Adaptive Global Power-of-Two Ternary Quantization Algorithm Based on Unfixed Boundary Thresholds.

Area of Science:

• Edge computing
• Deep learning optimization
• Hardware acceleration

Background:

• Quantizing convolutional neural networks (CNNs) to extremely low bit widths is crucial for reducing storage and computational demands in edge devices.
• However, aggressive quantization often leads to significant drops in detection accuracy due to limited representational capacity and misaligned quantization boundaries.
• Existing methods struggle to balance efficiency gains with accuracy preservation.

Purpose of the Study:

• To propose novel quantization algorithms that minimize accuracy loss in low-bit CNNs for edge computing.
• To develop hardware-friendly quantization techniques that enhance computational efficiency on embedded systems.
• To introduce adaptive quantization methods that accommodate varying weight distributions and bit widths.

Main Methods:

• Adaptive Global Power-of-Two Ternary Quantization (APTQ) quantizes filters into power-of-two binary subfilters with unfixed thresholds.
• The APQ algorithm extends APTQ for arbitrary quantization bit widths.
• Dedicated edge deployment convolutional computation modules were designed for quantized models.

Main Results:

• APTQ and APQ demonstrate superior accuracy performance compared to state-of-the-art quantization algorithms on CIFAR10, CIFAR100, and Mini-ImageNet datasets.
• Achieved minimal accuracy loss, e.g., 0.13% for APTQ ternary ResNet-56 on CIFAR10.
• Designed modules significantly reduce on-chip hardware resource usage for improved computational efficiency.

Conclusions:

• The proposed APTQ and APQ algorithms effectively balance quantization accuracy and deployment efficiency for edge computing.
• Power-of-two quantization and adaptive thresholds address key challenges in low-bit CNN quantization.
• These methods offer valuable insights for optimizing CNNs in industrial edge AI applications.