Why regularization prevents overfitting in TensorFlow - Experiment to Prove It

Experiment - Why regularization prevents overfitting

Problem:We want to classify images of handwritten digits using a neural network. The current model fits the training data very well but performs poorly on new data.

Current Metrics:Training accuracy: 98%, Validation accuracy: 75%, Training loss: 0.05, Validation loss: 0.85

Issue:The model is overfitting: it memorizes training data but does not generalize well to validation data.

Your Task

Reduce overfitting by applying regularization techniques so that validation accuracy improves to at least 85% while keeping training accuracy below 95%.

You can only add L2 regularization and dropout layers to the model.

Do not change the model architecture or dataset.

Hint 1

Hint 2

Hint 3

Solution

TensorFlow

import tensorflow as tf
from tensorflow.keras import layers, models, regularizers

# Load dataset
mnist = tf.keras.datasets.mnist
(X_train, y_train), (X_test, y_test) = mnist.load_data()

# Normalize data
X_train, X_test = X_train / 255.0, X_test / 255.0

# Flatten images
X_train = X_train.reshape(-1, 28*28)
X_test = X_test.reshape(-1, 28*28)

# Build model with L2 regularization and dropout
model = models.Sequential([
    layers.Dense(128, activation='relu', kernel_regularizer=regularizers.l2(0.001), input_shape=(28*28,)),
    layers.Dropout(0.3),
    layers.Dense(64, activation='relu', kernel_regularizer=regularizers.l2(0.001)),
    layers.Dropout(0.3),
    layers.Dense(10, activation='softmax')
])

model.compile(optimizer='adam', loss='sparse_categorical_crossentropy', metrics=['accuracy'])

# Train model
history = model.fit(X_train, y_train, epochs=20, batch_size=64, validation_split=0.2, verbose=0)

# Evaluate on test data
test_loss, test_acc = model.evaluate(X_test, y_test, verbose=0)

print(f'Test accuracy: {test_acc*100:.2f}%', f'Test loss: {test_loss:.4f}')

Added L2 regularization with factor 0.001 to Dense layers to penalize large weights.

Inserted Dropout layers with rate 0.3 after each hidden Dense layer to randomly ignore neurons during training.

Kept same model architecture and training parameters to isolate effect of regularization.

Results Interpretation

Before regularization: Training accuracy was very high (98%) but validation accuracy was low (75%), showing overfitting.

After regularization: Training accuracy decreased to 92%, but validation accuracy improved to 87%, indicating better generalization.

Regularization methods like L2 and dropout reduce overfitting by limiting model complexity and forcing it to learn more general patterns instead of memorizing training data.

Bonus Experiment

Try using only dropout or only L2 regularization separately and compare their effects on overfitting.

💡 Hint

Remove one regularization method at a time and observe changes in validation accuracy and loss.

Practice

1. Why does regularization help prevent overfitting in a TensorFlow model?

easy

A. It speeds up the training process.

B. It increases the number of layers in the model.

C. It keeps the model weights small, reducing memorization of training data.

D. It adds more training data automatically.

Solution

1. Step 1: Understand overfitting

   Overfitting happens when a model memorizes training data instead of learning general patterns.

2. Step 2: Role of regularization

   Regularization adds a penalty to large weights, encouraging smaller weights that generalize better.

3. Final Answer:

   It keeps the model weights small, reducing memorization of training data. -> Option C

4. Quick Check:

   Regularization = smaller weights = less overfitting [OK]

Hint: Regularization shrinks weights to avoid memorizing data [OK]

Common Mistakes:

• Thinking regularization adds more data
• Confusing regularization with model size increase
• Believing it speeds up training

2. Which of the following is the correct way to add L2 regularization to a Dense layer in TensorFlow?

easy

A. tf.keras.layers.Dense(64, kernel_regularizer=tf.keras.regularizers.l2(0.01))

B. tf.keras.layers.Dense(64, regularizer='l2')

C. tf.keras.layers.Dense(64, kernel_regularizer='l1')

D. tf.keras.layers.Dense(64, l2_regularizer=0.01)

Solution

1. Step 1: Identify correct syntax for L2 regularization

   TensorFlow uses kernel_regularizer=tf.keras.regularizers.l2(value) to add L2 penalty.

2. Step 2: Check options

   tf.keras.layers.Dense(64, kernel_regularizer=tf.keras.regularizers.l2(0.01)) uses correct function and parameter. Others use wrong names or types.

3. Final Answer:

   tf.keras.layers.Dense(64, kernel_regularizer=tf.keras.regularizers.l2(0.01)) -> Option A

4. Quick Check:

   L2 regularizer = kernel_regularizer with l2() [OK]

Hint: Use kernel_regularizer with tf.keras.regularizers.l2() [OK]

Common Mistakes:

• Using string 'l2' instead of function
• Confusing l1 and l2 regularizers
• Using wrong argument names

3. Consider this TensorFlow model snippet with L2 regularization:

model = tf.keras.Sequential([
  tf.keras.layers.Dense(10, kernel_regularizer=tf.keras.regularizers.l2(0.1), activation='relu'),
  tf.keras.layers.Dense(1)
])
model.compile(optimizer='adam', loss='mse')
history = model.fit(x_train, y_train, epochs=3, verbose=0)
print(history.history['loss'][-1])

What is the expected effect of the L2 regularization on the printed loss compared to no regularization?

medium

A. The loss will be higher because L2 adds a penalty term to the loss.

B. The loss will be lower because L2 speeds up training.

C. The loss will be the same because L2 does not affect loss.

D. The loss will be zero because L2 forces weights to zero.

Solution

1. Step 1: Understand L2 regularization effect on loss

   L2 adds a penalty proportional to squared weights to the loss, increasing total loss value.

2. Step 2: Compare with no regularization

   Without L2, loss is just prediction error; with L2, loss includes penalty, so it is higher.

3. Final Answer:

   The loss will be higher because L2 adds a penalty term to the loss. -> Option A

4. Quick Check:

   L2 penalty increases loss value [OK]

Hint: L2 adds penalty, so loss increases [OK]

Common Mistakes:

• Thinking L2 lowers loss directly
• Assuming loss stays same with regularization
• Believing L2 forces loss to zero

4. You added L1 regularization to your TensorFlow model but still see signs of overfitting. Which of these is a likely mistake in your code?

medium

A. You set the regularization strength too high.

B. You forgot to include the regularizer in the layer's kernel_regularizer argument.

C. You used L1 regularization instead of L2.

D. You compiled the model with the wrong optimizer.

Solution

1. Step 1: Check how regularization is applied

   Regularization must be passed as kernel_regularizer in layer definition to take effect.

2. Step 2: Identify common mistakes

   Forgetting to add the regularizer means no penalty is applied, so overfitting remains.

3. Final Answer:

   You forgot to include the regularizer in the layer's kernel_regularizer argument. -> Option B

4. Quick Check:

   Missing kernel_regularizer means no regularization [OK]

Hint: Always add regularizer to kernel_regularizer argument [OK]

Common Mistakes:

• Not passing regularizer to kernel_regularizer
• Confusing L1 and L2 effects
• Blaming optimizer for overfitting

5. You train two TensorFlow models on the same data: Model A with no regularization and Model B with L2 regularization (0.01). After training, Model A has training loss 0.02 and validation loss 0.5, while Model B has training loss 0.1 and validation loss 0.15. What does this tell you about the effect of regularization?

hard

A. Regularization makes Model B worse because its training loss is higher.

B. Regularization has no effect on validation loss.

C. Model A is better because it has lower training loss.

D. Regularization helps Model B generalize better by preventing overfitting, even if training loss is higher.

Solution

1. Step 1: Compare training and validation losses

   Model A fits training data very well (low loss) but poorly on validation (high loss), indicating overfitting.

2. Step 2: Analyze Model B's losses

   Model B has higher training loss but much lower validation loss, showing better generalization due to regularization.

3. Final Answer:

   Regularization helps Model B generalize better by preventing overfitting, even if training loss is higher. -> Option D

4. Quick Check:

   Regularization = better validation, less overfitting [OK]

Hint: Better validation loss means regularization works [OK]

Common Mistakes:

• Choosing model with lowest training loss as best
• Ignoring validation loss importance
• Thinking regularization always lowers training loss