Logarithmic Learning Differential Convolutional Neural Network.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Convolutional Neural Networks (CNNs) are pivotal in image classification but can incur high computational costs.
• Differential CNNs offer performance improvements but still face calculation challenges.

Purpose of the Study:

• To introduce logarithmic learning into differential CNNs to enhance performance and reduce computational overhead.
• To develop a novel Logarithmic Differential Convolutional Neural Network (LDiffCNN) for more efficient image classification.

Main Methods:

• Integration of LogRelu activation function into CNNs and differential CNNs.
• Development of a Logarithmic Cost Function and a unique logarithmic learning method.
• Evaluation using various datasets and optimizers (SGD/Adam).

Main Results:

• LogRelu integration improved CNN and differential CNN performance by 1.61%–5.44%.
• ResNet architectures with LogRelu showed enhanced top-1 accuracy (3.07%–9.96%).
• The LDiffCNN achieved up to 3.02% higher accuracy than standard CNNs and reduced training iterations by 38%.

Conclusions:

• Logarithmic learning integration effectively addresses the computational cost drawbacks of differential CNNs.
• The proposed LDiffCNN demonstrates superior performance, faster convergence, and reduced training time.
• The study validates the efficiency and benefits of logarithmic approaches in deep learning for image classification.