Hyperdimensional Computing

This repository implements Hyperdimensional Computing (HDC) models in C++. The models provide solutions to common AI problems but use the HDC framework instead of other usual models such as NNs. The implementations available in this repository are based on papers available in HDC literature.

Models implemented:

• Language recognition (language): Identify the language an unknown string.
• VoiceHD (voicehd): Speech recognition with ISOLET dataset.
• Digit recognition (mnist): Detect a handwritten digit using the MNIST dataset.
• EMG (emg): Hand gesture recognition.

Running

cmake -B build -S .
cmake --build build # -j<nprocs> to build in parallel

You can create optimized builds by executing:

cmake -B build -S . -DCMAKE_BUILD_TYPE=Release

This generates a binary for each implemented HDC model in the path build/hdc. Execute the binary from the repository's root directory using:

./build/language dataset/language/train dataset/language/test
./build/voicehd dataset/voicehd/train* dataset/voicehd/test*
./build/mnist dataset/mnist dataset/mnist/train* dataset/mnist/test*
./build/emg dataset/emg

Experimental AVX-256 for binary HDC

It is possible to accelerate binary HDC by using its experimental AVX-256 implementation. While it can greatly reduce execution time, it is not safe and guaranteed that the executables will run correctly. You can enable it by defining __ASM_LIBBIN at configure time:

cmake -B build -S . -DCMAKE_BUILD_TYPE=Release -DCMAKE_CXX_FLAGS='-march=native -D__ASM_LIBBIN'