AML SA1

What evaluation metric would you most likely avoid in an imbalanced binary classification problem?

Accuracy

Dropout is used to increase model complexity.

False

Adam optimizer uses both first and second moment estimates.

True

L1 regularization uses which mathematical approach?

Absolute value of weights

Regularization is most helpful when:

The model is overfitting

Training loss should always be lower than test loss in a good model.

True

A confusion matrix provides insight into:

Classification performance

What type of validation helps reduce model variance by rotating the validation set?

K-fold Cross Validation

Mini-batch gradient descent typically converges faster than batch gradient descent.

True

When is learning rate scheduling particularly useful?

When training loss plateaus

Batch Gradient Descent differs from Mini-Batch in that it:

Uses the entire dataset to compute a single update

What does increasing the dropout rate typically do?

Reduce overfitting by increasing neuron variability

In gradient descent, a smaller learning rate generally leads to:

Slower, more stable convergence

AUC measures the model's ability to classify correctly at various thresholds.

True

Which technique disables random neurons during training?

Dropout

The area under the ROC curve indicates:

Discriminative ability of a model

Regularization increases the model's training accuracy.

False

Feature engineering is not part of the ML pipeline.

False

The Adam optimizer is considered superior to vanilla SGD because it:

Adapts learning rates and includes momentum

What does early stopping monitor to determine when to halt training?

Validation performance

R-squared is a metric used in classification.

False

Cross-validation helps detect if a model is underfitting.

True

Which loss function penalizes larger errors more significantly?

MSE

What is the primary trade-off involved in setting a learning rate too high?

Risk of overshooting the minimum

A regularization technique that results in feature selection is:

Regularization

Validation loss is often used to trigger early stopping.

True

Which of the following describes a characteristic of supervised learning?

It predicts outputs using labeled datasets

A low learning rate can result in slow but stable convergence.

True

ReLU is a commonly used loss function in classification.

False

Overfitting usually occurs when the model is too simple.

False

What is the role of the test set in model development?

To estimate real-world performance

Which of the following optimizers maintains a running average of past squared gradients?

RMSProp

What component in optimization helps reduce oscillation and improve directionality?

Momentum

In classification tasks, which metric is most concerned with minimizing false negatives?

Recall

Which metric is best suited for regression tasks?

Mean Absolute Error

In the machine learning workflow, what is the main purpose of feature engineering?

Enhance data representation for better learning

The main goal of optimization is to increase accuracy on the test set.

False

RMSProp improves over SGD by:

Scaling learning rates by past gradient magnitudes 

RMSProp maintains a history of past gradients.

True

What is the primary reason for using L2 regularization?

It penalizes large weights to reduce overfitting