Dropout & Batch Normalization — Checkpoint Quiz

A is correct. By randomly dropping different neurons on every forward pass, dropout forces the network to spread useful information across many neurons rather than depending on a small, fragile set of them, which improves generalization.

A is correct. During training, dropout creates an ensemble-like effect over random sub-networks. At test time you want a single, stable, deterministic prediction using the full network, so dropout is disabled and outputs are rescaled to account for the difference.

A is correct. As weights update during training, the distribution of activations entering each layer keeps shifting, forcing later layers to constantly readapt. Batch normalization re-centers and rescales activations per batch, stabilizing this and generally allowing faster, more stable training.