A Hardware-Friendly Low-Bit Power-of-Two Quantization Method for CNNs and Its FPGA Implementation.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Convolutional Neural Networks (CNNs) face challenges in balancing accuracy, efficiency, and hardware resource consumption (DSPs, RAMs on FPGAs), limiting their industrial applicability.
• Existing quantization methods struggle to meet the stringent demands of industrial applications for low-bit CNNs.

Purpose of the Study:

• To propose an innovative low-bit power-of-two quantization method, Global Sign-based Network Quantization (GSNQ), to address the resource and accuracy limitations of CNNs.
• To develop a novel convolutional algorithm for efficient FPGA deployment of quantized CNN models.
• To validate the effectiveness of GSNQ in reducing accuracy loss and improving hardware adaptability.

Main Methods:

• Introduced Global Sign-based Network Quantization (GSNQ), a method that quantizes weights based on their sign to expand the value range.
• Developed a fine-grained, multi-scale global retraining strategy to mitigate accuracy degradation from low-bit quantization.
• Proposed a novel convolutional algorithm replacing multiplication with shift operations for FPGA implementation.

Main Results:

• GSNQ demonstrated superior accuracy compared to existing methods across various CNN architectures (LeNet-5, AlexNet, VGG-Net, ResNet, GoogLeNet).
• Achieved "lossless" quantization (accuracy higher than baseline) in most low-bit scenarios.
• FPGA experiments confirmed the convolutional algorithm's minimal DSP usage and low occupancy of LUTs/FFs, enhancing computational parallelism and demonstrating strong hardware adaptability.

Conclusions:

• GSNQ effectively reduces hardware resource requirements for CNNs on FPGAs without compromising accuracy, often improving it.
• The proposed method and algorithm provide a viable solution for deploying efficient and accurate CNNs in industrial settings.
• GSNQ offers theoretical and experimental validation for the practical industrial application of low-bit quantized CNNs.