High-Performance Hyperdimensional Computing in JAX
Open-source JAX library for hyperdimensional computing (HDC) and vector symbolic architectures (VSA). Published in the Journal of Machine Learning Research, 2025.
Overview
JAX-HDC provides a unified functional API for four VSA models — Binary Spatter Codes (BSC), Multiply-Add-Permute (MAP), Holographic Reduced Representations (HRR), and Fourier HRR — fully compatible with JAX transforms (jit, vmap, pmap, grad).
Key Results
- 8–10x CPU speedups over NumPy, 40–80x GPU speedups over PyTorch baselines
- XLA compilation enables kernel fusion and hardware acceleration across CPUs, GPUs, and TPUs
- Benchmarked on EU languages, EMG gestures, and VoiceHD classification tasks
Features
- Random and feature encoders, centroid classifiers, gradient-based learning integration
- Designed for large-scale HDC research in ML, neuro-symbolic AI, and edge computing
ArXiv link coming soon.