Applying Regularization: Dropout and Weight Decay

As you've learned to build and train models, you might encounter a common challenge: your model performs exceptionally well on the data it was trained on, but its performance drops significantly when faced with new, unseen data. This phenomenon is known as overfitting, and it indicates that your model has learned the training data's noise and specific quirks rather than the underlying general patterns. Regularization techniques are designed to combat overfitting, helping your models generalize better to new data. This section focuses on two widely used regularization methods in deep learning: dropout and weight decay (often referred to as $L_2$ regularization).

Understanding Overfitting

Before we discuss the solutions, let's visualize what overfitting looks like. When a model overfits, it's essentially memorizing the training set. Its capacity is so high, or it has been trained for too long on the same data, that it starts fitting to the random fluctuations present in the training samples. This leads to complex models that don't perform well on data they haven't seen before.

The following chart illustrates the typical behavior of training and validation loss for an overfit model versus a well-regularized one.

Behavior of training and validation loss curves. Overfitting occurs when validation loss starts to increase while training loss continues to decrease. Regularization aims to keep both losses low and converging.

Notice how the overfit model's validation loss starts to increase after a certain point, even as its training loss continues to decrease. A regularized model, however, tends to show better alignment between training and validation loss, or at least the validation loss plateaus at a lower value.

Dropout

Dropout is a simple yet effective regularization technique introduced by Srivastava et al. (2014). During each training iteration, dropout randomly sets the outputs of a fraction of neurons in a layer to zero. This "dropping out" of neurons changes the network architecture slightly for each training batch.

Why does this help?

1. Prevents Co-adaptation: Neurons learn to be more independent. They cannot rely on specific other neurons being active because any neuron might be dropped out. This encourages them to learn features that are individually more useful.
2. Ensemble Effect: Training with dropout can be seen as training a large number of "thinned" networks (networks with a subset of neurons). At test time, using the full network can be interpreted as averaging the predictions of these many thinned networks, which often improves performance.

Implementation in Flux.jl

Flux.jl provides the Dropout(p) layer, where p is the probability that each neuron's output is set to zero during training. It's typically inserted between other layers in your model, often after activation functions in fully connected layers.

using Flux

# Define a model with Dropout
model = Chain(
  Dense(784, 256, relu),
  Dropout(0.5),  # Apply dropout with a 50% probability
  Dense(256, 128, relu),
  Dropout(0.3),  # Apply dropout with a 30% probability
  Dense(128, 10)
)

During training, Flux automatically enables dropout. When you evaluate the model (e.g., using model(x) outside a training loop or explicitly with Flux.testmode!), dropout is automatically disabled, and the layer outputs are scaled appropriately to account for the neurons that were dropped during training. This scaling ensures that the expected sum of inputs to the next layer remains the same during training and inference. Specifically, the outputs of the remaining active neurons are scaled up by a factor of $1/(1-p)$.

The dropout rate p is a hyperparameter you'll need to tune. Common values range from 0.2 to 0.5. A higher p means more aggressive regularization.

Weight Decay (L2 Regularization)

Weight decay, also known as $L_2$ regularization, is another common technique to prevent overfitting. It works by adding a penalty term to the model's loss function. This penalty is proportional to the sum of the squares of the model's weights.

The modified loss function becomes:

$$L_{total} = L_{original} + \frac{\lambda}{2} \sum_{i} w_i^2$$

Here, $L_{original}$ is the original loss (e.g., cross-entropy or mean squared error), $w_i$ are the individual weights in the model, and $\lambda$ (lambda) is the regularization strength or weight decay coefficient. The factor of $1/2$ is often included for mathematical convenience when taking derivatives.

Why does this help?

1. Discourages Large Weights: Large weights can make the model very sensitive to small changes in the input, potentially leading to sharp decision boundaries and overfitting to noise. By penalizing large weights, $L_2$ regularization encourages the model to find solutions with smaller, more distributed weights.
2. Simpler Models: Models with smaller weights are often considered "simpler" and tend to generalize better. The penalty term pushes weights towards zero, but not exactly to zero unless a weight is truly uninformative.

Implementation in Flux.jl

In Flux.jl, weight decay is typically applied as part of the optimizer. The Optimiser struct can wrap an existing optimizer with a WeightDecay component.

using Flux
using Flux: Optimise

# Original optimizer
opt_rule = Adam(0.001) # Learning rate of 0.001

# Add weight decay
lambda = 0.01 # Weight decay coefficient
opt = Optimiser(opt_rule, WeightDecay(lambda))

# Example model
model = Dense(10, 5)
ps = Flux.params(model)
gs = gradient(() -> Flux.mse(model(rand(10)), rand(5)), ps)

# Update parameters using the optimizer with weight decay
Flux.update!(opt, ps, gs)

In this setup, during the parameter update step, the optimizer will not only move parameters in the direction that minimizes $L_{original}$ but also implicitly reduce the magnitude of the weights due to the $L_2$ penalty term. The gradient of the $L_2$ penalty with respect to a weight $w$ is $\lambda w$. So, the update rule effectively becomes $w \leftarrow w - \eta (\frac{\partial L_{original}}{\partial w} + \lambda w)$, which causes weights to decay towards zero.

The weight decay coefficient $\lambda$ is a hyperparameter. Typical values are small, such as $10^{-2}$, $10^{-4}$, or $10^{-5}$. Finding the right value often requires experimentation.

Choosing and Using Regularization

Both dropout and weight decay are powerful tools for improving model generalization.

• Dropout is particularly effective in large, deep networks where co-adaptation of features can be a problem. It's common in fully connected layers.
• Weight decay ($L_2$) is a more general regularizer that can be applied to almost any model with learnable weights.

It's also possible, and sometimes beneficial, to use both techniques together. The optimal strength for each (the dropout probability p and the weight decay coefficient $\lambda$) will depend on your specific dataset and model architecture. These are hyperparameters that you'll typically tune using a validation set, a topic we'll cover in more detail when discussing hyperparameter tuning strategies.

When applying regularization, it's important to monitor both training and validation performance.

• If your model is still overfitting (validation loss much higher than training loss, or validation loss increasing), you might need to increase the regularization strength (e.g., higher p for dropout, larger $\lambda$ for weight decay).
• Conversely, if your model is underfitting (both training and validation losses are high and not improving, or validation performance is poor), you might be applying too much regularization. In this case, try reducing the regularization strength or even removing it.

By carefully applying regularization techniques like dropout and weight decay, you can build deep learning models in Julia that perform well not just on the data they've seen, but also on new, unseen examples. This is a significant step towards developing reliable and effective machine learning solutions.