CEModule: A Computation Efficient Module for Lightweight Convolutional Neural Networks.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Lightweight convolutional neural networks (CNNs) are crucial for efficient AI, with lightweight design based on repetitive feature maps (LoR) being a key approach.
• Existing LoR methods focus on feature map regeneration (RO) but neglect issues in feature extraction (CE), such as poor generalization and high computational costs.
• Interpreting CNNs reveals the importance of 'key features' for model performance.

Purpose of the Study:

• To address limitations in the feature extraction (CE) part of lightweight design based on repetitive feature maps (LoR) in CNNs.
• To introduce a novel lightweight convolutional module (LCM) that enhances feature extraction while maintaining efficiency.
• To improve the generalization and reduce the computational workload of lightweight CNNs.

Main Methods:

• Introduced the concept of 'key features' from a CNN model interpretation perspective.
• Developed CEModule, a novel LCM focusing on improving the feature extraction (CE) process.
• Incorporated group convolution for reduced floating-point operations (FLOPs) and a dynamic adaptation algorithm (α-DAM) for enhanced generalization.

Main Results:

• CEModule reduces FLOPs by up to 54% on CIFAR-10 while maintaining classification accuracy.
• The developed CENet, utilizing CEModule, increased accuracy by 1.2% on ImageNet with comparable FLOPs and training strategies.
• The α-DAM algorithm improved the generalization of CEModule-enabled lightweight CNNs across different dataset scales.

Conclusions:

• CEModule offers an effective solution for improving the feature extraction component of lightweight CNNs.
• The proposed method enhances model generalization and significantly reduces computational costs.
• CEModule and CENet represent advancements in efficient and accurate deep learning model design.