NysHD Bridges Hyperdimensional Computing with Kernel Methods for AI.

Hyperdimensional Computing (HDC) is an emerging approach in cognitive science that processes information using high-dimensional random vectors, known for its mathematical rigor and hardware efficiency. A key challenge in HDC's machine learning applications is effectively mapping raw data into this high-dimensional space. Researchers have developed NysHD, a novel method that leverages the Nyström method, a technique from kernel approximation literature, to create these essential data mappings for HDC. This innovation provides a straightforward way to convert any positive-semidefinite similarity function into an equivalent HDC mapping. By enabling the import of a vast array of well-established kernel functions into the HDC framework, NysHD significantly broadens the types of problems HDC can address. Empirical evaluations demonstrate that NysHD achieves notable improvements in classification accuracy, averaging 11% better on graph datasets and 17% better on string datasets, compared to current HDC encoding techniques.