Dropout is a regularization technique for neural networks in which random neurons are temporarily “switched off” during training, forcing the network to learn more robust, redundant patterns.
What it means in plain English
During training with dropout, a random subset of the network’s neurons is ignored on each pass. This prevents the network from becoming overly reliant on any single neuron or path, because it can never count on all of them being present. The result is a more robust network that generalises better and is less prone to overfitting. At prediction time, all neurons are used again.
It is a simple but highly effective and widely used technique.
A simple example
Imagine training a team where, each practice, a few random members sit out. Everyone learns to cover multiple roles, so the team stays strong no matter who is missing. Dropout does the same for a neural network’s neurons.
Why it matters
Dropout is one of the most popular and effective ways to combat overfitting in deep learning. Its simplicity and effectiveness made it a standard component in training robust neural networks.
Related terms
- Regularization — the category dropout belongs to.
- Overfitting — the problem dropout helps prevent.
- Neural Network — where dropout is applied.
Frequently asked questions
What does dropout do?
During training it randomly “drops” (ignores) some neurons on each pass, which prevents the network from relying too heavily on any single neuron and reduces overfitting.
Is dropout used during prediction?
No — dropout is only applied during training. At prediction time the full network is used, so it behaves consistently.
