Memristor-based RDBO-CNN circuit design and application of image multi-classification recognition

Area of Science:

• Hardware Implementation of AI Algorithms
• Neuromorphic Computing
• Optimization Algorithms

Background:

• Traditional convolutional neural networks (CNNs) require extensive hyperparameter tuning and are not readily implementable in hardware.
• Existing dung beetle optimization (DBO) algorithms face challenges with exploration-exploitation balance and local optima, hindering their application in CNN parameter optimization.

Purpose of the Study:

• To propose a novel memristor crossbar architecture circuit for implementing a reinforced dung beetle optimization (RDBO) algorithm integrated with a convolutional neural network (CNN).
• To enhance the DBO algorithm to overcome limitations such as local optima and unbalanced exploration/exploitation for efficient CNN parameter optimization.
• To validate the performance of the proposed RDBO-CNN circuit for image recognition and classification tasks.

Main Methods:

• Development of a memristor crossbar architecture circuit comprising feeding, storage, ball rolling, dance, subpopulation, and CNN modules.
• Implementation of an enhanced dung beetle optimization (RDBO) algorithm incorporating giant dung beetle and spiral search strategies.
• Integration of the RDBO algorithm with a CNN module (convolutional, pooling, and fully connected layers) for image classification.

Main Results:

• The RDBO-CNN circuit demonstrated feasibility and accuracy in image classification tasks on the MNIST dataset.
• Further validation on the satellite image recognition dataset (RSI-CB) confirmed the circuit's effectiveness and good accuracy.
• The proposed hardware implementation offers a viable solution for optimizing and deploying CNNs.

Conclusions:

• The developed memristor-based RDBO-CNN circuit provides an efficient hardware implementation for neural network optimization and image classification.
• The enhanced RDBO algorithm effectively addresses the limitations of traditional DBO, improving CNN parameter tuning.
• This work contributes to the advancement of neural network technology and its hardware applications.