Hyperdimensional Computing Encoding Schemes for Improved Image Classification

Abstract

We introduce a novel encoding scheme for hyperdimensional computing (HDC) image classification tasks that takes advantage of both spatial awareness of pixels and nonlinear relationships between pixel values using a Siamese Neural Network (SNN) architecture. We demonstrate that, using this encoding scheme, we can achieve improved classification accuracy on the MNIST and CIFAR datasets over the current state-of-the-art binary HDC encoding scheme.