Skip to content
Sign in

What is Dropout?

Dropout is a regularization technique used during neural network training that randomly sets a fraction of neurons to zero on each forward pass to reduce overfitting.

During training, dropout randomly drops out (sets to zero) a chosen percentage of neurons in a layer for each mini-batch. This prevents the network from relying too heavily on any single neuron and forces it to learn more robust features.

At inference time, all neurons are used but their outputs are scaled by the dropout probability so the expected value remains the same as during training. This simple change improves generalization without changing the model architecture.

Dropout can be viewed as training an implicit ensemble of many thinned networks that share weights, which is why it often leads to better performance on unseen data.

Example

In a fully-connected layer with 100 neurons and a dropout rate of 0.5, roughly 50 neurons are randomly ignored during each training step. The remaining neurons must still produce useful activations, making the model less sensitive to the removal of any particular neuron.

Why it matters

Dropout remains one of the simplest and most effective ways to combat overfitting in deep networks and is still widely used or adapted in modern architectures such as transformers and CNNs.

Frequently asked questions

It can slightly increase the number of epochs needed, but the extra robustness usually outweighs the modest extra compute.