Q$^2$: Quantization-Aware Gradient Balancing and Attention Alignment for Low-Bit Quantization

Researchers propose Q², a new framework that addresses gradient imbalance issues in quantization-aware training for complex visual tasks like object detection and image segmentation. The method achieves significant performance improvements (+2.5% mAP for object detection, +3.7% mDICE for segmentation) while introducing no inference-time overhead.

• →Current quantization-aware training (QAT) suffers from gradient imbalance at feature fusion stages when applied to complex visual tasks.
• →Q² framework introduces two components: Quantization-aware Gradient Balancing Fusion and Quantization-aware Attention Distribution Alignment.
• →The method achieves +2.5% mAP improvement on object detection and +3.7% mDICE improvement on image segmentation.
• →Q² is a plug-and-play solution that can be integrated into existing QAT pipelines without inference overhead.
• →The approach only applies during training, making it highly practical for real-world deployment scenarios.