L1 and L2 regularization in TensorFlow - Model Metrics & Evaluation

Metrics & Evaluation - L1 and L2 regularization

Which metric matters for L1 and L2 regularization and WHY

L1 and L2 regularization help prevent overfitting by adding a penalty to large weights in the model. To check if regularization works well, we look at validation loss and validation accuracy. If validation loss decreases and accuracy improves or stays stable, regularization is helping the model generalize better to new data.

We also watch training loss and training accuracy. If training loss is low but validation loss is high, the model is overfitting. Regularization aims to reduce this gap.

Confusion matrix example with regularization effect

Without regularization:
          Predicted
          P     N
Actual P 90    30
       N 20    60

With regularization:
          Predicted
          P     N
Actual P 85    25
       N 15    65

Total samples = 200

Here, regularization reduces false positives (FP) and false negatives (FN), improving model generalization.

Precision vs Recall tradeoff with L1 and L2 regularization

L1 regularization tends to produce sparse models by pushing some weights to zero, which can simplify the model and improve interpretability. This might slightly reduce recall but improve precision by focusing on important features.

L2 regularization spreads out the penalty, shrinking weights but rarely to zero. This often improves recall by keeping more features but may reduce precision if noisy features remain.

Choosing between L1 and L2 depends on whether you want a simpler model (L1) or a smoother model (L2). Both help balance precision and recall by reducing overfitting.

Good vs Bad metric values when using L1 and L2 regularization

Good: Validation loss close to training loss, stable or improved validation accuracy, and balanced precision and recall. For example, precision = 0.85, recall = 0.80, F1 = 0.82.

Bad: Large gap between training and validation loss (overfitting), very low recall or precision, or validation accuracy dropping after adding regularization. For example, precision = 0.95 but recall = 0.30 means the model misses many true positives.

Common pitfalls with metrics and regularization

• Accuracy paradox: High accuracy can be misleading if data is imbalanced. Regularization might not fix this alone.
• Data leakage: If validation data leaks into training, metrics look better but model won't generalize.
• Over-regularization: Too strong L1 or L2 can underfit, causing high training and validation loss.
• Ignoring metric tradeoffs: Focusing only on accuracy without checking precision and recall can hide problems.

Self-check question

Your model has 98% accuracy but 12% recall on fraud detection. Is it good for production? Why or why not?

Answer: No, it is not good. The very low recall means the model misses most fraud cases, which is dangerous. Even with high accuracy, the model fails to catch fraud, so it needs improvement, possibly by adjusting regularization or model design.

Key Result

Validation loss and accuracy show if L1/L2 regularization helps reduce overfitting and improves model generalization.

Practice

1. What is the main purpose of using L1 and L2 regularization in TensorFlow models?

easy

A. To change the activation function of the model

B. To increase the size of the model weights

C. To prevent overfitting by adding penalties to large weights

D. To speed up the training process by skipping layers

Solution

1. Step 1: Understand the role of regularization

   L1 and L2 regularization add penalties to the loss function based on the size of the weights to reduce overfitting.

2. Step 2: Identify the effect on model weights

   These penalties discourage large weights, helping the model generalize better on new data.

3. Final Answer:

   To prevent overfitting by adding penalties to large weights -> Option C

4. Quick Check:

   Regularization = Prevent overfitting [OK]

Hint: Regularization adds penalty to weights to avoid overfitting [OK]

Common Mistakes:

• Thinking regularization speeds up training
• Confusing regularization with activation functions
• Believing regularization increases weight size

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

easy

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

B. Dense(64, regularizer=tf.keras.l2(0.01))

C. Dense(64, kernel_regularizer=tf.keras.regularizers.l1(0.01))

D. Dense(64, kernel_regularizer=tf.keras.regularizers.l1_l2(0.01))

Solution

1. Step 1: Recall the correct argument name

   The argument to add regularization in a Dense layer is kernel_regularizer.

2. Step 2: Use the correct regularizer function

   For L2 regularization, use tf.keras.regularizers.l2() with the penalty value.

3. Final Answer:

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

4. Quick Check:

   Correct syntax = Dense(64, kernel_regularizer=tf.keras.regularizers.l2(0.01)) [OK]

Hint: Use kernel_regularizer with tf.keras.regularizers.l2 for L2 [OK]

Common Mistakes:

• Using 'regularizer' instead of 'kernel_regularizer'
• Using L1 regularizer when L2 is needed
• Wrong function call like tf.keras.l2 instead of tf.keras.regularizers.l2

3. What will be the effect on the weights after training a model with L1 regularization compared to no regularization?

medium

A. Weights will be larger to fit the training data better

B. Weights will not change compared to no regularization

C. Weights will be small but rarely zero

D. Weights will be exactly zero for some features, creating sparsity

Solution

1. Step 1: Understand L1 regularization effect

   L1 regularization adds absolute value penalties, encouraging some weights to become zero.

2. Step 2: Compare with no regularization

   Without regularization, weights can be large and non-sparse; L1 creates sparsity by zeroing some weights.

3. Final Answer:

   Weights will be exactly zero for some features, creating sparsity -> Option D

4. Quick Check:

   L1 = Sparsity with zero weights [OK]

Hint: L1 makes some weights zero, creating sparse models [OK]

Common Mistakes:

• Confusing L1 with L2 effects
• Thinking weights get larger with regularization
• Assuming weights remain unchanged

4. You wrote this code to add L2 regularization but get an error:

model.add(Dense(32, kernel_regularizer=tf.keras.regularizers.l2))

What is the error and how to fix it?

medium

A. Missing parentheses after l2; fix with l2(0.01)

B. kernel_regularizer should be regularizer

C. Dense layer does not support regularization

D. l2 should be replaced with l1

Solution

1. Step 1: Identify the error in regularizer usage

   The code uses tf.keras.regularizers.l2 without calling it as a function.

2. Step 2: Correct the syntax by adding parentheses and penalty value

   Use tf.keras.regularizers.l2(0.01) to create the regularizer instance.

3. Final Answer:

   Missing parentheses after l2; fix with l2(0.01) -> Option A

4. Quick Check:

   Call l2() with penalty value [OK]

Hint: Always call l1 or l2 with parentheses and penalty [OK]

Common Mistakes:

• Forgetting parentheses after l1 or l2
• Using 'regularizer' instead of 'kernel_regularizer'
• Thinking Dense layer can't have regularization

5. You want to build a TensorFlow model that uses both L1 and L2 regularization on the kernel weights of a Dense layer with penalties 0.001 and 0.01 respectively. Which code snippet correctly applies this?

hard

A. Dense(128, kernel_regularizer=tf.keras.regularizers.l2(0.01), kernel_regularizer=tf.keras.regularizers.l1(0.001))

B. Dense(128, kernel_regularizer=tf.keras.regularizers.l1_l2(l1=0.001, l2=0.01))

C. Dense(128, kernel_regularizer=tf.keras.regularizers.l1(0.001), bias_regularizer=tf.keras.regularizers.l2(0.01))

D. Dense(128, kernel_regularizer=tf.keras.regularizers.l1(0.001) + tf.keras.regularizers.l2(0.01))

Solution

1. Step 1: Understand how to combine L1 and L2 regularization

   TensorFlow provides l1_l2 regularizer to apply both penalties together on the same weights.

2. Step 2: Check the correct syntax for l1_l2

   Use tf.keras.regularizers.l1_l2(l1=0.001, l2=0.01) as the kernel_regularizer argument.

3. Final Answer:

   Dense(128, kernel_regularizer=tf.keras.regularizers.l1_l2(l1=0.001, l2=0.01)) -> Option B

4. Quick Check:

   Use l1_l2() for combined regularization [OK]

Hint: Use l1_l2() to combine L1 and L2 penalties [OK]

Common Mistakes:

• Trying to add two kernel_regularizer arguments
• Adding regularizers instead of using l1_l2
• Applying L1 to kernel and L2 to bias instead