Regularization techniques like $L1$ and $L2$ work by adding penalties to the loss function based on the magnitude of the network's weights. Dropout, introduced by Srivastava et al. in 2014, takes a fundamentally different approach. Instead of modifying the loss function or the weights directly based on their values, dropout modifies the network architecture itself during training, randomly removing units (along with their connections) from the network for each training step.

The Idea: Training Thinned Networks

Imagine you have a large neural network. During each training iteration (typically for each mini-batch), dropout randomly "drops" or deactivates a certain fraction of the neurons in a layer. This means these neurons, for that specific training iteration, do not participate in the forward pass (their output is effectively zero) and do not contribute to the backpropagation step (no gradient is calculated for them).

Consider a layer with neurons. For each neuron, a random decision is made: it is kept active with a probability $p$ (the "keep probability"), or it is deactivated (dropped) with probability $1-p$ . This process is repeated for every training example or mini-batch, so the specific set of neurons that are active changes constantly.

A simplified view of a network layer before and during a single dropout training step. Greyed-out nodes (H2, H4) are randomly deactivated for this iteration.

Why Does Dropout Work?

Dropout acts as a powerful regularizer for several reasons:

Prevents Co-adaptation: Since any neuron can be randomly dropped, other neurons cannot rely too heavily on its presence. This forces each neuron to learn features that are more robust and useful in conjunction with different random subsets of other neurons. It discourages complex co-adaptations where multiple neurons only work well together in a very specific way, which can be a sign of overfitting to the training data.
Encourages Redundancy: The network learns redundant representations. If one pathway or neuron is dropped, other pathways or neurons should still be able to produce a reasonable output.
Implicit Ensemble: Training a network with dropout can be interpreted as training a large collection of "thinned" networks (networks with different subsets of neurons removed). At test time, using the full network approximates averaging the predictions of this vast ensemble of thinned networks, which generally improves performance and generalization.

Implementation: Keep Probability and Inverted Dropout

Dropout is controlled by a hyperparameter, typically the "keep probability" $p$ , which is the probability that any given neuron is kept active. Alternatively, implementations might use the "dropout rate" ( $1-p$ ), the probability that a neuron is dropped. Common values for $p$ are between 0.5 and 0.8 for hidden layers. A value of $p=1.0$ means no dropout is applied.

A crucial detail is how to handle the network during evaluation or prediction (inference time). We want to use the full network capacity then, so we don't randomly drop neurons. However, simply using all neurons would mean the expected output magnitude of a layer during inference would be higher than during training (since more neurons are active).

To compensate for this, a technique called inverted dropout is commonly used. During the training phase, after applying dropout (setting some activations to zero), the activations of the remaining active neurons are scaled up by dividing by the keep probability $p$ .

\text{Activation}_{scaled} = \frac{\text{Activation}_{original}}{p} \quad (\text{if neuron kept})

\text{Activation}_{scaled} = 0 \quad (\text{if neuron dropped})

By doing this scaling during training, the expected sum of activations remains consistent between training and inference. At test time, dropout is simply turned off, and no scaling is needed because the weights have already been trained to account for the fact that, on average, only a fraction $p$ of neurons were active during training.

Applying Dropout in Practice

Dropout is typically applied to the outputs of hidden layers, often the fully connected layers in a network, as these tend to have the most parameters and are prone to overfitting. It's less common, though not unheard of, to apply it directly to input layers (where it might discard input features) or convolutional layers (where specialized dropout variants sometimes perform better).

Most deep learning frameworks provide straightforward implementations:

# Conceptual Example using a Keras-like API
import tensorflow as tf

model = tf.keras.Sequential([
    tf.keras.layers.Dense(128, activation='relu', input_shape=(input_dim,)),
    tf.keras.layers.Dropout(0.5), # Apply dropout with rate 0.5 (p=0.5)
    tf.keras.layers.Dense(64, activation='relu'),
    tf.keras.layers.Dropout(0.5), # Apply dropout again
    tf.keras.layers.Dense(num_classes, activation='softmax')
])

# During model.fit(), dropout is active.
# During model.predict() or model.evaluate(), dropout is automatically disabled.

In summary, dropout is an effective and computationally inexpensive regularization technique that prevents overfitting by randomly disabling neurons during training, forcing the network to learn more robust and less interdependent features. Its implementation via inverted dropout ensures consistency between training and inference phases.