HDCOG: A Lightweight Hyperdimensional Computing Framework with Feature Extraction

Shijin Duan, Xiaolin Xu

Abstract: As an emerging classification and recognition paradigms on nanoscale edge-devices, the performance (e.g., inference accuracy) of Hyperdimensional Computing (HDC) is usually obstructed by the limited computing resources. This paper proposes HDCOG, a lightweight framework leveraging feature extraction as a preliminary of binary HDC model. HDCOG extracts distinguishable feature information from the data input during the encoding phase, which significantly eliminates information redundancy in object classification and recognition. We validate HDCOG framework with several popular datasets against other state-of-the-art binary HDC models and machine learning methods. Experimental evaluation and analysis demonstrate that, compared with other state-of-the-art binary HDC models, HDCOG significantly reduces the memory overhead to lower than 10% and consumes comparable or even less inference time. Moreover, HDCOG also achieves about 9% accuracy improvement on complex datasets against other binary HDC models, which is comparable to lightweight neural networks.