Transform Quantization for CNN Compression.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Convolutional Neural Networks (CNNs) require significant computational resources.
• Existing CNN quantization methods often yield suboptimal performance or are limited to training-time compression.
• Efficient compression of pre-trained CNN models is crucial for deployment on resource-constrained devices.

Purpose of the Study:

• To develop a novel post-training compression technique for CNNs.
• To improve compression efficiency and performance at low bit-rates.
• To unify quantization and dimensionality reduction into a single framework.

Main Methods:

• Developed a rate-distortion framework for CNN quantization.
• Introduced transform quantization, combining decorrelation and quantization.
• Derived an optimal End-to-end Learned Transform (ELT) for weight transformation.
• Applied optimal bit-depth allocation post-decorrelation.

Main Results:

• Transform quantization significantly advances the state of the art in CNN compression.
• Achieved superior compression at any given bit-rate compared to previous methods.
• Demonstrated effective compression of models like AlexNet, ResNet, and DenseNet to 1-2 bits.
• Showcased improved performance in both retrained and non-retrained quantization scenarios.

Conclusions:

• Transform quantization offers a unified and effective approach for CNN compression.
• The proposed method enables efficient low bit-rate compression and inference in the transform domain.
• This technique facilitates the deployment of large CNN models on edge devices.