The Image Identification Application with HfO₂-Based Replaceable 1T1R Neural Networks

Area of Science:

• Hardware implementation of artificial intelligence
• Neuromorphic computing
• Memristor-based computing

Background:

• Fully connected neural networks (FCNNs) are crucial for image recognition tasks.
• Traditional hardware implementations face limitations in power consumption and speed.
• Memristor-based computing offers a promising alternative for efficient neural network hardware.

Purpose of the Study:

• To investigate the hardware implementation of an FCNN using one-transistor-one-resistor (1T1R) arrays.
• To evaluate the performance of this hardware in handwritten digital image recognition.
• To analyze the impact of device non-idealities on network accuracy.

Main Methods:

• Fabrication of 1T1R arrays by series connection of memristors and nMOSFETs.
• Establishment of single-layer and double-layer FCNN architectures.
• Testing the network with 8x8 handwritten digital images.
• Simulating network accuracy based on memristor conductivity adjustment range and precision.

Main Results:

• Achieved a recognition accuracy of 95.19% for 8x8 handwritten digital images.
• Determined that stuck-off devices have minimal impact, while stuck-on devices significantly reduce accuracy.
• Simulation results closely matched experimental findings, with less than 1% difference compared to 32-bit floating-point precision.

Conclusions:

• The 1T1R array hardware implementation is effective for handwritten digit recognition.
• Device failure analysis provides insights into the robustness of memristor-based neural networks.
• The proposed hardware demonstrates high accuracy and efficiency comparable to traditional digital systems.