Batch Normalization (BN) has been widely adopted in deep neural networks for its ability to stabilize training and improve convergence. However, BN relies on batch-wise mean and variance estimates, which can become inaccurate during inference, particularly in Few-shot Learning (FSL) and domain adaptation scenarios where the test distribution differs from training or the available batch size is small. This dependency often causes performance degradation due to mismatched or outdated statistics. In this work, we introduce Meta Affine Transformation (MetaAFN), a batch-statistics-free normalization strategy that replaces the normalization step in BN with a meta-network-generated affine transformation. By entirely removing the reliance on batch statistics, MetaAFN avoids mismatched training-set statistics and instead uses a lightweight meta-network to dynamically produce scale ( γ ) and shift ( β ) parameters conditioned on the current input features. This design enables the model to adaptively modulate representations without explicit BN, improving robustness to distribution shifts. We evaluate MetaAFN on two representative tasks — FSL and source-free domain adaptation — using multiple benchmark datasets. Experimental results show that MetaAFN consistently outperforms or matches BN and MetaBN, with clear advantages under significant distributional shifts. These findings highlight MetaAFN as an effective and practical alternative to BN, offering improved adaptability and generalization in heterogeneous data scenarios.
Relation:
International Journal of Pattern Recognition and Artificial Intelligence 39(16) , p.2551033
