Abstract
Machine Learning algorithms based on Brain-inspired Hyperdimensional (HD)computing imitate cognition by exploiting statistical properties ofhigh-dimensional vector spaces. It is a promising solution for achieving highenergy-efficiency in different machine learning tasks, such as classification,semi-supervised learning and clustering. A weakness of existing HDcomputing-based ML algorithms is the fact that they have to be binarized forachieving very high energy-efficiency. At the same time, binarized models reachlower classification accuracies. To solve the problem of the trade-off betweenenergy-efficiency and classification accuracy, we propose the QubitHDalgorithm. It stochastically binarizes HD-based algorithms, while maintainingcomparable classification accuracies to their non-binarized counterparts. TheFPGA implementation of QubitHD provides a 65% improvement in terms ofenergy-efficiency, and a 95% improvement in terms of the training time, ascompared to state-of-the-art HD-based ML algorithms. It also outperformsstate-of-the-art low-cost classifiers (like Binarized Neural Networks) in termsof speed and energy-efficiency by an order of magnitude during training andinference.