NEW
Font size
WorksheetsContent Section | Worksheet Questions (Grade 13)
Total questions: 20
Worksheet time: 4mins
What is the main purpose of data augmentation?
Reduce training time
Increase dataset size artificially
Remove noise from data
Reduce model parameters
Data augmentation is most commonly used in which domain?
Text processing
Image processing
Time series analysis
Reinforcement learning
Early stopping is used to prevent:
Underfitting
High bias
Overfitting
Gradient explosion
L1 regularization adds which penalty term to the loss function?
Sum of squared weights
Sum of absolute weights
Maximum weight value
Mean of weights
L2 regularization penalizes:
Absolute value of weights
Number of features
Squared value of weights
Dropout probability
Dropout works by:
Increasing neurons
Removing layers
Randomly disabling neurons during training
Normalizing inputs
Dropout is applied during:
Inference only
Training only
Validation only
Both training and inference
Batch normalization normalizes:
Weights
Loss values
Activations of a layer
Labels
Batch normalization is usually applied:
After activation
Before activation
At output layer only
After loss computation
Which technique randomly drops neurons to reduce overfitting?
L2 Regularization
Batch Normalization
Dropout
Early Stopping
How does data augmentation help generalization?
Increasing model complexity
Exposing model to data variations
Reducing training epochs
Increasing batch size
Which statement correctly compares L1 and L2 regularization?
Both produce sparse weights
L1 produces sparse weights, L2 shrinks weights
L2 sparse, L1 shrinks weights
Both remove weights
Why is dropout turned off during inference?
Speed up computation
Use full network capacity
Reduce memory usage
Avoid normalization
One benefit of batch normalization is:
Increased overfitting
Slower convergence
Reduced sensitivity to initialization
Larger gradients
Early stopping typically monitors:
Training accuracy
Validation loss
Learning rate
Weight decay
Why does L1 regularization lead to sparse solutions?
Squares weights
Uniform scaling
Sharp penalty at zero
Increases learning rate
L2 regularization affects weight updates by:
Increasing magnitude
Pushing weights toward zero
Removing weights randomly
Normalizing gradients
Batch normalization allows higher learning rates because it:
Reduces model size
Stabilizes input distributions of layers
Increases batch size
Removes regularization
Why are dropout and batch normalization sometimes not used together?
Increase training time
Dropout disrupts batch statistics
BN removes dropout effect
Both reduce capacity
When can data augmentation hurt performance?
Distorts true data distribution
Large batch size
L2 regularization used
Early stopping applied
