Batch Normalization is a technique used in deep neural networks to improve training speed, stability, and convergence. Its primary purpose is to address the issue of internal covariate shift, which refers to the change in the distribution of each layer's inputs during training due to changes in the previous layer's parameters. This shift can slow down the training process and make it more challenging for each layer to learn effectively.
How Batch Normalization Works
-
Normalization within Mini-batches: During training, batch normalization normalizes the input of each layer by subtracting the mini-batch mean and dividing by the mini-batch standard deviation. This helps in reducing the internal covariate shift, making the network more stable and allowing for faster training.
-
Learnable Parameters: Batch normalization introduces two learnable parameters per activation, typically referred to as scale and shift parameters. These parameters allow the model to adapt and learn the optimal scale and shift for each layer's inputs.
-
Normalization Across Features: In addition to normalizing across the mini-batch dimension, batch normalization also normalizes across the features for each sample within the batch. This normalization is performed independently for each feature dimension.
Impact on Training
-
Faster Convergence: Batch normalization often leads to faster convergence during training by allowing the usage of higher learning rates without the risk of divergence.
-
Reduction in Overfitting: It acts as a form of regularization, reducing the dependence on dropout or other regularization techniques, thereby helping prevent overfitting to some extent.
-
Stability and Gradient Flow: It stabilizes the training process by reducing the likelihood of vanishing or exploding gradients, enabling more robust gradient flow through the network.
Drawbacks and Limitations
-
Batch Size Dependency: Batch normalization's effectiveness can be influenced by the batch size used during training. Very small batch sizes may lead to inaccurate estimates of the mini-batch statistics, affecting its performance.
-
Difficulty in Applying to Some Architectures: Batch normalization might not work optimally with recurrent neural networks (RNNs) due to the sequential nature of their computation.
-
Impact on Inference: During inference, the mean and standard deviation used for normalization must be estimated from the entire training dataset or from running statistics, which could introduce some discrepancy, especially in cases where the inference data distribution significantly differs from the training data.
While batch normalization is a powerful technique and commonly used in many deep learning architectures, its effectiveness can vary based on network architecture, data distribution, and specific use cases. In some scenarios, alternatives like layer normalization or instance normalization might be preferred.
