Related Experiment Video
Updated: Jun 12, 2026
09:49
Assessing Changes in Synaptic Plasticity Using an Awake Closed-Head Injury Model of Mild Traumatic Brain Injury
Published on: January 20, 2023
Efficient and accurate neural-field reconstruction using resistive memory
Yifei Yu1,2,3, Xinyuan Zhang1,2,3, Shaocong Wang1,2,3
1Department of Electrical and Computer Engineering, the University of Hong Kong, Hong Kong, China.
Nature
|June 10, 2026
Summary
This study introduces a new framework for reconstructing signals from incomplete data, improving efficiency and parallelism for AI applications like medical imaging and virtual reality.
Area of Science:
- Computer Science
- Electrical Engineering
- Artificial Intelligence
Background:
- Complex signal reconstruction from sparse data is crucial for AI, medical imaging, and AR/VR.
- Traditional hardware faces limitations in sampling, storage, data movement, and parallel efficiency.
- Existing methods struggle with incomplete measurements and limited computational resources.
Purpose of the Study:
- To present a software-hardware co-optimization framework for sparse-input signal reconstruction.
- To address the challenges of traditional digital hardware in signal reconstruction tasks.
- To enhance energy efficiency and parallelism in AI-driven signal processing.
Main Methods:
- Utilized neural fields for implicit signal representation via neural networks.
- Applied low-rank decomposition and structured pruning for signal compression.
- Developed a resistive-memory-based computing-in-memory platform with a Gaussian encoder and MLP processing engine.
- Implemented a hardware-aware quantization circuit for precise weight mapping.
Main Results:
- Achieved significant projected gains in energy efficiency (23.5×, 21.0×, 32.3×) and parallelism (10.8×, 38.8×, 6.2×) across three applications.
- Demonstrated comparable reconstruction quality to traditional methods.
- Showcased the system's performance on a 40-nm 256 Kb resistive-memory macro.
Conclusions:
- The developed framework offers substantial improvements in energy efficiency and parallelism for sparse-input signal reconstruction.
- This work advances AI-driven signal reconstruction, enabling efficient medical AI and 3D vision applications.
- The resistive-memory-based platform provides a viable solution for resource-constrained AI tasks.
