A geometric approach for accelerating neural networks designed for classification problems.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Convolutional Neural Networks (CNNs) are computationally intensive, limiting their deployment on resource-constrained devices.
• Model compression techniques are crucial for accelerating CNN computations and enhancing generalization.
• Existing methods often lack a systematic approach to identify and remove non-informative network components.

Purpose of the Study:

• To propose a novel geometric-based technique for compressing CNNs.
• To accelerate computations and improve generalization by eliminating non-informative network elements.
• To systematically compress convolutional and fully connected layers.

Main Methods:

• Utilizing a geometric index, the separation index, to evaluate network element functionality.
• Applying center-based separation index for optimal compression of convolutional and fully connected layers.
• Developing an algorithm to exclude low-performance layers, select optimal filters, and tune fully connected layer parameters.

Main Results:

• Demonstrated effective compression on CIFAR-10 and ImageNet datasets.
• Achieved significant parameter pruning: 87.5% for VGG16, 77.6% for GoogLeNet, and 78.8% for DenseNet.
• Outperformed state-of-the-art methods in network compression effectiveness.

Conclusions:

• The proposed geometric-based compression technique offers an effective way to reduce CNN complexity.
• The method systematically identifies and removes redundant network components, leading to improved efficiency and generalization.
• This approach provides a valuable tool for deploying deep learning models in practical applications.