MBFQuant: A Multiplier-Bitwidth-Fixed, Mixed-Precision Quantization Method for Mobile CNN-Based Applications.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deploying Convolutional Neural Networks (CNNs) on mobile devices is hindered by computational constraints.
• Network quantization reduces CNN computational load but can degrade accuracy, particularly in compact architectures.
• Existing quantization methods often struggle to balance efficiency and accuracy for mobile CNNs.

Purpose of the Study:

• To present MBFQuant, an efficient mixed-precision quantization pipeline for CNNs on mobile platforms.
• To address the accuracy degradation issue in quantized compact CNN architectures.
• To redefine the mixed-precision quantization design space for improved performance.

Main Methods:

• MBFQuant fixes the multiplier bitwidth, focusing on the sum of weight and activation quantization bitwidths.
• A Simulated Annealing (SA)-based optimizer is employed to automatically search and determine optimal bitwidth assignments.
• The pipeline is evaluated across ten CNN architectures and four datasets.

Main Results:

• MBFQuant achieves significant accuracy improvements compared to uniform bitwidth quantization.
• Up to 19.34% accuracy gain in image classification tasks.
• Up to 1.12% accuracy gain in object detection tasks, while maintaining running efficiency.

Conclusions:

• MBFQuant offers an effective solution for deploying accurate CNNs on resource-constrained mobile devices.
• The proposed fixed-multiplier bitwidth strategy and SA optimizer successfully navigate the trade-off between quantization and accuracy.
• This pipeline demonstrates superior performance for mobile AI applications.