Hyperdimensional Computing for Robust and Efficient Unsupervised Learning

Abstract

Clustering has emerged as a critical tool in diverse fields. Nevertheless, its high computational cost has been a persistent challenge, particularly for large-scale datasets. To address this, various compute-in-memory (CiM) approaches have been proposed, including the use of Ferroelectric FET (FeFET) technology due to its ultra-efficient and compact CiM architecture. However, non-idealities resulting from cell thickness and device temperature have impeded the scaling of FeFETs and thus hindered their potential to be used for clustering. In light of this, we propose a Hyper-Dimensional Computing (HDC) framework specifically for FeFET technology in the context of clustering. Our approach involves a cross-layer FeFET reliability model that captures the effects of scaling on multi-bit FeFETs, taking into account the impact of process variation and inherent stochasticity. We use two models in our HDC framework, a full-precision, ideal model for training, and a quantized error-impacted version for validation and inference. This iterative adaptation strategy helps to overcome the challenges associated with the non-idealities of FeFET technology. Our results demonstrate the proposed HDC framework performs better than traditional algorithms such as k-means and BIRCH. Moreover, our model can function as its ideal counterpart without noise, proving its potential to scale FeFET technology for clustering applications.