different parameters of HDC, including: Dimension, epochs of retraining, training and testing set sizes and quantization bits.
• Random train/test split is not implemented yet. You need to implement by yourself.
• This script will generate files of considerable size in the directory, make sure you have proper access and disk space.

Citation

Dongning Ma, Jianmin Guo, Yu Jiang, Xun Jiao, "HDTest: Differential Fuzz Testing of Brain-Inspired Hyperdimensional Computing". 58th 2021 ACM/EDAC/IEEE Design Automation Conference (DAC), San Francisco, CA, 2021.