Batch Normalization (BatchNorm) is a widely adopted technique that enables faster and more stable training of deep neural networks (DNNs). Despite its pervasiveness, the exact reasons for BatchNorm's effectiveness are still poorly understood. The popular belief is that this effectiveness stems from controlling the change of the layers' input distributions during training to reduce the so-called "internal covariate shift". In this work, we demonstrate that such distributional stability of layer inputs has little to do with the success of BatchNorm. Instead, we uncover a more fundamental impact of BatchNorm on the training process: it makes the optimization landscape significantly smoother. This smoothness induces a more predictive and stable behavior of the gradients, allowing for faster training.
核心贡献 · Key contributions
证明 BatchNorm 的成功并非源于减少内部协变量偏移。 Demonstrates that BatchNorm's success is not due to reducing internal covariate shift.
发现 BatchNorm 平滑优化曲面，提升梯度的可预测性。 Identifies that BatchNorm smooths the optimization landscape, improving gradient predictiveness.
表明 BatchNorm 改善损失和梯度的 Lipschitz 连续性，允许更大学习率。 Shows that BatchNorm improves Lipschitzness of loss and gradients, enabling larger learning rates.