What is dropout in neural networks?

• A. A mechanism to permanently disable neurons during training.
• B. A stochastic optimization algorithm variant.
• C. A data augmentation method used during training.
• D. A regularization technique that randomly disables a subset of neurons during training to prevent co-adaptation.

Answer: (D)

Dropout is a regularization technique used during training to prevent overfitting by randomly deactivating a subset of neurons on each training pass. By dropping different neurons each time, the network can’t rely on any specific set of features and must learn redundant representations that work well across many sub-networks. This reduces co-adaptation, making the model more robust and better at generalizing to new data. At inference time, dropout isn’t applied; instead, the network uses all neurons and activations are scaled to account for the dropout during training. Typical dropout rates for hidden layers are around 0.5, though this can vary by architecture.

It’s not a mechanism to permanently disable neurons, not a stochastic optimization algorithm variant, and not a data augmentation technique.