Classification with binary hyperdimensional computing

Hyperdimensional computing (HDC) has been introduced as a brain-inspired, lightweight, and energy-efficient algorithmic framework suitable for wearable internet of things (IoT) devices, near-sensor artificial intelligence (AI) applications, and on-device processing. It originates at the intersection of symbolic AI and connectionism, with the potential of combining the advantages of both, i.e., the transparency of symbolic representations, and flexibility and robustness of connectionist representations. HDC maps input data to vectors in a hyperdimensional (HD) space, i.e., hyperdimensional vectors (HVs) with typically thousands of components, and relies on simple component-wise arithmetic vector operations. Several frameworks of HDC exist, distinguished by their vector component types. This dissertation focuses on dense binary HDC since it has been reported to offer higher energy efficiency and more robustness compared to its non-binary counterpart, at the cost of a slightly lower classification accuracy. Therefore, this dissertation investigates adjustments and improvements to the dense binary HDC classification model to enhance its classification performance in terms of accuracy and transparency. First, a novel uniform framework is developed to encode binarized images into HVs. The framework uses a local feature extraction method relying only on native HD arithmetic vector operations. The proposed encoding framework outperforms other studies using native HDC with different encoding approaches, and is on par with more complex hybrid HDC models and lightweight binary neural networks (BNNs). It also demonstrates higher robustness to noise and blur compared to the baseline encoding. Second, two models that enhance previous HDC classifiers are introduced and explored. The models use a confidence metric that measures the confidence with which a sample has been classified during the training procedure. This enhanced training procedure consistently improves classification accuracy compared to the baseline and increases confidence in the classifier’s predictions. Third, a generic post-hoc analysis is proposed, which determines the importance of individual features in the HDC classification model. This analysis provides insight and transparency in the decision-making process at every deployment of the HDC classification model.