Quantized CNN: A Unified Approach to Accelerate and Compress Convolutional Networks

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep neural networks (DNNs), particularly convolutional neural networks (CNNs), are computationally intensive and require high-performance hardware.
• This dependency limits the deployment of DNNs on resource-constrained devices like mobile phones.

Purpose of the Study:

• To present a unified approach for accelerating and compressing convolutional neural networks.
• To enable efficient inference computation on quantized networks with reduced memory and storage footprint.

Main Methods:

• A quantized CNN approach is proposed, minimizing approximation error in both fully connected and convolutional layers.
• The method focuses on careful quantization of network layers for efficient computation.

Main Results:

• The quantized CNN achieves 4-6x acceleration and 15-20x compression compared to standard CNNs.
• Comparable classification accuracy is maintained post-quantization.
• Accurate image classification was demonstrated on mobile devices within 1 second.

Conclusions:

• Quantized CNNs offer a viable solution for accelerating and compressing deep learning models.
• The approach effectively reduces computational and memory requirements, facilitating deployment on mobile devices.
• This work enables efficient and accurate image classification on edge devices.