Loading...
Search
Search in this resource
sort by
Send to Friend
 Digital Object List
 Bookmark
 Review List
  1. ارائه بستر پردازشی بهینه برای سیگنال‌ های زیستی مبتنی بر مدل‌ های پردازشی ابربعدی با استفاده از پردازش درون حافظه
ارائه بستر پردازشی بهینه برای سیگنال‌ های زیستی مبتنی بر مدل‌ های پردازشی ابربعدی با استفاده از پردازش درون حافظه
24 viewed

ارائه بستر پردازشی بهینه برای سیگنال‌ های زیستی مبتنی بر مدل‌ های پردازشی ابربعدی با استفاده از پردازش درون حافظه

جسوری، هادی Jasouri, Hadi

Proposing an Optimized Hyper-Dimensional Computing Platform Customized for Biosignals using In-Memory Computing

Jasouri, Hadi | 2025

24 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 58624 (19)
  4. University: Sharif University of Technology
  5. Department: Computer Engineering
  6. Advisor(s): Koohi, Somayyeh
  7. Abstract:
  8. The use of hyperdimensional computing models in edge applications and biosignal processing has been steadily growing due to their inherent robustness against noise and the simplicity of their operations. Nevertheless, these models face a fundamental challenge: extremely high dimensionality and the associated hardware cost during inference, which lead to substantial energy consumption and reduced efficiency. In contrast, processing-in-memory (PIM) architectures have recently emerged as a promising approach to reduce data movement and accelerate computation; however, they require algorithms that can adapt to their structural constraints. To tackle this issue, we propose ADAPT, a framework that employs adaptive dimension-level pruning. ADAPT not only eliminates redundant and non-informative features but also restores saturated dimensions when necessary. Moreover, we extend for the first time the concept of eliminating all-zero and all-one columns from the CMOS architecture domain to the PIM paradigm, thereby reducing both computation and memory-access overheads. By combining these mechanisms, the model dynamically adapts to data characteristics and hardware requirements, striking an effective balance between accuracy and efficiency. Extensive experiments on ISOLET, UCIHAR, and MNIST datasets demonstrate that the enhanced framework, Adaptive-Pruning, consistently outperforms baseline approaches by simultaneously achieving higher accuracy (up to a ∼2% improvement) and greater sparsity. These algorithmic gains directly translate into hardware-level benefits, yielding up to 65% energy savings and 2.87× speedup on ISOLET, 61% energy savings and 2.58× speedup on UCIHAR, and 76% energy savings and 4.24× speedup on MNIST. Overall, ADAPT bridges the gap between algorithm and hardware, underscoring the potential of algorithm–architecture co-design as an effective pathway toward developing intelligent, energy-efficient, and low-power PIM-based systems for humancentric and biomedical signal processing
  9. Keywords:
  10. Hyperdimensional Computing (HD) ; In-Memory Computing ; Pruning Method ; Energy Efficiency ; Biosignal Processing ; Adaptive Pruning ; Algorithm–Architecture Co-Design

 Digital Object List

 Bookmark

...see more