Training Lightweight Deep Convolutional Neural Networks Using Bag-of-Features Pooling

Area of Science:

• Computer Vision
• Deep Learning
• Machine Learning

Background:

• Convolutional Neural Networks (CNNs) excel in computer vision but demand significant computational resources for training and deployment.
• Existing CNN architectures often lack efficiency, posing challenges for hardware-constrained applications.

Purpose of the Study:

• To develop a novel quantization-based pooling method for creating more lightweight and efficient CNNs.
• To enable end-to-end trainability and native classification of variable-sized images within CNN architectures.

Main Methods:

• A bag-of-features-inspired approach using trainable radial basis function neurons for quantizing convolutional layer activations.
• Implementation of differentiable quantization and aggregation layers for seamless integration into CNNs.
• Introduction of a fast linear variant for enhanced understanding of convolutional neural architectures.

Main Results:

• The proposed method significantly reduces the number of parameters in CNNs, leading to smaller model sizes.
• Demonstrated improved performance compared to competitive methods across seven datasets and three learning tasks (classification, regression, retrieval).
• Enabled native classification of images with various sizes without requiring input resizing.

Conclusions:

• The proposed quantization-based pooling method offers an effective strategy for developing lightweight and high-performing CNNs.
• This approach provides a viable solution for deploying deep learning models on resource-limited hardware.
• The method contributes to a deeper understanding of convolutional neural network architectures and their optimization.