AI0-001 Practice Test 47 — 15 Questions

Question 1

A healthcare organization wants to use patient data to predict disease risk. They are concerned about bias in the model. Which step is most critical during the data preparation phase to mitigate bias?

Accepted Answer

Ensuring the training data is representative of the target population. Ensuring the training data is representative of the target population is the most critical step during data preparation to mitigate bias because bias often originates from skewed or incomplete data that does not reflect the real-world distribution of patient demographics, conditions, and outcomes. Without a representative dataset, any subsequent preprocessing or modeling will propagate and potentially amplify existing disparities, leading to unfair or inaccurate predictions for underrepresented groups.

Answer

Applying SMOTE to oversample minority classes

Answer

Using a more complex algorithm

Answer

Removing all demographic features

Question 2

A security team needs to ensure that all data used for AI model training in the cloud is encrypted at rest and in transit. Which set of measures meets this requirement on AWS?

Accepted Answer

Enable S3 default encryption with SSE-S3 and use HTTPS for API calls. Encryption at rest is achieved by enabling default encryption on S3 (SSE-S3 or SSE-KMS), and encryption in transit is achieved by enforcing HTTPS/TLS for all API calls. Together, these two measures directly satisfy the requirement for data used in AI training on AWS.

Answer

Use Security Groups and Network ACLs

Answer

Use client-side encryption and store keys in AWS Secrets Manager

Answer

Enable VPC peering and use VPN connections

Question 3

Which TWO of the following are common methods for mitigating bias in AI models?

Accepted Answer

Reweighting training samples based on sensitive attributes. Option B (Reweighting training samples based on sensitive attributes) is correct because it is a standard pre-processing bias-mitigation technique: by assigning higher weights to underrepresented or historically disadvantaged groups, the model's loss function is adjusted so those samples contribute more to the learned parameters, reducing disparate impact across sensitive attributes. Option D (Adding fairness constraints during training) is correct because it is a standard in-processing technique: fairness metrics such as demographic parity, equalized odds, or disparate impact are encoded as constraints or penalty terms in the objective function, forcing the optimizer to trade off accuracy against a quantified fairness criterion. The other options do not belong: A (adversarial training) is primarily used to improve robustness against adversarial examples, not to mitigate bias; C (L1 regularization) induces sparsity in weights for feature selection and generalization, not fairness; and E (k-fold cross-validation) is a model-evaluation/resampling method for estimating generalization performance, not a bias-mitigation method.

Answer

Using adversarial training

Answer

Applying L1 regularization

Answer

Performing k-fold cross-validation

Question 4

A company is choosing between fine-tuning and RAG for a legal document assistant. Which TWO factors would MOST strongly favor RAG over fine-tuning?

Accepted Answer

The legal documents are updated frequently (weekly). Option A is correct because RAG retrieves from an external knowledge base at query time, so weekly-updated legal documents can be re-indexed without retraining the model, whereas fine-tuning would require repeated, costly retraining to keep pace. Option D is correct because RAG naturally returns the retrieved passages that grounded the answer, enabling precise source citations, while a fine-tuned model's parametric knowledge offers no traceable provenance. Option B does not favor RAG, since understanding complex legal terminology is largely a function of the base model's pretraining and can be addressed by either approach. Option C does not favor RAG either, as deep multi-document reasoning is a model capability rather than a retrieval benefit, and RAG alone does not guarantee it. Option E is not a strong RAG advantage, because building and operating a vector store and retriever also incurs infrastructure and compute costs, so a limited training budget does not decisively favor RAG.

Answer

The model needs to understand complex legal terminology

Answer

The queries require deep reasoning across multiple documents

Answer

The company has limited compute budget for training

Question 5

A company deploys a deep learning model for real-time object detection in autonomous vehicles. The model was trained on high-end GPUs but needs to run on edge devices with limited computational resources. Which technique is most effective for reducing model size and inference latency while maintaining acceptable accuracy?

Accepted Answer

Quantization. Quantization reduces the precision of the model's weights and activations (e.g., from 32-bit floating point to 8-bit integers), which significantly decreases model size and speeds up inference on edge devices with limited computational resources. This technique directly addresses the constraints of edge deployment while often maintaining acceptable accuracy through careful calibration.

Answer

Hyperparameter tuning

Answer

Batch normalization

Answer

Dropout

Question 6

Which TWO techniques are commonly used for feature selection in machine learning? (Choose 2)

Accepted Answer

L1 regularization (Lasso). L1 regularization (Lasso) is correct because it adds a penalty equal to the absolute value of the coefficients to the loss function, which drives less important feature coefficients exactly to zero, effectively performing embedded feature selection. Recursive Feature Elimination (RFE) is correct because it is a wrapper method that repeatedly trains a model, ranks features by importance (e.g., coefficients or feature_importances_), removes the weakest feature(s), and recurses until the desired number of features remains. PCA is not a feature-selection technique but a dimensionality-reduction method that creates new uncorrelated components from the original features, so it does not select among them. SMOTE is a data-level technique for handling class imbalance by synthesizing minority-class samples, not for selecting features. Dropout is a neural-network regularization method that randomly deactivates neurons during training to reduce overfitting, and it does not perform feature selection.

Answer

Principal Component Analysis (PCA)

Answer

SMOTE

Answer

Dropout

Question 7

A data scientist is monitoring a deployed image classification model. Which TWO actions are best practices for detecting model drift? (Choose 2.)

Accepted Answer

Use a holdout test set to periodically evaluate model accuracy.. Option C is correct because periodically scoring the model against a fixed holdout test set gives a direct, repeatable measurement of accuracy degradation, which is the core signal of model drift. Option E is correct because tracking the distribution of input features over time (e.g., via histograms or statistical tests like KL divergence or PSI) detects data drift, which typically precedes and causes model drift. Option A is not a detection practice but a remediation action, and retraining blindly on a schedule does not identify drift. Option B is a modeling change that may affect generalization but does nothing to detect drift in a deployed model. Option D is weaker than C and E because average prediction confidence can shift for reasons unrelated to drift and is not a reliable standalone drift indicator.

Answer

Schedule automatic weekly retraining of the model.

Answer

Increase the model's complexity to improve generalization.

Answer

Monitor the average prediction confidence of the model.

Question 8

Which TWO of the following are appropriate uses of unsupervised learning?

Accepted Answer

Detecting unusual patterns in network traffic that may indicate a cyberattack. Option C is correct because anomaly detection in network traffic is a classic unsupervised learning task: the model is trained on normal traffic without labels and flags deviations that may indicate a cyberattack, using techniques like clustering or autoencoders. Option E is correct because customer segmentation is typically performed with unsupervised clustering algorithms such as k-means or hierarchical clustering, which group customers by purchasing behavior without predefined labels. Option A is incorrect because spam classification is supervised learning, requiring labeled examples of spam and non-spam emails. Option B is incorrect because predicting a house's sale price is supervised regression with labeled target values. Option D is incorrect because identifying a person from a photo is supervised classification (or a related supervised recognition task) trained on labeled images of known individuals.

Answer

Classifying emails as spam or not spam

Answer

Predicting the sale price of a house given its features

Answer

Identifying a person from a photo

Question 9

Which THREE are common pitfalls when operationalizing AI models? (Select THREE.)

Accepted Answer

Training-serving skew due to differences in data preprocessing. Option A is correct because training-serving skew occurs when the preprocessing, feature engineering, or transformations applied during training differ from those applied at inference time, causing the model to receive inputs that do not match its learned distribution and degrading predictions. Option C is correct because deployed models degrade over time due to data drift, concept drift, and changing user behavior, so without monitoring for performance drift (e.g., tracking accuracy, latency, and input distributions) failures go undetected. Option D is correct because operationalizing AI requires serving infrastructure that can handle production traffic, scaling, and latency requirements; ignoring scalability leads to outages or unacceptable response times under load. Option B is not a pitfall but often a deliberate, sound engineering choice, since simpler models are easier to debug, maintain, and explain. Option E is not a pitfall either; automating retraining is a recommended MLOps practice that helps keep models current, provided it is paired with validation and monitoring.

Answer

Using simpler models that are easier to debug

Answer

Automating the model retraining process

Question 10

Which THREE are common causes of data leakage in machine learning pipelines?

Accepted Answer

Using future information to predict the present. Option B is correct because using future information to predict the present is the classic definition of temporal leakage: when training features contain values that would not have been available at prediction time (e.g., tomorrow's price used to predict today's), the model learns relationships that cannot generalize. Option D is correct because applying normalization (e.g., StandardScaler or MinMaxScaler) before splitting lets the scaler compute statistics such as the mean, variance, or min/max over the test data, so test-set information leaks into the training transformation; the scaler must be fit only on the training split. Option E is correct because features directly derived from the target variable (e.g., a 'remaining balance' column computed from the label, or target-encoded aggregates that include the current row's label) encode the answer into the inputs, producing artificially high validation scores that collapse in production. Option A is not a leakage cause but a mitigation: time-based splitting is the recommended approach for sequential data precisely to prevent temporal leakage. Option C is also not inherently leakage: cross-validation on the entire dataset is standard practice as long as the preprocessing is fit within each fold; leakage arises only if transformations are fit on the full dataset before cross-validation.

Answer

Using time-based splitting for sequential data

Answer

Using cross-validation on the entire dataset

Question 11

A machine learning engineer is training a logistic regression model and notices that the loss is decreasing very slowly. The learning rate is set to 0.001. What is the MOST likely cause and appropriate fix?

Accepted Answer

The learning rate is too low; increase it to 0.01. A learning rate of 0.001 is very low for many logistic regression implementations, causing the gradient descent algorithm to take extremely small steps toward the minimum of the loss function. This results in a slow decrease in loss because each weight update is minimal. Increasing the learning rate to 0.01 allows larger steps per iteration, accelerating convergence without typically causing divergence in well-scaled data.

Answer

The learning rate is too high; decrease it to 0.0001

Answer

The model is overfitting; add L2 regularisation

Answer

The batch size is too large; reduce it

Question 12

An organization uses a fine-tuned LLM for generating financial reports. An attacker gains access to the model's API and sends a series of queries that gradually reconstruct the training data of the fine-tuned model. This is an example of which attack?

Accepted Answer

Model inversion. Model inversion is an attack where an adversary queries a model repeatedly to reconstruct the training data or infer sensitive features of the training set. In this scenario, the attacker uses API access to gradually reconstruct the fine-tuned model's training data, which is the defining characteristic of model inversion. The gradual querying pattern is typical of inversion attacks that exploit the model's memorization of training examples.

Answer

Membership inference

Answer

Data poisoning

Answer

Model extraction

Question 13

Refer to the exhibit. A data scientist observes the training output. Which issue is most likely?

Accepted Answer

Overfitting. The exhibit shows training loss decreasing while validation loss increases after a certain epoch, which is the classic signature of overfitting. The model is memorizing the training data rather than learning generalizable patterns, leading to poor performance on unseen data.

Answer

Underfitting

Answer

Data augmentation failure

Answer

Model compression

Question 14

A team is deploying a deep learning model that uses a convolutional neural network (CNN) for image recognition. The model achieves high accuracy but is very slow to infer on edge devices. Which THREE optimization techniques should the team consider to speed up inference without significant accuracy loss? (Select three.)

Accepted Answer

Use weight pruning to remove unnecessary connections in the network.. Weight pruning removes redundant or less important connections (weights) from the neural network, reducing the number of computations required during inference. This directly speeds up inference on edge devices while typically causing only a minor drop in accuracy if done carefully, making it a standard optimization technique for deploying CNNs on resource-constrained hardware.

Answer

Use larger convolutional filters (e.g., 7x7 instead of 3x3) to capture more context.

Answer

Increase the number of convolutional layers to improve feature extraction.

Question 15

Which TWO of the following are effective defenses against adversarial examples in AI systems?

Accepted Answer

Train the model with adversarial examples (adversarial training). Adversarial training (option A) is a correct defense because it augments the training set with adversarial examples generated by attacks such as FGSM or PGD, so the model learns to classify perturbed inputs correctly and its decision boundary becomes more robust. Input sanitization and feature squeezing (option D) are also correct because they preprocess inputs to remove or reduce the adversarial perturbation — for example, by quantizing pixel values, spatial smoothing, or reducing color bit depth — which shrinks the attacker's effective search space and can neutralize small perturbations. Option B is not marked correct because model ensembles with majority voting can sometimes improve robustness but are not a reliable standalone defense; attackers can craft transferable or ensemble-aware adversarial examples that fool all members. Option C is wrong because increasing sensitivity to input changes is the opposite of what is wanted — it makes the model easier to fool with tiny perturbations. Option E is wrong because pruning reduces model complexity for efficiency and may slightly alter robustness, but it is not an established defense against adversarial examples and can even degrade robustness.

Answer

Use an ensemble of models and majority voting

Answer

Increase the model's sensitivity to input changes

Answer

Reduce model complexity through pruning