Data Quality-Aware Mixed-Precision Quantization via Hybrid Reinforcement Learning.

Area of Science:

• Deep Learning
• Computer Vision
• Model Quantization

Background:

• Mixed-precision quantization often uses fixed bit-widths, leading to suboptimal performance.
• Conventional methods overlook data quality variations, impacting model robustness in real-world scenarios.

Purpose of the Study:

• To propose a data quality-aware mixed-precision quantization framework (DQMQ) for dynamic bit-width adaptation.
• To improve the robustness and performance of quantized models by considering varying data qualities.

Main Methods:

• DQMQ employs a hybrid reinforcement learning approach, combining model-based policy optimization with supervised quantization training.
• Bit-width sampling is relaxed to a continuous probability distribution, enabling end-to-end differentiable optimization.
• The framework jointly learns a bit-width decision policy with quantization training.

Main Results:

• DQMQ successfully adapts quantization bit-widths to different data qualities, selecting optimal settings per layer.
• Experiments on mixed-quality image datasets demonstrate DQMQ's ability to handle uneven input qualities.
• DQMQ outperforms existing fixed and mixed-precision quantization methods on benchmark datasets and networks.

Conclusions:

• DQMQ offers a superior approach to mixed-precision quantization by incorporating data quality awareness.
• The framework enhances model robustness and performance, particularly in dynamic or real-world application environments.
• DQMQ represents a significant advancement in developing more adaptable and efficient quantized deep learning models.