배치 정규화와 층 정규화

As signals flow through layers, their scale drifts and training slows. Normalization keeps activations in a stable, well-behaved range. ⚖️

Batch normalization rescales each feature using the mean and variance computed across the current mini-batch.

You drop in a batch norm layer after a linear or conv layer. Use nn.BatchNorm1d for dense features and BatchNorm2d for images.

Batch norm adds two trainable parameters, gamma and beta, so the network can rescale and shift the normalized output as it needs.

By steadying activations, batch norm allows higher learning rates and deeper networks that would otherwise be hard to train.

During training, batch norm uses batch statistics; at eval it switches to running averages, so call model.eval() before testing.

Batch norm depends on the batch. With very small batches the statistics get noisy and the benefit can disappear.

Layer normalization normalizes across the features of a single sample, so it does not depend on the batch at all.

Layer norm shines in transformers and RNNs. You add it with nn.LayerNorm, passing the size of the features to normalize.

Because layer norm uses per-sample statistics, it behaves the same in training and eval, which suits variable batch sizes.

Reach for batch norm in conv vision models and layer norm in sequence and transformer models, where batches and lengths vary.

You compared normalization: batch norm uses batch statistics and helps CNNs, while layer norm works per-sample and powers transformers. 🚀