Abstract
This paper studies the geometry of binary hyperdimensional computing (HDC), acomputational scheme in which data are encoded using high-dimensional binaryvectors. We establish a result about the similarity structure induced by theHDC binding operator and show that the Laplace kernel naturally arises in thissetting, motivating our new encoding method Laplace-HDC, which improves uponprevious methods. We describe how our results indicate limitations of binaryHDC in encoding spatial information from images and discuss potentialsolutions, including using Haar convolutional features and the definition of atranslation-equivariant HDC encoding. Several numerical experimentshighlighting the improved accuracy of Laplace-HDC in contrast to alternativemethods are presented. We also numerically study other aspects of the proposedframework such as robustness and the underlying translation-equivariantencoding.