Convolution Filter Compression via Sparse Linear Combinations of Quantized Basis.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Convolutional Neural Networks (CNNs) excel in real-world tasks but demand substantial computational resources.
• Large parameter counts in CNNs hinder deployment on memory-limited embedded systems.

Purpose of the Study:

• To propose a novel CNN compression method for efficient deployment on resource-constrained devices.
• To significantly reduce storage and computation requirements of CNNs while maintaining accuracy.

Main Methods:

• A novel compression technique generating convolution filters using learnable, low-dimensional quantized filter bases.
• Reconstruction of convolution filters via linear combinations of these filter bases.
• Incorporation of L1-ball projection for coefficient sparsity to reduce storage and prevent overfitting.

Main Results:

• The proposed method achieves a higher compression ratio compared to existing filter decomposition and network quantization techniques.
• Evaluations on image classification and object detection tasks show comparable accuracy to state-of-the-art methods.
• Demonstrated significant reduction in storage and computational needs for CNNs.

Conclusions:

• The novel filter reconstruction method offers an effective approach to compress CNNs.
• This technique enables efficient deployment of deep learning models on embedded devices.
• The method provides a strong balance between model compression and predictive accuracy.