AI Concepts and Foundations practice questions

It typically has lower inference latency.

Why wrong: Latency depends on implementation, not architecture choice.

It handles novel scenarios better due to joint training.

Why wrong: End-to-end learning may generalize differently, but modular allows targeted improvements.

It requires less engineering effort.

Why wrong: Modular systems often require more engineering for interfaces.

Overfitting

Why wrong: Incorrect; overfitting relates to training vs. generalization performance, not bias.

Model drift

Why wrong: Incorrect; drift refers to changes in data or relationships over time, not inherent bias.

Underfitting

Why wrong: Incorrect; underfitting is about insufficient model capacity.

DBSCAN

Why wrong: DBSCAN is an unsupervised clustering algorithm, not appropriate for classification.

Apriori

Why wrong: Apriori is used for association rule mining, not text classification.

Principal component analysis (PCA)

Why wrong: PCA is for dimensionality reduction, not classification.

Insufficient epochs

Why wrong: Insufficient epochs would show both training and validation loss still decreasing at end, not diverging.

Underfitting

Why wrong: Underfitting would show high loss on both sets, not a gap.

Data leakage

Why wrong: Data leakage would cause unrealistically high validation accuracy from the start.

Retrain the model using only the most recent three months of data.

Why wrong: Retraining is not possible because original training data is unavailable; also, this action contradicts the constraint.

Collect new labeled data and perform transfer learning from the original model.

Why wrong: This requires new data and retraining, which is not feasible given the constraint of no retraining.

Decrease the decision threshold to a lower value, such as 0.3.

Why wrong: Lowering the threshold would classify more customers as churn, increasing false positives.

Transfer learning

Why wrong: Transfer learning is a technique, not a problem description.

Underfitting

Why wrong: Underfitting would show poor performance even on training data.

Bias-variance tradeoff

Why wrong: This describes the balance but not the specific failure to generalize.

Apply L2 regularization

Why wrong: Regularization increases bias to reduce variance.

Increase the training dataset size

Why wrong: More data typically reduces variance, not bias.

Remove irrelevant features

Why wrong: Removing features simplifies the model, increasing bias.

Increase data augmentation

Why wrong: Augmentation helps but dropout is more targeted for overfitting.

Decrease the learning rate

Why wrong: Learning rate affects convergence, not generalization.

Increase the number of layers

Why wrong: More layers increase capacity, likely worsening overfitting.

Expert system

Why wrong: Expert systems rely on predefined rules, same limitation.

Natural language processing (NLP)

Why wrong: NLP is needed but the core improvement is ML.

Robotic process automation

Why wrong: RPA automates repetitive tasks, not conversational AI.

The model has no impact on fraud detection

Why wrong: It does detect some fraud, but misses many.

The model detects all fraudulent transactions

Why wrong: Low recall means many are missed.

Many legitimate transactions are flagged as fraud

Why wrong: That would be low precision, not low recall.

Collect more data

Why wrong: More data can help but is not the first step; regularization is quicker.

Use cross-validation

Why wrong: Cross-validation helps evaluate but does not fix overfitting.

Increase model complexity

Why wrong: This would likely worsen overfitting.

Unsupervised learning

Why wrong: Unsupervised learning finds patterns without labels, not classification.

Reinforcement learning

Why wrong: Reinforcement learning is for agent-based sequential decisions.

Regression

Why wrong: Regression predicts continuous values, not categories.

Feature scaling

Why wrong: Scaling does not address missing values.

One-hot encoding

Why wrong: Encoding is for categorical features, not missing values.

Principal component analysis (PCA)

Why wrong: PCA reduces dimensionality, not imputes missing data.

Cross-validation

Why wrong: Cross-validation is a method to evaluate models, not a cause of bias.

High regularization

Why wrong: Regularization reduces variance, not bias.

Supervised learning is only used for classification

Why wrong: Supervised learning includes regression as well.

Unsupervised learning always requires a target variable

Why wrong: Unsupervised learning has no target variable.

Supervised learning is a subset of reinforcement learning

Why wrong: They are distinct paradigms.

Vanishing gradients

Why wrong: Vanishing gradients would cause slow decrease in loss.

Underfitting

Why wrong: Underfitting would have high training loss as well.

Data leakage

Why wrong: Data leakage would cause abnormally high accuracy.

Training hyperparameters

Why wrong: Hyperparameters are already included.

Model type

Why wrong: Model type is already specified.

Evaluation metrics

Why wrong: Metrics can be defined separately; preprocessing is more fundamental.

Use a different dataset from a similar hospital without checking demographics

Why wrong: Using a different dataset without validation may perpetuate or introduce new biases.

Manually increase the weight of age-related features in the model

Why wrong: Manual weighting can introduce new biases and lacks systematic validation.

Replace the neural network with a decision tree to simplify decision logic

Why wrong: Model simplification does not guarantee fairness and may reduce performance.

Apply matrix factorization with higher latent factors

Why wrong: Matrix factorization still depends on interaction data and may not capture niche item features.

Increase the weight of popular items in the recommendation score

Why wrong: This would further bias toward popular items and worsen the long-tail problem.

Collect more user interaction data over time

Why wrong: More data won't help if niche items inherently have few interactions.

F1 score

Why wrong: The F1 score is incorrect because it balances precision and recall via their harmonic mean, yet in this 99:1 class imbalance, the model’s 99% accuracy is driven by correctly classifying the majority class, leaving recall for fraud near zero. The F1 score would still appear high if precision is also high, masking the failure to detect fraud. It is tempting because F1 is designed for imbalanced datasets where both false positives and false negatives matter equally; it would be correct if the cost of missing fraud and falsely flagging legitimate transactions were symmetric.

Precision

Why wrong: Precision focuses on false positives, not on catching fraud.

Accuracy

Why wrong: Accuracy is misleading on imbalanced datasets.