Convolution Kernel Operations on a Two-Dimensional Spin Memristor Cross Array.

Area of Science:

• Neuromorphic Engineering
• Computer Engineering
• Materials Science

Background:

• Convolution operations in deep learning are computationally intensive, demanding significant time and energy.
• These operations, crucial for image processing tasks like noise reduction and edge extraction, create substantial memory and communication burdens in data-intensive scenarios.

Purpose of the Study:

• To propose a novel circuit architecture based on a spin memristor crossbar array to address the computational and memory bottlenecks of convolution operations.
• To demonstrate the efficacy of this circuit in accelerating image processing tasks within deep learning algorithms.

Main Methods:

• Development of a logic switch utilizing spin memristors for precise control of the crossbar array.
• Design of a new spin memristor crossbar array and associated peripheral circuits capable of performing multiplication and addition operations inherent to convolution.
• Simulation of color image filtering and edge extraction using the proposed circuit.

Main Results:

• The proposed spin memristor circuit effectively performs multiplication and addition, core components of convolution.
• Significant alleviation of the computational memory bottleneck was observed.
• Image processing simulations for filtering and edge extraction demonstrated the circuit's correctness and effectiveness, validated by Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM) metrics.

Conclusions:

• The spin memristor-based circuit offers a promising solution for accelerating convolution operations in deep learning.
• This approach effectively mitigates the memory/communication burden, paving the way for more efficient AI hardware.
• The validated simulations confirm the circuit's capability for accurate image processing tasks.