Abstract
Artificial intelligence technologies have become prevalent in diversified applications such as clustering and classification. However, their high demands on computing power and resources often become a barrier to deployment, especially in resource-constrained computing environments like edge and embedded devices. To address this issue, unconventional computing methods like hyperdimensional computing (HDC) have emerged. In contrast to multi-pass learning methods such as neural networks, HDC is a one-pass learning method that consumes much less time and energy. In HDC, independent symbols are encoded into high-dimensional random vectors called hypervectors, which are expected to be orthogonal to each other due to the blessing of dimensionality. However, the high dimensionality of hypervectors often amplifies memory space usage and IO traffic, which degrades the HDC performance. In this work, we propose CompactHD, an energy-efficient HDC scheme for classification problems on emerging nonvolatile memories, such as racetrack memories. By encoding hypervectors into smaller but imprecise summaries, irrelevant classes may be efficiently excluded, preventing the expensive comparison of query and test hypervectors. According to experimental results, CompactHD reduces the IO traffic by 37.78% with a minor accuracy loss of 0.16%. When the inference performance is a top-priority concern, CompactHD dramatically reduces the IO traffic by 72.56% with an accuracy loss of 9.03%. The enhanced balance between the inference performance and accuracy makes CompactHD especially suitable for resource-constrained computing environments.
| Original language | English |
|---|---|
| Title of host publication | Proceedings - 2025 14th IEEE Non-Volatile Memory Systems and Applications Symposium, NVMSA 2025 |
| Publisher | Institute of Electrical and Electronics Engineers Inc. |
| Pages | 12-17 |
| Number of pages | 6 |
| ISBN (Electronic) | 9798331585273 |
| DOIs | |
| Publication status | Published - 2025 |
| Event | 2025 14th IEEE Non-Volatile Memory Systems and Applications Symposium, NVMSA 2025 - Singapore, Singapore Duration: 2025 Aug 20 → 2025 Aug 22 |
Publication series
| Name | Proceedings - 2025 14th IEEE Non-Volatile Memory Systems and Applications Symposium, NVMSA 2025 |
|---|
Conference
| Conference | 2025 14th IEEE Non-Volatile Memory Systems and Applications Symposium, NVMSA 2025 |
|---|---|
| Country/Territory | Singapore |
| City | Singapore |
| Period | 25-08-20 → 25-08-22 |
UN SDGs
This output contributes to the following UN Sustainable Development Goals (SDGs)
-
SDG 7 Affordable and Clean Energy
All Science Journal Classification (ASJC) codes
- Computer Science Applications
- Hardware and Architecture
- Electrical and Electronic Engineering
- Safety, Risk, Reliability and Quality
- Electronic, Optical and Magnetic Materials
Cite this
- APA
- Author
- BIBTEX
- Harvard
- Standard
- RIS
- Vancouver