Dropout Regularization
Dropout is a regularization technique that randomly deactivates a fraction of a neural network's neurons during each training step, preventing the network from over-relying on any single unit and reducing overfitting.
Definition
Dropout is a regularization technique that randomly deactivates a fraction of a neural network's neurons during each training step, preventing the network from over-relying on any single unit and reducing overfitting.
Overview
Dropout was introduced by Nitish Srivastava, Geoffrey Hinton, and colleagues in a 2014 paper as a computationally cheap alternative to training and averaging many different neural network architectures. During each forward pass in training, every neuron in a dropout layer is independently "dropped" (set to zero) with some fixed probability, typically between 0.2 and 0.5, and the remaining active neurons are rescaled to keep the expected output magnitude constant — a technique known as inverted dropout. Because a different random subset of neurons is active on each training step, no single neuron can become overly specialized or co-dependent on specific other neurons, which in practice is described as reducing "co-adaptation." The overall effect approximates training an ensemble of many thinner sub-networks that share weights, and averaging their predictions at test time. At inference, dropout is turned off entirely and all neurons are used, relying on the rescaling done during training to keep outputs calibrated. Dropout is most commonly applied to fully connected layers, and less commonly to convolutional layers, where alternatives like spatial dropout (dropping entire feature maps) are often preferred. In transformer architectures, dropout is typically applied to attention weights and feed-forward sublayers, though very large language models increasingly train with little or no dropout because their datasets are large enough that overfitting is less of a concern, and dropout can slow convergence. Its main alternatives and complements include L2 regularization, L1 regularization, early stopping, and data augmentation.
Key Concepts
- Randomly zeroes out a fraction of neurons on each training step
- Typical dropout rates range from 0.2 to 0.5 depending on layer and architecture
- Uses inverted dropout to rescale surviving activations and keep expected magnitude stable
- Disabled entirely at inference time, using the full network
- Approximates ensembling many thinner sub-networks with shared weights
- Reduces co-adaptation between neurons, mitigating overfitting
- Commonly applied to fully connected layers; spatial dropout used for convolutional layers
- Often reduced or omitted in very large models trained on massive datasets