Title: Achieving Multi-Domain and Long-Tailed Quantization Through Feature Alignment and Scaling

Original: arXiv:2606.04920v1 Announcement Type: New

Abstract: Deploying deep neural networks on devices with limited resources requires efficient inference, a goal typically achieved through quantization. Despite this, current approaches primarily target single-domain and evenly distributed datasets, often neglecting real-world scenarios characterized by domain shifts or significant class imbalances. To overcome these limitations, we introduce Efficient Multi-Domain Alignment Quantization (EmaQ). This method stabilizes quantization across multiple domains by employing sensitivity-aware weight aggregation and aligning domain distributions via CDF-based projection. Additionally, we adapt EmaQ into EmaQ-LT to handle long-tailed distributions, utilizing class-conditioned variance scaling and confidence-based logit adjustment to reduce the overconfidence of majority classes. Our theoretical framework provides convergence guarantees and justifies the design of the proposed scaling and sensitivity mechanisms. Empirical evaluations on standard benchmarks, as well as multi-domain datasets (Office-31, Digits) and long-tailed datasets (SynDigits-LT, CIFAR-10-LT, CIFAR-100-LT), demonstrate that both EmaQ and EmaQ-LT deliver robust low-bit performance in the presence of domain shifts and class imbalance.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC