Towards Efficient Learning on Edge by Hyperdimensional Computing

Recent advancements in machine learning, while powerful, are often burdened by significant computational and memory requirements, limiting their deployment in resource-constrained settings. Hyper dimensional Computing (HDC) emerges as an alternative with its simplicity, lightweight operations, and robustness to errors. By encoding data into high-dimensional vectors and performing efficient algebraic computations, HDC opens a new avenue as an efficient learning paradigm. In this talk, Dr. Fatemeh Asgarinejad will introduce the fundamentals of HDC and briefly discuss existing research that has extensively explored various stages of HDC algorithm. Then, she will present three key domains of her research: First, she will discuss PIONEER, a novel approach that employs learned projection vectors to optimize the encoding process. By leveraging neural networks to learn these vectors, PIONEER enables HDC to achieve high accuracy with significant computational efficiency. Second, she will present HDXpose, an adversarial attack framework that exploits an advantage of HDC: “explainability”. By strategically analyzing and perturbing influential input points, HDXpose effectively unveils vulnerabilities within HDC models, underscoring the need for robust security measures in HDC system design. Lastly, Dr. Asgarinejad will show an application of HDC in developing a cost-effective and noise-resilient pressure mat system for human activity recognition. The HDC-based system surpasses CNNs in accuracy and efficiency. Co-sponsored by: Media Partner: Open Research Institute (ORI) Speaker(s): Fatemeh Asgarinejad Agenda: – Invited talk from Dr. Fatemeh Asgarinejad, an incoming Assistant Professor of Teaching in the Electrical and Computer Engineering Department at the University of California, Riverside. – Q/A Session Virtual: https://events.vtools.ieee.org/m/496707