CS5720 - Week 12
Slide 225 of 240

Regularization Techniques Review

Why Regularization Matters

Regularization prevents overfitting by adding constraints or penalties to the model, helping it generalize better to unseen data.
The Overfitting Problem:

• Model memorizes training data instead of learning patterns
• High training accuracy, poor test performance
• Complex models are especially vulnerable
• Regularization provides the solution
⚖️ Key Insight
Regularization is like adding "common sense" to your model - it prevents learning overly specific patterns that don't generalize.

Essential Techniques

🎲 Dropout
Randomly sets neurons to zero during training to prevent co-adaptation.
✓ Simple and effective • Works with any architecture
⚖️ Weight Decay (L2)
Penalizes large weights by adding L2 norm to the loss function.
✓ Prevents weight explosion • Smooth gradients
📊 Batch Normalization
Normalizes layer inputs to stabilize and accelerate training.
✓ Faster training • Less sensitive to initialization
⏹️ Early Stopping
Stops training when validation performance plateaus or degrades.
✓ Prevents overfitting • Saves compute time
🔄 Data Augmentation
Artificially expands training data with transformations.
✓ More robust models • Better generalization

Impact of Regularization on Model Performance

😰
No Regularization
Model overfits to training data
98%
Training Acc
65%
Test Acc
33%
Gap
High
Variance
🤔
Some Regularization
Better but still room for improvement
88%
Training Acc
82%
Test Acc
6%
Gap
Medium
Variance
😊
Optimal Regularization
Well-generalized model
85%
Training Acc
84%
Test Acc
1%
Gap
Low
Variance
Click on any technique or effect to learn more about implementation!
Prepared by Dr. Gorkem Kar