Related Experiment Video
Updated: Oct 2, 2025
08:07
Assembly and Characterization of Biomolecular Memristors Consisting of Ion Channel-doped Lipid Membranes
Published on: March 9, 2019
8.0K
Long-Term Accuracy Enhancement of Binary Neural Networks Based on Optimized Three-Dimensional Memristor Array.
Jie Yu1,2, Woyu Zhang1,2, Danian Dong1,2
1Key Laboratory of Microelectronics Device & Integrated Technology, Institute of Microelectronics of Chinese Academy of Sciences, Beijing 100029, China.
Micromachines
|February 25, 2022
Summary
Researchers optimized self-selected devices (SSDs) for neural network (NN) edge devices, significantly improving data retention and reducing energy consumption in neuromorphic Internet of Things (IoT) systems.
Area of Science:
- Materials Science
- Computer Engineering
- Artificial Intelligence
Background:
- Embedded neuromorphic Internet of Things (IoT) systems require efficient neural network (NN) inference on edge devices.
- Key challenges include device integration, data retention, and power consumption for edge computing.
- Non-volatile weight storage is crucial for binary neural networks (BNNs) in embedded NN applications.
Purpose of the Study:
- To improve the efficiency and data retention characteristics of self-selected devices (SSDs) for NN edge devices.
- To elucidate the data loss mechanism in SSDs and develop an optimization method to enhance non-volatile weight storage.
- To reduce energy consumption during training while maintaining recognition accuracy in embedded neuromorphic systems.
Main Methods:
- Utilized self-selected devices (SSDs) as base cells for dense three-dimensional (3D) architectures to store non-volatile weights in BNNs.
- Investigated the data loss mechanism in SSDs and introduced a titanium interfacial layer to retain oxygen ions and prevent diffusion.
- Simulated the performance of a 3D VRRAM array constructed with optimized SSDs by mapping pre-trained BNN weights.
Main Results:
- The introduction of a titanium interfacial layer significantly improved the retention characteristics of the SSDs by reducing oxygen ion diffusion.
- Optimized SSDs demonstrated a 24% improvement in long-term recognition accuracy (over 10^5 seconds) for pre-trained BNNs.
- Achieved a 25,000-fold reduction in system energy consumption during training without compromising accuracy.
Conclusions:
- The optimized SSDs offer a high storage density and non-volatile solution for embedded neuromorphic applications.
- This approach effectively addresses the low power consumption and miniaturization requirements of edge computing.
- The study provides a pathway to enhance the performance and energy efficiency of future IoT systems.
