Hessian-based mixed-precision quantization with transition aware training for neural networks.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Model quantization is crucial for deploying deep neural networks on resource-constrained embedded systems.
• Mixed-precision quantization offers performance benefits but often requires extensive search for optimal bit configurations, increasing computational cost.

Purpose of the Study:

• To introduce Hessian-based Mixed-Precision Quantization Aware Training (HMQAT) to reduce the search overhead for optimal bit configurations in quantized neural networks.
• To develop an efficient method for determining the best mixed-precision settings to balance accuracy and model size.

Main Methods:

• HMQAT utilizes a sensitivity metric based on the joint average Hessian trace and parameter size to guide the bit configuration search.
• An automated Pareto frontier method is employed to solve the bit configuration optimization problem.
• Quantization transition-aware fine-tuning of the scale factor is incorporated to maintain inference performance.

Main Results:

• HMQAT significantly reduces the search overhead for mixed-precision quantization.
• Evaluations on ImageNet and CIFAR10 demonstrated substantial model size reduction (e.g., 10.34x for ResNet18 on ImageNet) while preserving high Top-1 accuracy (99.81%).
• The method outperforms existing state-of-the-art mixed-precision techniques in terms of search cost and accuracy-size trade-off.

Conclusions:

• HMQAT provides an efficient and effective approach for mixed-precision quantization-aware training.
• The proposed method enables superior compression of neural networks, facilitating their deployment on lightweight devices.
• This research contributes to the advancement of efficient deep learning inference for embedded applications.