Parameter Distribution Balanced CNNs.

Area of Science:

• Computer Vision
• Deep Learning
• Machine Learning

Background:

• Convolutional Neural Networks (CNNs) are pivotal in computer vision.
• Limited research exists on optimal parameter allocation within CNN convolution layers.
• Enhancing CNN discriminative performance often involves architectural changes, not parameter distribution optimization.

Purpose of the Study:

• To investigate the relationship between CNN parameter distribution and its discriminative performance.
• To propose a method for enhancing CNN performance by optimizing parameter allocation without altering network architecture.
• To establish a guideline for designing CNNs with improved performance under size constraints.

Main Methods:

• Proposed an energy function to quantify CNN parameter distribution.
• Related parameter distribution to CNN discriminative performance.
• Transformed optimal parameter distribution into an energy maximization problem.
• Developed a balanced parameter distribution guideline for CNN design.

Main Results:

• CNN parameter distribution with higher energy values correlates with improved performance.
• Experiments on shallow CNNs and public datasets confirmed the energy function's effectiveness.
• Guideline demonstrated consistent improvements on AlexNet, ResNet34, and ResNet101 on ImageNet under size constraints.

Conclusions:

• CNN parameter distribution significantly impacts discriminative performance.
• The proposed energy function provides a quantifiable link between distribution and performance.
• Balanced parameter distribution offers an effective and simple guideline for designing high-performing CNNs.