IBM-style AI Engineer Exam Verified Questions,
Correct Answers, and Detailed Explanations for
Computer Science Students||Already Graded A+
1. Which of the following best describes supervised learning?
A. Learning patterns without labeled outputs
B. Learning from both labeled and unlabeled data
C. Learning a function that maps inputs to labeled outputs
D. Learning via reinforcement feedback
Answer: C
Rationale: Supervised learning uses input–output pairs (features →
labels) and trains a model to predict labels from inputs.
2. In a binary classification model, which metric is most
appropriate when classes are heavily imbalanced and the
positive class is rare but important?
A. Accuracy
B. F1-score
C. Mean Squared Error
D. ROC AUC
Answer: B
Rationale: F1-score balances precision and recall, which is more
meaningful than accuracy when positives are rare.
3. Which of the following is a regularization technique to reduce
overfitting in a neural network?
A. Dropout
B. Batch normalization only
C. Learning rate decay only
D. Data shuffling during training
Answer: A
Rationale: Dropout randomly deactivates neurons during training to
prevent co-adaptation and overfitting.
, 4. When using a pre-trained transformer model for NLP tasks,
what is “fine-tuning”?
A. Re-training the entire model from scratch on a new dataset
B. Adapting only the final layers of the pre-trained model on a
new task
C. Converting the model to use fewer parameters
D. Adding new classes without retraining
Answer: B
Rationale: Fine-tuning means taking a pre-trained model and
updating some (or all) weights on a new task dataset.
5. What is the purpose of a validation set during machine
learning model development?
A. To train the model
B. To evaluate final model performance before deployment
C. To tune hyperparameters and prevent overfitting
D. To collect new data
Answer: C
Rationale: A validation set is used during training to adjust
hyperparameters and detect overfitting before final evaluation.
6. Which technique helps address data skew when a dataset’s
features follow different scales?
A. One-hot encoding
B. Data augmentation
C. Feature scaling (e.g. standardization or normalization)
D. Increasing batch size
Answer: C
Rationale: Feature scaling ensures that features on different scales
contribute proportionally to learning.
7. In the context of computer vision, what does “data
augmentation” commonly refer to?
, A. Adding more labels
B. Transforming existing images (rotations, flips, color
variations)
C. Increasing image resolution only
D. Reducing dataset size
Answer: B
Rationale: Data augmentation creates variations of training images
to improve generalization and avoid overfitting.
8. What is “transfer learning” most useful for?
A. Training a model from scratch on a huge dataset
B. Leveraging knowledge from one task/domain to improve
learning in another with limited data
C. Automatically cleaning data
D. Increasing model interpretability
Answer: B
Rationale: Transfer learning reuses learned patterns from one
domain/task to help learning in a related but data-scarce domain.
9. Which of the following is an unsupervised learning algorithm?
A. Logistic Regression
B. k-means Clustering
C. Decision Tree Classifier
D. Support Vector Machine
Answer: B
Rationale: k-means groups unlabeled data into clusters; no
pre-assigned labels involved.
10. Which loss function is commonly used for regression
tasks?
A. Cross-entropy loss
B. Hinge loss
, C. Mean Squared Error (MSE)
D. Kullback-Leibler divergence
Answer: C
Rationale: MSE measures average squared difference between
predictions and true continuous values, good for regression.
11. In an ML pipeline, where does “feature engineering”
belong?
A. After model evaluation
B. Before model training, during data preprocessing
C. Only after deployment
D. During model inference
Answer: B
Rationale: Feature engineering transforms raw data into features
before training to improve model performance.
12. If a model memorizes the training data but fails to
generalize, this is known as:
A. Underfitting
B. Overfitting
C. Data drift
D. Bias
Answer: B
Rationale: Overfitting is when a model fits noise/peculiarities in
training data and doesn’t generalize to new data.
13. Which of the following best defines “model drift” in a
deployed ML system?
A. The training data changes over time
B. The model’s predictions gradually degrade because data
distribution changes
C. The model weights update themselves automatically
D. The model is re-trained periodically
Correct Answers, and Detailed Explanations for
Computer Science Students||Already Graded A+
1. Which of the following best describes supervised learning?
A. Learning patterns without labeled outputs
B. Learning from both labeled and unlabeled data
C. Learning a function that maps inputs to labeled outputs
D. Learning via reinforcement feedback
Answer: C
Rationale: Supervised learning uses input–output pairs (features →
labels) and trains a model to predict labels from inputs.
2. In a binary classification model, which metric is most
appropriate when classes are heavily imbalanced and the
positive class is rare but important?
A. Accuracy
B. F1-score
C. Mean Squared Error
D. ROC AUC
Answer: B
Rationale: F1-score balances precision and recall, which is more
meaningful than accuracy when positives are rare.
3. Which of the following is a regularization technique to reduce
overfitting in a neural network?
A. Dropout
B. Batch normalization only
C. Learning rate decay only
D. Data shuffling during training
Answer: A
Rationale: Dropout randomly deactivates neurons during training to
prevent co-adaptation and overfitting.
, 4. When using a pre-trained transformer model for NLP tasks,
what is “fine-tuning”?
A. Re-training the entire model from scratch on a new dataset
B. Adapting only the final layers of the pre-trained model on a
new task
C. Converting the model to use fewer parameters
D. Adding new classes without retraining
Answer: B
Rationale: Fine-tuning means taking a pre-trained model and
updating some (or all) weights on a new task dataset.
5. What is the purpose of a validation set during machine
learning model development?
A. To train the model
B. To evaluate final model performance before deployment
C. To tune hyperparameters and prevent overfitting
D. To collect new data
Answer: C
Rationale: A validation set is used during training to adjust
hyperparameters and detect overfitting before final evaluation.
6. Which technique helps address data skew when a dataset’s
features follow different scales?
A. One-hot encoding
B. Data augmentation
C. Feature scaling (e.g. standardization or normalization)
D. Increasing batch size
Answer: C
Rationale: Feature scaling ensures that features on different scales
contribute proportionally to learning.
7. In the context of computer vision, what does “data
augmentation” commonly refer to?
, A. Adding more labels
B. Transforming existing images (rotations, flips, color
variations)
C. Increasing image resolution only
D. Reducing dataset size
Answer: B
Rationale: Data augmentation creates variations of training images
to improve generalization and avoid overfitting.
8. What is “transfer learning” most useful for?
A. Training a model from scratch on a huge dataset
B. Leveraging knowledge from one task/domain to improve
learning in another with limited data
C. Automatically cleaning data
D. Increasing model interpretability
Answer: B
Rationale: Transfer learning reuses learned patterns from one
domain/task to help learning in a related but data-scarce domain.
9. Which of the following is an unsupervised learning algorithm?
A. Logistic Regression
B. k-means Clustering
C. Decision Tree Classifier
D. Support Vector Machine
Answer: B
Rationale: k-means groups unlabeled data into clusters; no
pre-assigned labels involved.
10. Which loss function is commonly used for regression
tasks?
A. Cross-entropy loss
B. Hinge loss
, C. Mean Squared Error (MSE)
D. Kullback-Leibler divergence
Answer: C
Rationale: MSE measures average squared difference between
predictions and true continuous values, good for regression.
11. In an ML pipeline, where does “feature engineering”
belong?
A. After model evaluation
B. Before model training, during data preprocessing
C. Only after deployment
D. During model inference
Answer: B
Rationale: Feature engineering transforms raw data into features
before training to improve model performance.
12. If a model memorizes the training data but fails to
generalize, this is known as:
A. Underfitting
B. Overfitting
C. Data drift
D. Bias
Answer: B
Rationale: Overfitting is when a model fits noise/peculiarities in
training data and doesn’t generalize to new data.
13. Which of the following best defines “model drift” in a
deployed ML system?
A. The training data changes over time
B. The model’s predictions gradually degrade because data
distribution changes
C. The model weights update themselves automatically
D. The model is re-trained periodically
