How Batch Normalization parameters enable parameter-efficient CNN transfer learning via adaptive feature selection

抄録

Transfer learning with pre-trained models is essential in image classification, yet understanding why specific parameters enable effective adaptation remains limited. We investigate the role of Batch Normalization (BN) γ parameters in transfer learning, where training only BN parameters and the classifier achieves performance comparable to full fine-tuning. Through systematic experiments across multiple domains combining ablation studies, distribution analysis, overlap rate analysis, and correlation studies, we demonstrate that γ parameters function as adaptive feature selectors. Our findings reveal dataset-specific yet highly reproducible feature selection patterns, with quantitative evidence consistently supporting γ’s capability for deterministic task-specific feature selection from pre-trained models.