Mitigating Class Imbalance in Neural Network Quantization via Rank-Based Variance Regularization

Citations

WEB OF SCIENCE

0
Citations

SCOPUS

0

초록

While Quantization Aware Training (QAT) improves inference efficiency, it struggles with performance degradation on class-imbalanced datasets due to limited representation of tail classes and an overemphasis on head class optimization. To address the issue, we propose a Rank-based variance regularization method, which introduces a regularization term for variance consistency across classes and applies a rank-based statistical test for outlier robustness. Our method also uses a class-inverse scaling strategy to enhance tail class representation. Experimental results show that our approach outperforms existing methods up to 6.3% and achieves stable performance across head and tail classes under severe class imbalance and high compression rates.

키워드

Class Imbalance; Neural Network Quantization; Variance Regularization
제목
Mitigating Class Imbalance in Neural Network Quantization via Rank-Based Variance Regularization
저자
Kim, Suk Hyun; Youn, Jongsu; Yoon, Yeonghun; Choi, Jongwon
DOI
10.1007/978-3-032-31930-2_40
발행일
2027
유형
Proceedings Paper
저널명
Lecture Notes in Computer Science
권
16825
페이지
593 ~ 607