Investigation of Effectiveness of Shuffled Frog-Leaping Optimizer in Training a Convolution Neural Network

Area of Science:

• Computer Vision
• Deep Learning
• Machine Learning

Background:

• Convolutional Neural Networks (CNNs) are leading deep learning architectures for image processing, object detection, and classification.
• CNNs utilize filters, nonlinear functions, and pooling layers, with weight-sharing to manage parameters for spatial and temporal data.
• Training CNNs presents challenges, leading to the exploration of optimization techniques like Ant Colony Optimization and Genetic Algorithms.

Purpose of the Study:

• To investigate the efficacy of the Shuffled Frog-Leaping Algorithm (SFLA) for training a classical CNN structure, LeNet-5.
• To evaluate SFLA's performance across four diverse datasets.
• To compare SFLA against other evolutionary optimization algorithms.

Main Methods:

• The study employed the Shuffled Frog-Leaping Algorithm (SFLA) to train the LeNet-5 Convolutional Neural Network (CNN) architecture.
• Four distinct datasets were utilized to assess the training methodology.
• Performance was benchmarked against Whale Optimization Algorithm (WO), Bacteria Swarm Foraging Optimization (BFSO), and Ant Colony Optimization (ACO).

Main Results:

• The Shuffled Frog-Leaping Algorithm (SFLA) significantly improved the performance of the LeNet-5 CNN.
• The SFLA demonstrated high accuracy in classification and approximation tasks.
• A slight increase in training computation time was observed with the SFLA approach.

Conclusions:

• SFLA is a viable and effective optimization technique for training Convolutional Neural Networks (CNNs).
• The proposed SFLA method offers superior classification and approximation accuracy compared to other tested evolutionary algorithms.
• Future research could explore further optimizations to mitigate the increase in computation time while retaining accuracy benefits.