Dynamic Hyperdimensional Computing for Improving Accuracy-Energy Efficiency Trade-Offs

Abstract

Brain-inspired Hyperdimensional (HD) computing is an emerging technique that computes with either binary or integer HD vectors. However, both vector representations confront an extreme trade-off between accuracy and energy efficiency. This issue limits the generalizability of HD computing for many applications. In this paper, we propose a threshold-based dynamic HD computing framework (TD-HDC) to improve the accuracy-energy efficiency trade-offs. Standard HD computing always executes the same processing flow regardless of input data. On the contrary, TD-HDC dynamically selects the execution path between the binary and integer HD models based on the classification difficulty of input data. In other words, TD-HDC utilizes both HD models and manages to efficiently allocate their computational resources. On the MNIST dataset, we demonstrate that our proposed framework is flexible and can reduce energy consumption and execution time by 51.3% and 15%, respectively, under the same accuracy level.