HyperSpace: A Generalized Framework for Spatial Encoding in Hyperdimensional Representations
Researchers from George Mason University, Neya Robotics, and the US Army Ground Vehicle Systems Center have introduced HyperSpace, an open-source framework designed for Vector Symbolic Architectures (VSAs). This tool decomposes VSA systems into modular operators, including encoding, binding, bundling, similarity, cleanup, and regression. The study utilizes HyperSpace to benchmark two prominent VSA backends: Holographic Reduced Representations (HRR) and Fourier Holographic Reduced Representations (FHRR). While FHRR theoretically offers lower complexity for individual operations, the framework reveals that similarity and cleanup processes dominate runtime in spatial domains. Consequently, both models demonstrate comparable end-to-end performance. However, significant differences in memory footprint exist, with HRR requiring approximately half the memory of FHRR vectors. By facilitating modular, system-level evaluation, HyperSpace exposes practical deployment trade-offs in VSA pipelines that are often obscured by theoretical or operator-level comparisons alone. This development provides valuable insights for optimizing hyperdimensional computing applications, particularly where memory efficiency and runtime performance are critical factors.
Wire timeline
HyperSpace: A Generalized Framework for Spatial Encoding in Hyperdimensional Representations
Researchers from George Mason University, Neya Robotics, and the US Army Ground Vehicle Systems Center have introduced HyperSpace, an open-source framework designed for Vector Symbolic Architectures (VSAs). This tool decomposes VSA systems into modular operators, including encoding, binding, bundling, similarity, cleanup, and regression. The study utilizes HyperSpace to benchmark two prominent VSA backends: Holographic Reduced Representations (HRR) and Fourier Holographic Reduced Representations (FHRR). While FHRR theoretically offers lower complexity for individual operations, the framework reveals that similarity and cleanup processes dominate runtime in spatial domains. Consequently, both models demonstrate comparable end-to-end performance. However, significant differences in memory footprint exist, with HRR requiring approximately half the memory of FHRR vectors. By facilitating modular, system-level evaluation, HyperSpace exposes practical deployment trade-offs in VSA pipelines that are often obscured by theoretical or operator-level comparisons alone. This development provides valuable insights for optimizing hyperdimensional computing applications, particularly where memory efficiency and runtime performance are critical factors.
cs.AI updates on arXiv.org