Shallowing Deep Networks: Layer-Wise Pruning Based on Feature Representations.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Convolutional Neural Networks (CNNs) have achieved significant success in various applications.
• The increasing complexity of CNNs leads to high computational costs, limiting their deployment on resource-constrained devices.

Purpose of the Study:

• To propose a novel feature representation based layer-wise pruning method for CNNs.
• To reduce the computational cost and model size of complex CNNs without compromising performance.

Main Methods:

• A layer-wise pruning strategy is employed, focusing on feature representation within convolutional layers.
• Redundant parameters are identified by investigating learned features, differing from traditional weight-based pruning methods.
• The pruning process is conducted at the layer level to optimize network compactness.

Main Results:

• The proposed method significantly reduces the computational cost of CNN models.
• Pruned models demonstrate equivalent or improved performance compared to original models across various datasets.
• The approach effectively achieves model compression through feature-based layer pruning.

Conclusions:

• Feature representation based layer-wise pruning is an effective technique for creating compact CNNs.
• This method addresses the challenge of deploying complex models on devices with limited computational power.
• The proposed approach offers a viable solution for efficient deep learning model deployment.