Courseiva

CompTIA AI+ AI0-001 (AI0-001) — Questions 601–675

1000 questions total · 14pages · All types, answers revealed

Take a mock examExam hub
Page 8

Page 9 of 14

Page 10
601
MCQmedium

A company is building an AI-powered document intelligence system to extract key fields from scanned invoices. The data contains 95% of invoices from one vendor and 5% from others. During model training, the F1 score is 0.95 on the overall test set, but the performance on the minority vendor invoices is very poor. What is the MOST likely cause?

A.The model is overfitting on the minority class
B.The dataset is imbalanced, and the model is biased toward the majority class
C.The data has a train/test leakage problem
D.The feature extraction is incorrect for the minority vendor invoices
AnswerB

Imbalanced data causes the model to optimize for the majority class, ignoring minority classes.

Why this answer

The imbalanced dataset causes the model to learn mostly from the majority class, leading to poor performance on the minority class. The other options are less likely given the high overall F1 score.

Full explanation →
602
MCQhard

A model trained on a dataset with imbalanced classes achieves 98% accuracy but only 50% recall for the minority class. Which technique should be applied first to address the imbalance?

A.Apply cost-sensitive learning
B.Reduce the majority class size
C.Use SMOTE to generate synthetic samples
D.Collect more data for the minority class
AnswerA

Cost-sensitive learning adjusts class weights in the loss function, directly tackling imbalance without data modification.

Why this answer

Cost-sensitive learning directly modifies the model's loss function to penalize misclassifications of the minority class more heavily than those of the majority class. This approach addresses the root cause of the imbalance—the model's bias toward the majority class—without altering the dataset distribution, making it the most immediate and effective first step.

Exam trap

CompTIA often tests the misconception that data-level techniques like SMOTE or undersampling should always be the first approach, when in fact cost-sensitive learning is a simpler, less invasive, and often more effective initial step that directly adjusts the model's learning objective.

How to eliminate wrong answers

Option B is wrong because reducing the majority class size (random undersampling) discards potentially valuable data, which can lead to loss of information and increased variance in the model, and it is not typically the first technique applied. Option C is wrong because SMOTE generates synthetic samples for the minority class, which can introduce noise and is a data-level augmentation technique that should be considered after cost-sensitive adjustments or as a complementary method, not as the first step. Option D is wrong because collecting more data for the minority class is often impractical, time-consuming, and may not be feasible in real-world scenarios; it is not a guaranteed or immediate solution to the imbalance.

Full explanation →
603
MCQeasy

A healthcare AI system uses patient data to predict disease risk. To comply with HIPAA, which privacy technique should be applied to the training data?

A.Differential privacy
B.Anonymisation
C.De-identification (Safe Harbor method)
D.Pseudonymisation
AnswerC

HIPAA's Safe Harbor method removes 18 identifiers, meeting the de-identification standard.

Why this answer

The Safe Harbor method of de-identification, as defined by the HIPAA Privacy Rule, removes 18 specific identifiers (e.g., names, dates, geographic subdivisions smaller than a state) from the training data. This ensures the data no longer qualifies as Protected Health Information (PHI), allowing it to be used for AI model training without violating HIPAA. Unlike other techniques, Safe Harbor provides a clear, legally safe harbor from privacy breach notification requirements when properly applied.

Exam trap

Cisco often tests the distinction between de-identification (Safe Harbor) and pseudonymisation, trapping candidates who think pseudonymisation alone satisfies HIPAA because it obscures direct identifiers, when in fact pseudonymised data remains PHI and requires the same privacy protections.

How to eliminate wrong answers

Option A is wrong because differential privacy adds mathematical noise to query outputs to protect individual records, but it does not remove all 18 HIPAA-required identifiers from the training data itself, and applying it alone would not satisfy the Safe Harbor standard for de-identification. Option B is wrong because anonymisation is a broader, often irreversible process that may go beyond HIPAA's Safe Harbor requirements, but the question specifically asks for a technique to comply with HIPAA, and the Safe Harbor method is the explicit, codified standard under HIPAA, not generic anonymisation. Option D is wrong because pseudonymisation replaces identifiers with pseudonyms but retains the ability to re-identify individuals via a key, which means the data is still considered PHI under HIPAA and does not meet the de-identification standard required for compliance.

Full explanation →
604
Multi-Selectmedium

Which THREE of the following are types of machine learning paradigms? (Choose three.)

Select 3 answers
A.Gradient boosting
B.Reinforcement learning
C.Unsupervised learning
D.Quantum computing
E.Supervised learning
AnswersB, C, E

Reinforcement learning involves an agent learning from rewards.

Why this answer

Reinforcement learning is a correct machine learning paradigm where an agent learns to make decisions by interacting with an environment, receiving rewards or penalties based on its actions. This trial-and-error approach is distinct from supervised and unsupervised learning, as it focuses on maximizing cumulative reward through exploration and exploitation.

Exam trap

CompTIA often tests candidates by listing specific algorithms (like gradient boosting) or adjacent technologies (like quantum computing) as distractors, hoping you confuse a technique or enabling technology with a fundamental learning paradigm.

Full explanation →
605
MCQmedium

A company uses AI to generate marketing images. They want to ensure that the images are clearly identified as AI-generated to comply with transparency obligations. Which approach is most effective?

A.Add a disclaimer in the platform's terms of service
B.Include metadata in the image file indicating it is AI-generated
C.Embed a visible watermark stating 'AI-generated' in each image
D.Rely on deepfake detection algorithms to flag the images
AnswerC

Visible watermarks are easily noticed and cannot be removed without degrading the image, ensuring clear disclosure.

Why this answer

Watermarking AI-generated content with visible or invisible markers is a direct way to disclose AI origin. Using metadata alone can be stripped; disclaimers in terms of service are not visible to users; relying on human detection is unreliable.

Full explanation →
606
MCQmedium

An AI model is being developed for medical diagnosis from X-ray images. The dataset contains only frontal chest X-rays. The model achieves high accuracy on test set but fails on lateral views. What is the most likely cause?

A.Dataset bias
B.Underfitting
C.Label noise
D.Overfitting
AnswerA

The training set lacks lateral views, causing bias; the model has not learned to recognize features specific to lateral X-rays.

Why this answer

The model was trained exclusively on frontal chest X-rays, so it never learned features specific to lateral views. When tested on lateral views, the distribution shift causes poor performance, which is a classic case of dataset bias (sampling bias). The high accuracy on the test set is misleading because the test set also only contained frontal views, masking the model's inability to generalize to other X-ray orientations.

Exam trap

Cisco often tests the distinction between overfitting and dataset bias by presenting a scenario where the model performs well on the test set (which shares the same bias) but fails on a different data distribution, leading candidates to mistakenly choose overfitting instead of recognizing the sampling bias.

How to eliminate wrong answers

Option B (Underfitting) is wrong because underfitting would cause poor performance on both the training and test sets, not just on a different distribution of data. Option C (Label noise) is wrong because label noise refers to incorrect ground-truth labels in the training data, which would degrade performance across all data types, not selectively on lateral views. Option D (Overfitting) is wrong because overfitting would manifest as high training accuracy but low test accuracy on the same distribution (e.g., frontal views), not as a failure on an entirely unseen data orientation.

Full explanation →
607
Multi-Selectmedium

A machine learning engineer wants to track hyperparameter experiments and compare results across runs. Which TWO tools are best suited for this purpose? (Choose 2)

Select 2 answers
A.MLflow
B.Weights & Biases
C.Apache Airflow
D.Docker
E.Kubeflow
AnswersA, B

MLflow provides experiment tracking, logging, and comparison UI.

Why this answer

MLflow is correct because it provides a centralized tracking server and API to log hyperparameters, metrics, and artifacts for each run, enabling easy comparison across experiments. Weights & Biases is correct because it offers a cloud-hosted dashboard with real-time logging, hyperparameter sweeps, and collaborative comparison features, making it ideal for tracking and comparing runs.

Exam trap

Cisco often tests the distinction between infrastructure tools (orchestration, containerization) and purpose-built experiment tracking tools; the trap here is that candidates may confuse Kubeflow’s pipeline capabilities with dedicated experiment tracking, or assume Docker/Airflow can serve as tracking solutions because they are used in ML workflows.

Full explanation →
608
MCQeasy

In the AI lifecycle, which phase involves splitting data into training, validation, and test sets?

A.Model training
B.Data preprocessing
C.Data collection
D.Model evaluation
AnswerB

Correct; preprocessing includes cleaning, transforming, and splitting data.

Why this answer

Data preprocessing is the phase where raw data is cleaned, transformed, and prepared for modeling. Splitting the dataset into training, validation, and test sets is a critical step during this phase to ensure unbiased evaluation and prevent data leakage. This split occurs before any model training begins, making it part of preprocessing rather than training or evaluation.

Exam trap

CompTIA often tests the misconception that data splitting belongs to model training or evaluation, when in fact it is a preprocessing step that must occur before any model sees the data.

How to eliminate wrong answers

Option A is wrong because model training is the phase where the algorithm learns patterns from the training data, not where the data is split; splitting must happen beforehand to avoid contaminating the evaluation. Option C is wrong because data collection is the initial gathering of raw data from sources, which occurs before any splitting or preprocessing. Option D is wrong because model evaluation uses the already-split test set to assess performance, but the split itself is established during data preprocessing.

Full explanation →
609
MCQeasy

A model's training accuracy is 99% but validation accuracy drops to 60%. What is the most likely issue?

A.Data leakage
B.Overfitting
C.Multicollinearity
D.Underfitting
AnswerB

Overfitting leads to high training accuracy but low validation accuracy.

Why this answer

A training accuracy of 99% with a validation accuracy of only 60% is a classic symptom of overfitting. The model has memorized the training data, including noise and outliers, rather than learning generalizable patterns, causing it to perform poorly on unseen validation data.

Exam trap

CompTIA often tests the distinction between overfitting and data leakage by presenting a large accuracy gap, where candidates might mistakenly attribute the issue to data leakage instead of recognizing that leakage typically inflates both accuracies rather than creating a divergence.

How to eliminate wrong answers

Option A is wrong because data leakage typically causes both training and validation accuracy to be artificially high, not a large gap between them; it occurs when information from outside the training set inadvertently influences the model. Option C is wrong because multicollinearity refers to high correlation among input features in regression models, which affects coefficient stability and interpretability, not a drastic accuracy drop between training and validation sets. Option D is wrong because underfitting would result in low accuracy on both training and validation sets (e.g., both below 70%), not a high training accuracy with a low validation accuracy.

Full explanation →
610
MCQhard

A retail company uses a time-series model to forecast daily sales for inventory management. The model is a seasonal ARIMA trained on three years of daily data. It performed well during initial validation but after deployment, forecasts became inaccurate during holiday seasons, often underestimating demand by up to 40%. The data science team examined the features and found that the training data did not include any holiday indicators. They also discovered that the model's residuals show strong autocorrelation during holiday weeks. The company needs to improve the forecast for the upcoming holiday season. They have access to historical sales data with holiday dates and are considering several approaches. Which approach will BEST address the issue?

A.Engineer holiday-related features (e.g., holiday flag, days before/after) and retrain the ARIMA model with these features as exogenous variables.
B.Decrease the learning rate and increase the number of training epochs.
C.Create an ensemble of the ARIMA model and a seasonal naive model to average forecasts.
D.Replace the ARIMA model with a deep learning LSTM model trained on the same data.
AnswerA

Incorporating holiday information directly addresses the root cause of the forecast error.

Why this answer

The core issue is that the ARIMA model lacks holiday-related information, causing systematic underestimation during holiday periods. By engineering holiday features (e.g., binary flags, days-before/after indicators) and including them as exogenous variables in a SARIMAX model, the model can directly learn the demand spikes associated with holidays. This directly addresses the residual autocorrelation during holiday weeks and improves forecast accuracy without changing the underlying time-series framework.

Exam trap

CompTIA often tests the misconception that adding more data or switching to a more complex model (like LSTM) automatically fixes forecasting issues, when the real problem is missing relevant features that can be addressed with a simpler, interpretable approach like SARIMAX with exogenous variables.

How to eliminate wrong answers

Option B is wrong because decreasing the learning rate and increasing epochs is a hyperparameter tuning technique for gradient-based models (e.g., neural networks), not applicable to ARIMA which is trained via maximum likelihood estimation or least squares. Option C is wrong because averaging forecasts from an ARIMA and a seasonal naive model would dilute the ARIMA's signal and still fail to capture holiday effects, as the naive model also lacks holiday features. Option D is wrong because replacing ARIMA with an LSTM does not automatically solve the missing holiday indicator problem; the LSTM would also need engineered holiday features to learn the pattern, and the question asks for the best approach given available data, not a complete model swap.

Full explanation →
611
MCQmedium

A data engineer is reviewing an S3 bucket policy for a machine learning project. The policy is intended to allow access to training data only from the corporate network (10.0.0.0/16). However, users in the corporate network report access denied. Which issue is most likely causing the problem?

A.The policy is missing a Deny statement.
B.The resource ARN is incorrect.
C.The policy does not include s3:ListBucket action, which is needed to list objects.
D.The IP address condition is not matching because the corporate network uses a different CIDR.
AnswerC

Users need ListBucket to see objects before getting them.

Why this answer

Option C is correct. The condition key uses "aws:SourceIp" but the correct key for IP address condition is "aws:SourceIp". However, the exhibit shows "aws:SourceIp" which is correct? Wait, the exhibit shows "aws:SourceIp" — that is correct.

But the problem might be that the policy uses "GetObject" but the users might be trying to list objects (s3:ListBucket). Option C says the action does not include s3:ListBucket, which is often needed first. Option A is wrong because the resource is correct.

Option B is wrong because the policy allows access from 10.0.0.0/16. Option D is wrong because the effect is Allow.

Full explanation →
612
MCQmedium

A government agency is deploying an AI model to screen loan applications. The model uses features like income, credit score, employment history, and zip code. During fairness auditing, the model is found to deny a disproportionately high number of applicants from a particular demographic group, even when controlling for legitimate financial factors. The agency wants to mitigate this bias without significantly reducing overall accuracy. Which approach should the data scientist prioritize?

A.Adjust the decision threshold for the affected group
B.Remove the zip code feature from the model
C.Apply sample weighting to balance the demographic groups
D.Use adversarial debiasing during model training
AnswerD

Adversarial debiasing forces the model to learn representations that are invariant to sensitive attributes, reducing bias with minimal accuracy loss.

Why this answer

Adversarial debiasing is the correct approach because it directly optimizes the model to remove sensitive information (e.g., demographic group membership) from its internal representations while preserving predictive accuracy. This technique trains a primary model to predict the target (loan approval) and an adversary to predict the protected attribute from the model's learned features, forcing the primary model to learn representations that are both accurate and unbiased. It addresses the root cause of bias—correlation between protected attributes and model predictions—without requiring post-hoc threshold adjustments or sacrificing overall performance.

Exam trap

Cisco often tests the misconception that simply removing a sensitive feature (like zip code) or reweighting data is sufficient to eliminate bias, when in reality bias can be encoded through correlated proxies and requires algorithmic debiasing during training.

How to eliminate wrong answers

Option A is wrong because adjusting the decision threshold for the affected group is a post-hoc fairness intervention that can reduce accuracy for that group and may violate regulatory requirements for equal treatment; it does not address the underlying model bias. Option B is wrong because removing the zip code feature alone is insufficient—bias can still propagate through correlated features like income or employment history, and this approach may reduce model accuracy without guaranteeing fairness. Option C is wrong because sample weighting can help balance representation but does not prevent the model from learning biased correlations; it may also distort the training distribution and degrade accuracy on the majority group.

Full explanation →
613
MCQeasy

Based on the exhibit, what does this indicate about the model?

A.The model has balanced performance
B.The model is underfitting
C.The model is overfitting
D.The model has high precision but low recall, missing many positives
AnswerD

Correct: Precision is high, recall is low, indicating the model is conservative in labeling positives.

Why this answer

The exhibit shows a confusion matrix or performance metrics where the model correctly identifies some positives but has a high number of false negatives, indicating low recall. This means the model misses many actual positive cases, which aligns with high precision (few false positives) but low recall. Option D correctly describes this imbalance.

Exam trap

Cisco often tests the distinction between precision and recall by presenting a confusion matrix where candidates confuse high precision with overall good performance, ignoring the low recall that indicates many missed positives.

How to eliminate wrong answers

Option A is wrong because balanced performance would show similar precision and recall values, not a significant disparity where recall is low. Option B is wrong because underfitting would result in poor performance on both training and validation data, not specifically high precision with low recall. Option C is wrong because overfitting typically shows high training accuracy but poor generalization, not a specific pattern of high precision and low recall; overfitting often leads to high variance, not this specific precision-recall trade-off.

Full explanation →
614
MCQmedium

An AI chatbot is being developed to assist with customer support. The team is concerned about prompt injection attacks where malicious users try to override the system's instructions. Which defense is MOST effective against direct prompt injection?

A.Input validation and sanitization
B.Rate limiting API requests
C.Output filtering
D.Using a larger model
AnswerA

Sanitizing inputs removes or neutralizes injection attempts, directly mitigating prompt injection.

Why this answer

Input validation and sanitization can detect and block attempts to inject malicious instructions. While guardrails help, the primary defense is to filter and validate user inputs before they reach the model.

Full explanation →
615
MCQeasy

A data engineer needs to store training data in a format that supports columnar pruning during model training. Which storage format should they use?

A.Parquet
B.XML
C.JSON
D.CSV
AnswerA

Parquet is columnar, enabling compression and pruning, reducing I/O.

Why this answer

Parquet is the correct choice because it is a columnar storage format that enables column pruning, allowing the training process to read only the columns needed for model training rather than entire rows. This reduces I/O and speeds up data loading, which is critical for large-scale AI/ML workloads. Unlike row-oriented formats, Parquet stores data by columns, making it efficient for analytical queries and feature selection.

Exam trap

CompTIA often tests the misconception that JSON or CSV are acceptable for columnar pruning because they are common and human-readable, but the trap here is that only columnar formats like Parquet or ORC support efficient column-level access, while row-oriented formats require full record scans.

How to eliminate wrong answers

Option B (XML) is wrong because XML is a verbose, hierarchical text format that stores data row-wise and lacks columnar pruning capabilities, leading to high storage overhead and slow read performance for tabular data. Option C (JSON) is wrong because JSON is a row-oriented, self-describing format that requires parsing entire records even when only a subset of fields is needed, making it unsuitable for column pruning. Option D (CSV) is wrong because CSV is a flat, row-oriented text format that forces reading entire rows into memory, with no support for columnar storage or predicate pushdown, resulting in inefficient I/O for selective column access.

Full explanation →
616
MCQmedium

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?

A.Applying SMOTE to oversample minority classes
B.Using a more complex algorithm
C.Removing all demographic features
D.Ensuring the training data is representative of the target population
AnswerD

Representative data prevents bias from skewed sampling.

Why this answer

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.

Exam trap

Cisco often tests the misconception that bias can be fixed by technical tweaks like oversampling or removing sensitive attributes, when in fact the root cause is almost always unrepresentative training data that must be addressed at the collection or sampling stage.

How to eliminate wrong answers

Option A is wrong because SMOTE (Synthetic Minority Oversampling Technique) addresses class imbalance by generating synthetic samples for minority classes, but it does not correct for broader representational bias (e.g., missing demographic subgroups) and can introduce artifacts if the minority class itself is not representative of the target population. Option B is wrong because using a more complex algorithm does not mitigate bias; in fact, complex models can overfit to spurious correlations in biased data, making bias worse rather than reducing it. Option C is wrong because removing all demographic features can mask bias but does not eliminate it—protected attributes like race or age may be correlated with other features (e.g., zip code, medical history), leading to proxy discrimination, and this approach can also remove clinically relevant information needed for accurate risk prediction.

Full explanation →
617
MCQmedium

A company built a speech-to-text model using a recurrent neural network (RNN). During deployment, the model performs poorly on accented speech. Which action would most effectively improve model robustness?

A.Collect a small sample of accented speech and fine-tune the model on that sample only.
B.Add dropout and reduce the number of RNN layers to prevent overfitting to the current data.
C.Augment the training dataset with various accented audio samples and retrain the model.
D.Replace the RNN with a convolutional neural network (CNN) for feature extraction.
AnswerC

Data augmentation with accents directly addresses the performance gap.

Why this answer

Option C is correct because augmenting the training dataset with diverse accented audio samples directly addresses the root cause of poor performance—distribution shift between training and deployment data. Retraining the model on this enriched dataset allows the RNN to learn invariant features across accents, improving generalization without altering the model architecture or risking catastrophic forgetting from fine-tuning on a tiny sample.

Exam trap

CompTIA often tests the misconception that architectural changes (like switching to CNN or adding regularization) can fix data distribution mismatches, when the real solution is to address the missing data diversity through augmentation or retraining.

How to eliminate wrong answers

Option A is wrong because fine-tuning on a small sample of accented speech can cause catastrophic forgetting of the original training distribution and does not provide enough diversity to learn robust accent-invariant features. Option B is wrong because adding dropout and reducing layers addresses overfitting to the current data, but the core problem is underfitting to accented speech due to missing representative training examples, not overfitting. Option D is wrong because replacing the RNN with a CNN for feature extraction does not inherently solve the accent robustness issue; CNNs are effective for spatial patterns but less suited for sequential temporal dependencies in speech, and the fundamental problem remains the lack of accented training data.

Full explanation →
618
MCQmedium

A security analyst is investigating a potential adversarial attack on a production image classifier. The attack involves tiny perturbations that are invisible to the human eye but cause the model to misclassify a stop sign as a speed limit sign. Which type of attack is this?

A.Data poisoning
B.Model inversion
C.Membership inference
D.Adversarial example
AnswerD

Adversarial examples are inputs with imperceptible perturbations that cause misclassification.

Why this answer

This is an adversarial example attack, where imperceptible perturbations are added to the input (e.g., a stop sign) to cause the model to misclassify it (e.g., as a speed limit sign). The perturbations are crafted using gradient-based methods (like FGSM or PGD) to maximize the model's loss, exploiting its linearity in high-dimensional spaces. This differs from other attacks because it targets the inference phase, not the training data or model parameters.

Exam trap

Cisco often tests the distinction between attacks that occur during training (e.g., data poisoning) versus inference (e.g., adversarial examples), and candidates mistakenly choose data poisoning because they confuse 'adding noise to input' with 'corrupting training data'.

How to eliminate wrong answers

Option A is wrong because data poisoning involves corrupting the training dataset (e.g., injecting mislabeled samples) to compromise the model during training, not adding perturbations to a single input at inference time. Option B is wrong because model inversion attempts to reconstruct private training data from the model's outputs (e.g., generating a face from a facial recognition model), not to cause misclassification of a specific input. Option C is wrong because membership inference determines whether a particular data point was part of the training set by analyzing the model's confidence scores, not by altering an input to induce a misclassification.

Full explanation →
619
MCQmedium

Which AI governance framework is specifically designed by the U.S. National Institute of Standards and Technology (NIST) to help organizations manage AI risks?

A.ISO/IEC 27001
B.COBIT
C.NIST AI Risk Management Framework
D.GDPR
AnswerC

NIST AI RMF is the U.S. standard for AI risk management.

Why this answer

The NIST AI Risk Management Framework (AI RMF) is the specific governance framework developed by the U.S. National Institute of Standards and Technology to help organizations manage AI risks, including those related to trustworthiness, fairness, and robustness. It provides a structured approach for identifying, assessing, and mitigating risks throughout the AI lifecycle, aligning with NIST's role in setting standards for cybersecurity and risk management.

Exam trap

Cisco often tests candidates by listing well-known frameworks or regulations (like ISO/IEC 27001 or GDPR) that are related to security or privacy but are not AI-specific, leading candidates to confuse general governance with the NIST AI RMF.

How to eliminate wrong answers

Option A is wrong because ISO/IEC 27001 is an international standard for information security management systems (ISMS), not an AI-specific risk management framework, and it focuses on general data security rather than AI risks. Option B is wrong because COBIT (Control Objectives for Information and Related Technologies) is a framework for IT governance and management, developed by ISACA, and does not address AI-specific risks or the NIST AI RMF. Option D is wrong because GDPR (General Data Protection Regulation) is a European Union regulation for data privacy and protection, not a risk management framework, and it is not designed by NIST.

Full explanation →
620
Multi-Selecthard

A company is deploying a generative AI system that produces text content. To comply with emerging transparency obligations, which THREE measures should they implement?

Select 3 answers
A.Watermark AI-generated content
B.Disclose AI involvement to users
C.Encrypt all training data
D.Provide deepfake detection tools
E.Limit model access to internal employees only
AnswersA, B, D

Watermarking helps identify AI-generated content and is a transparency best practice.

Why this answer

Watermarking AI-generated content (Option A) is correct because it embeds an imperceptible, machine-detectable signal into the output, enabling provenance verification. This directly addresses transparency obligations by allowing downstream systems to identify synthetic text, which is a key requirement in emerging AI regulations such as the EU AI Act.

Exam trap

Cisco often tests the distinction between security controls (encryption, access control) and governance/transparency measures, so candidates mistakenly select encryption or access restriction as fulfilling transparency obligations when they do not.

Full explanation →
621
Multi-Selectmedium

A company is building a RAG-based Q&A system for a large collection of technical manuals. They need to choose an embedding model and a similarity search method. Which TWO choices are most appropriate for this scenario? (Select TWO)

Select 2 answers
A.Use a general-purpose embedding model like text-embedding-ada-002
B.Use dot product as the similarity metric for non-normalized embeddings
C.Use Euclidean distance as the similarity metric for vector search
D.Use a domain-specific embedding model fine-tuned on technical documentation
E.Use cosine similarity as the similarity metric for vector search
AnswersD, E

A domain-specific model captures the nuances of technical language, improving retrieval precision for the Q&A system.

Why this answer

Cosine similarity is the standard metric for comparing embeddings (normalized vectors). Using a domain-specific embedding model (e.g., fine-tuned on technical text) yields better retrieval accuracy. Dot product can be used but is less common; Euclidean distance is not ideal for high-dimensional embeddings; a general-purpose model may perform poorly on domain-specific language.

Full explanation →
622
Multi-Selecthard

Which TWO techniques are specifically designed to protect individual privacy when training AI models?

Select 2 answers
A.Dropout
B.Model pruning
C.Differential privacy
D.Regularization
E.Anonymization
AnswersC, E

Differential privacy adds noise to training to prevent data leakage.

Why this answer

Differential privacy (C) is a technique that adds calibrated noise to training data or model updates, ensuring that the output of the model does not reveal whether any specific individual's data was included. This provides a formal mathematical guarantee of privacy, quantified by the epsilon parameter, making it a direct privacy-preserving method for AI training.

Exam trap

Cisco often tests the distinction between techniques that improve model performance (dropout, pruning, regularization) and those that explicitly safeguard privacy, leading candidates to confuse regularization with privacy protection.

Full explanation →
623
MCQmedium

A company wants to create an AI system that can identify objects in images. They have a large dataset of labeled images. Which type of neural network architecture is most suitable?

A.Transformer
B.Convolutional neural network (CNN)
C.Generative adversarial network (GAN)
D.Recurrent neural network (RNN)
AnswerB

Correct; CNNs excel at image recognition due to convolutional layers.

Why this answer

Convolutional neural networks (CNNs) are specifically designed to process grid-like data such as images. They use convolutional layers to automatically learn spatial hierarchies of features (edges, textures, objects) from pixel data, making them the most suitable architecture for image classification tasks with labeled datasets.

Exam trap

CompTIA often tests the misconception that any 'neural network' can handle images equally, but the trap is that RNNs and Transformers are sequence-based and not optimized for spatial feature extraction, while GANs are generative, not discriminative.

How to eliminate wrong answers

Option A is wrong because Transformers are primarily designed for sequential data (e.g., text) using self-attention mechanisms; while they can be adapted for vision (Vision Transformers), they require large datasets and are not the standard choice for traditional image classification. Option C is wrong because Generative adversarial networks (GANs) are used for generating new data (e.g., synthetic images) rather than classifying or identifying objects in existing images. Option D is wrong because Recurrent neural networks (RNNs) are designed for sequential or time-series data (e.g., text, speech) and struggle with spatial relationships in images due to vanishing gradients and lack of translation invariance.

Full explanation →
624
MCQmedium

A team is using a pre-trained BERT model for a sentiment analysis task on product reviews. They want to adapt it to their specific domain with limited labeled data. Which approach is MOST effective?

A.Use BERT as a feature extractor and train a logistic regression on top
B.Apply data augmentation to increase the dataset and then train from scratch
C.Train a new BERT model from scratch on the domain data
D.Fine-tune the pre-trained BERT model on the small labeled dataset
AnswerD

Fine-tuning leverages pre-trained knowledge and adapts effectively with limited data.

Why this answer

Fine-tuning the pre-trained model on the small labeled dataset is standard for transfer learning with BERT. Training from scratch would require massive data; feature extraction with a linear classifier is possible but less effective with limited data; data augmentation alone does not adapt the model.

Full explanation →
625
MCQmedium

A data scientist is training a binary classifier and observes that the training accuracy is 99% but the test accuracy is only 70%. Which of the following is the MOST likely cause?

A.The model is underfitting the training data
B.The learning rate is too high
C.The model is overfitting the training data
D.The test set contains data leakage from the training set
AnswerC

High training accuracy with much lower test accuracy is classic overfitting.

Why this answer

A large gap between training and test accuracy indicates overfitting: the model memorized training data but fails to generalize.

Full explanation →
626
MCQhard

A retail company uses a machine learning model to predict daily sales. The model takes features like past sales, promotions, holidays, and weather data. Recently, the model's accuracy dropped significantly. The data engineer checks the data pipeline and finds that the weather data source changed from a free API to a new paid API that provides more detailed data. The new data includes additional attributes like humidity and wind speed, but the existing pipeline only ingests temperature and precipitation. Also, the time zone format changed from UTC to local time. The model was trained on the old format. Which action should the engineer take first to restore model performance?

A.Add a new step to merge old and new weather data before feeding to the model.
B.Transform the new data to match the old format (time zone and selected features) and retrain the model.
C.Revert to the old weather API.
D.Retrain the model with the new data including all new features.
AnswerB

This aligns the data with the training pipeline, resolving the immediate mismatch.

Why this answer

The immediate problem is the time zone change causing misalignment between training and inference data. Transforming the new data to match the old format ensures consistency. Retraining with all new features may introduce drift; reverting may not be possible; merging data without alignment causes inconsistency.

Full explanation →
627
Multi-Selectmedium

A data scientist is preparing to train a convolutional neural network (CNN) for image classification. Which TWO actions are most effective for preventing overfitting? (Choose 2)

Select 2 answers
A.Use data augmentation
B.Increase the number of epochs
C.Use L2 regularization
D.Add more convolutional layers
E.Use dropout layers
AnswersA, E

Data augmentation generates variations of training images, effectively increasing the dataset size and reducing overfitting.

Why this answer

Data augmentation (A) is effective for preventing overfitting because it artificially expands the training dataset by applying random transformations (e.g., rotation, flipping, cropping, color jitter) to existing images. This exposes the CNN to a wider variety of input patterns, reducing the model's tendency to memorize noise or specific details and improving generalization to unseen data.

Exam trap

Cisco often tests the distinction between regularization techniques that directly reduce overfitting (like dropout and data augmentation) versus architectural changes (like adding layers) that increase capacity and may worsen overfitting if not balanced with regularization.

Full explanation →
628
Multi-Selecthard

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.)

Select 3 answers
A.Use larger convolutional filters (e.g., 7x7 instead of 3x3) to capture more context.
B.Use weight pruning to remove unnecessary connections in the network.
C.Implement knowledge distillation by training a smaller model to mimic the larger one.
D.Increase the number of convolutional layers to improve feature extraction.
E.Apply model quantization to reduce weight precision.
AnswersB, C, E

Pruning reduces computation and memory footprint.

Why this answer

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.

Exam trap

CompTIA often tests the misconception that increasing model capacity (larger filters or more layers) improves performance without considering the trade-off in inference speed, leading candidates to select options that actually worsen latency on edge devices.

Full explanation →
629
MCQmedium

A data scientist submits a model training job to a cloud ML platform. The job fails with an error: "Out of memory: Killed process." The training code is proven to work on the developer's local machine with 16GB RAM. The cloud instance has 32GB RAM. What is the most likely cause?

A.The cloud instance does not have enough memory
B.The training code is loading the entire dataset into memory without batch processing
C.The training model is too large for the GPU memory
D.The cloud instance ran out of disk space
AnswerB

If the dataset is larger than memory, the process may exhaust RAM even if instances have more memory than local.

Why this answer

Option B is correct because the error 'Out of memory: Killed process' occurs when the operating system's OOM killer terminates a process that exceeds available RAM. Even though the cloud instance has 32GB RAM, if the training code loads the entire dataset into memory without batch processing (e.g., using pandas.read_csv() without chunking), it can consume far more memory than the dataset size due to Python object overhead, temporary copies, and intermediate tensors. The local machine with 16GB RAM may have worked due to a smaller dataset or different memory pressure, but the cloud instance's 32GB is insufficient if the code is not memory-efficient.

Exam trap

CompTIA often tests the misconception that 'more RAM always solves memory errors', but the trap here is that the error is caused by inefficient data loading (no batching) rather than absolute memory size, so candidates must recognize the OOM killer message as a symptom of memory exhaustion from unbounded data consumption.

How to eliminate wrong answers

Option A is wrong because the cloud instance has 32GB RAM, which is double the local machine's 16GB, so the error is not simply due to insufficient memory—it's a code inefficiency. Option C is wrong because the error message 'Out of memory: Killed process' is a Linux kernel OOM killer message related to system RAM, not GPU memory; GPU out-of-memory errors typically appear as CUDA errors (e.g., 'CUDA out of memory'). Option D is wrong because disk space exhaustion would produce errors like 'No space left on device' or 'Disk quota exceeded', not an OOM killer process termination.

Full explanation →
630
Multi-Selectmedium

A team is designing an AI agent that needs to interact with external APIs, search the web, and perform multi-step reasoning. Which TWO architectural components are essential for this agentic workflow? (Choose TWO.)

Select 2 answers
A.Fine-tuning the base model
B.ReAct pattern (Reasoning + Acting)
C.Tool use / function calling
D.Single-turn response generation
E.Static prompt with no iterations
AnswersB, C

ReAct combines reasoning traces with actions, allowing the agent to plan and adapt.

Why this answer

Tool use allows the agent to call external APIs, and the ReAct pattern (Reasoning + Acting) enables iterative reasoning and action steps.

Full explanation →
631
MCQhard

A financial institution needs to integrate an AI-based credit scoring model into an existing mainframe system that processes transactions in COBOL. The model is deployed as a REST API. What is the best strategy to ensure minimal disruption and maintain data integrity?

A.Copy all transaction data to a cloud database for the model to access.
B.Use an API gateway with versioning and circuit breaker patterns.
C.Rewrite the mainframe system in Java to directly call the model.
D.Install a GPU on the mainframe to run the model natively.
AnswerB

This pattern enables loose coupling, resilience, and gradual integration without major changes to the mainframe.

Why this answer

Option B is correct because an API gateway with versioning and circuit breaker patterns allows the COBOL mainframe to call the REST API without modifying its core transaction logic. The gateway handles protocol translation, rate limiting, and failover, ensuring minimal disruption to the existing mainframe system while maintaining data integrity through transactional consistency and graceful degradation.

Exam trap

CompTIA often tests the misconception that legacy systems must be replaced or heavily modified to integrate with modern AI services, when in fact an API gateway provides a non-invasive integration layer that preserves the existing infrastructure.

How to eliminate wrong answers

Option A is wrong because copying all transaction data to a cloud database introduces latency, potential data inconsistency, and security risks, and does not address the integration between COBOL and the REST API. Option C is wrong because rewriting the mainframe system in Java is a massive, high-risk, and costly undertaking that would cause significant disruption and is unnecessary for integrating a REST API. Option D is wrong because installing a GPU on the mainframe does not enable native execution of the AI model, as mainframes lack the necessary software stack and the model is deployed as a REST API, not as a local executable.

Full explanation →
632
MCQhard

A model's predicted probabilities are well-calibrated overall but the model systematically assigns higher probabilities to one demographic group even when the actual outcome likelihood is the same. Which fairness issue is present?

A.Demographic parity violation
B.Equalised odds violation
C.Calibration disparity
D.Individual fairness violation
AnswerC

Calibration disparity occurs when predicted probabilities are not equally reliable across groups, as described.

Why this answer

Calibration fairness requires that for each predicted probability, the actual outcome rate is the same across groups. The scenario describes a calibration disparity. Demographic parity would involve different selection rates.

Equalised odds involves equal error rates. Individual fairness is about treating similar individuals similarly.

Full explanation →
633
Multi-Selectmedium

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

Select 2 answers
A.Train the model with adversarial examples (adversarial training)
B.Use an ensemble of models and majority voting
C.Increase the model's sensitivity to input changes
D.Implement input sanitization and feature squeezing
E.Reduce model complexity through pruning
AnswersA, D

Adversarial training hardens the model against perturbations.

Why this answer

Adversarial training (A) is effective because it exposes the model to perturbed inputs during training, forcing it to learn robust decision boundaries that are less sensitive to small, malicious perturbations. Input sanitization and feature squeezing (D) reduce the attack surface by compressing input features (e.g., reducing bit depth or spatial smoothing) to eliminate adversarial noise before inference, making it harder for an attacker to craft a successful perturbation.

Exam trap

Cisco often tests the misconception that ensemble methods or model simplification inherently improve adversarial robustness, when in fact they do not address the fundamental mechanism of adversarial perturbations and may even weaken defenses.

Full explanation →
634
Multi-Selecthard

Which TWO of the following are key characteristics of unsupervised learning?

Select 2 answers
A.It uses data without labeled responses
B.It predicts a target variable based on input features
C.It discovers hidden patterns or groupings in data
D.It requires a reward signal to learn optimal actions
E.It typically requires a separate validation set for tuning
AnswersA, C

Unsupervised learning works with unlabeled data.

Why this answer

Option A is correct because unsupervised learning algorithms, such as k-means clustering or hierarchical clustering, operate exclusively on input data that has no labeled responses. The model must infer the underlying structure directly from the features without any ground-truth outputs to guide it, which is the defining characteristic of unsupervised learning.

Exam trap

CompTIA often tests the distinction between supervised, unsupervised, and reinforcement learning by presenting a characteristic that is true for one paradigm but not the other, and the trap here is that candidates may confuse 'predicting a target variable' (supervised) with 'discovering hidden patterns' (unsupervised) because both involve analyzing input features.

Full explanation →
635
Multi-Selectmedium

Which TWO techniques should be considered when optimizing a deep learning model for deployment on edge devices with limited computational resources?

Select 2 answers
A.Apply adversarial training
B.Model quantization
C.Use a GPU for inference
D.Knowledge distillation
E.Increase the number of layers
AnswersB, D

Quantization reduces memory and computation requirements.

Why this answer

Model quantization reduces the precision of the model's weights and activations (e.g., from 32-bit floating point to 8-bit integer), which significantly decreases memory footprint and computational latency. This makes it a primary technique for deploying deep learning models on edge devices with limited resources.

Exam trap

CompTIA often tests the distinction between training-phase techniques (like adversarial training) and deployment-phase optimization techniques (like quantization and knowledge distillation), leading candidates to select options that improve model quality rather than reduce resource consumption.

Full explanation →
636
MCQeasy

A hospital's radiology department uses an AI model to detect lung nodules in CT scans. The model was trained on data from a specific brand of scanners and patient demographics common in Europe. Recently, the hospital acquired new scanners from a different manufacturer and started serving a more diverse patient population. Over the past month, the model's false-positive rate has increased by 15% and false-negative rate by 8%. The radiologists are losing confidence and are considering abandoning the AI tool altogether. The IT team has verified that the model inference is running correctly and the hardware is performing as expected. The data science team suspects the problem is related to the change in input data distribution. The hospital's AI operations policy requires that any model update must be validated on at least 500 recent cases before deployment. What is the BEST course of action for the AI operations team?

A.Roll back to the previous model version and restrict use of the AI tool to only European patients.
B.Collect 500 recent CT scans from the new scanners, retrain the model on a combined old and new dataset, and validate before deployment.
C.Retrain the model using the original training data but with increased regularization to avoid overfitting.
D.Adjust the model's decision threshold to reduce false positives and then monitor for two weeks.
AnswerB

Retraining with a representative sample addresses the data drift and meets the policy requirement of 500 validation cases.

Why this answer

Option C is correct because collecting 500 recent CT scans that represent the new scanner and patient distribution allows for validation of a retrained model that adapts to the new data. Option A is wrong because adjusting thresholds only addresses false-positive rate but not false-negative rate, and does not tackle the root cause. Option B is wrong because rolling back to the previous model where performance was good on old data would not improve performance on the new data.

Option D is wrong because retraining on the original training data would not include the new distribution and likely yield the same issues.

Full explanation →
637
MCQmedium

A machine learning engineer is tuning a neural network for image classification. The training loss decreases steadily, but the validation loss starts increasing after 50 epochs. Which action best addresses this issue?

A.Increase the number of hidden layers
B.Add more training data
C.Apply early stopping with a patience of 10 epochs
D.Increase the batch size
AnswerC

Early stopping monitors validation loss and stops training when it starts increasing, directly addressing overfitting.

Why this answer

The described behavior—decreasing training loss with increasing validation loss—is a classic sign of overfitting. Early stopping with a patience of 10 epochs directly addresses this by halting training when the validation loss fails to improve for a specified number of epochs, preventing further overfitting while retaining the best model weights.

Exam trap

Cisco often tests the distinction between underfitting and overfitting symptoms, and the trap here is that candidates may confuse a rising validation loss with a need for more data or a deeper network, when the correct action is to stop training early to combat overfitting.

How to eliminate wrong answers

Option A is wrong because increasing the number of hidden layers increases model capacity, which typically worsens overfitting by allowing the network to memorize training data more easily. Option B is wrong because adding more training data can help reduce overfitting in general, but it is not the most direct or immediate fix for the specific problem of validation loss increasing after 50 epochs; early stopping is a more targeted and efficient solution. Option D is wrong because increasing the batch size provides a more accurate gradient estimate but does not prevent overfitting; it may even lead to sharper minima and poorer generalization, making the validation loss issue worse.

Full explanation →
638
MCQeasy

A developer wants to deploy a scikit-learn model as a REST API endpoint with minimal infrastructure management. Which cloud service is MOST appropriate?

A.Use AWS Lambda with a custom runtime
B.Deploy on an EC2 instance manually
C.Use AWS SageMaker to create a real-time endpoint
D.Use Amazon ECS with manual Docker setup
AnswerC

SageMaker offers managed inference endpoints with automatic scaling, reducing operational overhead.

Why this answer

AWS SageMaker provides a fully managed service for deploying machine learning models as real-time endpoints with built-in scaling, monitoring, and automatic infrastructure management. It directly supports scikit-learn models via pre-built containers, eliminating the need for custom runtime setup or manual server configuration. This makes it the most appropriate choice for a developer seeking minimal infrastructure management.

Exam trap

Cisco often tests the misconception that serverless compute like AWS Lambda is the best choice for any API deployment, but the trap here is that Lambda's execution environment and constraints (timeout, payload size, cold starts) make it inappropriate for ML model inference, whereas SageMaker is purpose-built for this workload.

How to eliminate wrong answers

Option A is wrong because AWS Lambda with a custom runtime requires manual packaging of the scikit-learn model and dependencies, and Lambda has a 15-minute timeout and limited memory, making it unsuitable for real-time inference with larger models or payloads. Option B is wrong because deploying on an EC2 instance manually involves provisioning, patching, scaling, and managing the underlying server, which contradicts the requirement for minimal infrastructure management. Option D is wrong because Amazon ECS with manual Docker setup still requires managing the cluster, task definitions, and scaling policies, adding operational overhead compared to SageMaker's fully managed endpoint service.

Full explanation →
639
Multi-Selecthard

Which THREE data quality dimensions are critical for ensuring model reliability?

Select 3 answers
A.Timeliness.
B.Consistency.
C.Completeness.
D.Volume.
E.Accuracy.
AnswersA, B, E

Outdated data can cause predictions to be irrelevant.

Why this answer

Timeliness is critical because stale data can lead to incorrect model predictions, especially in dynamic environments like network traffic analysis or fraud detection. For AI models, data must reflect the current state of the system to ensure relevance and reliability. Without timely data, the model may act on outdated patterns, reducing its effectiveness.

Exam trap

CompTIA often tests the distinction between data quality dimensions and data characteristics, so the trap here is that candidates confuse completeness or volume with the three critical dimensions (timeliness, consistency, accuracy) that directly impact model reliability.

Full explanation →
640
MCQmedium

A team is evaluating a fine-tuned LLM for a code generation task. They notice the model rarely generates correct syntax but often produces plausible-looking code. Which evaluation metric is MOST appropriate to quantify this issue?

A.BLEU score
B.Perplexity
C.Pass@k (execution success rate)
D.Exact match accuracy
AnswerC

Pass@k runs the generated code against test cases, directly measuring functional correctness.

Why this answer

Pass@k measures the probability that at least one of k generated code samples passes a set of unit tests, directly quantifying execution correctness. Since the model produces plausible-looking but syntactically incorrect code, execution success rate (Pass@k) is the most appropriate metric to capture whether the code actually runs correctly, unlike surface-level similarity metrics.

Exam trap

Cisco often tests the misconception that BLEU or exact match are sufficient for code generation evaluation, but the trap here is that plausible-looking code can score high on n-gram overlap while being syntactically or semantically invalid, making execution-based metrics like Pass@k the only reliable measure of functional correctness.

How to eliminate wrong answers

Option A is wrong because BLEU score measures n-gram overlap between generated and reference text, which rewards lexical similarity but cannot detect syntax errors or functional correctness; plausible-looking code can have high BLEU while being invalid. Option B is wrong because perplexity measures how well a language model predicts a sequence, reflecting fluency rather than code execution validity; low perplexity can still yield syntactically invalid code. Option D is wrong because exact match accuracy requires the generated code to match a single reference exactly, which is too brittle for code generation where multiple valid solutions exist and does not assess runtime behavior.

Full explanation →
641
MCQeasy

A small business launched a customer support chatbot powered by a pre-trained language model. The chatbot was fine-tuned on a dataset of past support tickets. For the first week, it performed well, accurately answering 85% of queries. After a routine software update that included a new version of the underlying language model library, the chatbot's accuracy dropped to 60% and it began giving nonsensical responses to some questions. The update did not change any code or configuration specific to the chatbot. The business has a backup of the previous environment. What is the MOST appropriate immediate action?

A.Retrain the chatbot on the original dataset using the new library version.
B.Add more intents to the chatbot's configuration to cover the errors.
C.Increase the model's temperature parameter to 1.5 to encourage more varied responses.
D.Roll back the software update to the previous version of the language model library.
AnswerD

The update likely introduced a regression; rolling back restores functionality immediately.

Why this answer

The most appropriate immediate action is to roll back the software update to the previous version of the language model library (Option D). The accuracy drop and nonsensical responses are directly caused by the library update, which likely changed internal model behavior (e.g., tokenization, attention mechanisms, or default hyperparameters) without any code or configuration changes. Restoring the previous environment immediately resolves the issue and allows the business to investigate the library changes in a controlled manner.

Exam trap

CompTIA often tests the misconception that retraining or adjusting hyperparameters can fix a regression caused by an underlying library change, when the immediate and correct action is to roll back to the known-good environment.

How to eliminate wrong answers

Option A is wrong because retraining the chatbot on the original dataset using the new library version does not address the root cause—the library update itself may have altered model internals (e.g., tokenizer version, default parameters) that cannot be fixed by retraining alone, and retraining is time-consuming and not an immediate fix. Option B is wrong because adding more intents does not resolve the underlying model behavior change; the chatbot is producing nonsensical responses, not missing intents, so intent expansion is irrelevant to the core issue. Option C is wrong because increasing the temperature parameter to 1.5 would make responses more random and less coherent, worsening the nonsensical outputs; temperature controls output randomness, not model correctness or library compatibility.

Full explanation →
642
Multi-Selecthard

A data science team uses Vertex AI for model training and deployment. They want to implement CI/CD for ML pipelines. Which THREE Google Cloud services should they integrate?

Select 3 answers
A.Vertex AI Pipelines
B.Cloud Deploy
C.BigQuery
D.Cloud Build
E.Google Kubernetes Engine (GKE)
AnswersA, B, D

Vertex AI Pipelines is the workflow orchestrator for ML CI/CD.

Why this answer

Vertex AI Pipelines orchestrates ML workflows; Cloud Build automates builds; Cloud Deploy manages deployments. BigQuery is for analytics; GKE is for containers but not CI/CD specific.

Full explanation →
643
Multi-Selectmedium

A data scientist is building a recommendation system for an e-commerce platform. The dataset includes user purchase history, product descriptions, and user demographics. The goal is to recommend products that a user is likely to purchase. Which TWO techniques are most appropriate for this task? (Select TWO.)

Select 2 answers
A.Content-based filtering
B.Association rule mining
C.Linear regression
D.Anomaly detection
E.Collaborative filtering
AnswersA, E

Uses product descriptions and demographics to match user preferences.

Why this answer

Collaborative filtering uses user-item interactions; content-based filtering uses item features. Association rule mining is for basket analysis; regression is not typically used for recommendations; anomaly detection is for outliers.

Full explanation →
644
MCQeasy

An organization uses a third-party pre-trained model for a medical diagnosis system. Which supply chain security measure is MOST critical to verify before deployment?

A.Reviewing the model's software bill of materials (SBOM)
B.Ensuring the model is hosted on a private cloud
C.Scanning the model for backdoors
D.Checking for proper API key management
AnswerC

Backdoor detection is essential to ensure the model hasn't been maliciously altered to produce targeted failures.

Why this answer

Option C is correct because third-party pre-trained models can contain hidden backdoors inserted during training, such as poisoned weights that trigger misclassification on specific inputs. In a medical diagnosis system, such backdoors could cause life-threatening errors, making verification against adversarial manipulation the most critical supply chain security measure before deployment.

Exam trap

Cisco often tests the distinction between general security hygiene (SBOM, hosting, API keys) and the model-specific threat of backdoors, which is unique to AI supply chain security and the most critical to verify for pre-trained models.

How to eliminate wrong answers

Option A is wrong because while reviewing a software bill of materials (SBOM) is important for tracking dependencies, it does not detect backdoors or maliciously inserted behaviors in the model weights themselves. Option B is wrong because hosting the model on a private cloud addresses data privacy and access control but does not verify the integrity of the model's internal logic or protect against pre-existing backdoors. Option D is wrong because proper API key management secures access to the model endpoint but does not inspect the model for embedded threats like backdoors or trojans.

Full explanation →
645
MCQmedium

A team uses Kubeflow to manage ML workflows on Kubernetes. They want to automate hyperparameter tuning for a training job. Which Kubeflow component should they use?

A.KFServing
B.Kubeflow Notebooks
C.Kubeflow Pipelines
D.Kubeflow Katib
AnswerD

Katib provides automated hyperparameter tuning with various algorithms.

Why this answer

Katib is the hyperparameter tuning component in Kubeflow. Pipelines orchestrate workflows; KFServing is for inference; Notebooks are for development.

Full explanation →
646
Multi-Selectmedium

A DevOps team is deploying a machine learning model using a CI/CD pipeline. They want to ensure the model is reproducible and traceable. Which TWO practices should they implement?

Select 2 answers
A.Version the training dataset and code using Git and DVC.
B.Manually deploy the model to production after approval.
C.Store only the final model artifact in a shared drive.
D.Use a spreadsheet to record model version numbers.
E.Package the model in a Docker container with a fixed base image.
AnswersA, E

Version control for data and code ensures exact reproduction of training.

Why this answer

Versioning the training dataset and code using Git and DVC (Data Version Control) ensures that every model training run can be exactly reproduced by tracking changes to both the source code and the data. DVC stores metadata pointers in Git while the actual data is stored in a remote cache, enabling precise rollback and audit trails. This practice directly supports reproducibility and traceability, which are core requirements for MLOps and CI/CD pipelines.

Exam trap

Cisco often tests the misconception that manual steps or simple documentation (like spreadsheets) are sufficient for traceability, when in fact automated version control and containerization are required for true reproducibility in a CI/CD pipeline.

Full explanation →
647
MCQmedium

An LLM-powered application occasionally generates factual-sounding but incorrect information. Users rely on this output for decision-making. Which risk does this primarily represent?

A.Hallucinations and over-reliance
B.Sensitive information disclosure
C.Model denial of service
D.Prompt injection
AnswerA

Hallucinations produce false information, and over-reliance amplifies the risk.

Why this answer

The scenario describes an LLM generating plausible but incorrect information (hallucination) and users relying on it for decisions (over-reliance). This directly matches the combined risk of hallucinations and over-reliance, as the model's confident but false outputs can lead to poor decision-making without proper verification.

Exam trap

Cisco often tests the distinction between inherent model flaws (hallucinations) and external attacks (prompt injection), so candidates may confuse the two because both involve unexpected outputs, but the root cause differs—internal generation vs. external manipulation.

How to eliminate wrong answers

Option B is wrong because sensitive information disclosure involves the model leaking private data (e.g., PII, secrets) from its training set or context, not generating factually incorrect content. Option C is wrong because model denial of service refers to overwhelming the system with requests to cause resource exhaustion, not the quality or accuracy of outputs. Option D is wrong because prompt injection is an adversarial attack where crafted inputs manipulate the model's behavior (e.g., bypassing safeguards), not an inherent generation of incorrect facts.

Full explanation →
648
MCQmedium

A company wants to build a customer service chatbot that answers questions about their internal policy documents. The documents are updated monthly, and the team cannot afford to retrain a model each time. Which approach is MOST appropriate?

A.Train a custom model from scratch on the policy documents each month
B.Use Retrieval-Augmented Generation (RAG) with the policy documents indexed in a vector store
C.Fine-tune a base LLM on the policy documents monthly
D.Use a larger foundation model with a longer context window and paste all documents into each prompt
AnswerB

RAG retrieves relevant document chunks at query time, ensuring the chatbot always answers from the latest uploaded documents without any model retraining.

Why this answer

Retrieval-Augmented Generation (RAG) is the most appropriate approach because it allows the chatbot to answer questions based on the latest policy documents without retraining the model. By indexing the documents in a vector store, the system retrieves relevant chunks for each query and passes them to a pre-trained LLM for generation, ensuring up-to-date responses with minimal maintenance overhead.

Exam trap

Cisco often tests the misconception that fine-tuning or training from scratch is necessary for domain-specific knowledge, when in fact RAG provides a cost-effective, zero-retraining solution for frequently updated documents.

How to eliminate wrong answers

Option A is wrong because training a custom model from scratch each month is computationally expensive, time-consuming, and requires significant expertise and data, which contradicts the constraint of not being able to retrain a model each time. Option C is wrong because fine-tuning a base LLM monthly still requires retraining the model on new documents, which incurs similar costs and complexity as training from scratch, and does not solve the update problem efficiently. Option D is wrong because pasting all policy documents into each prompt exceeds typical context window limits (e.g., 4K-32K tokens), leading to high latency, token costs, and potential loss of relevant information due to the model's attention limitations.

Full explanation →
649
Multi-Selectmedium

A team is designing a RAG system for a large collection of PDFs. They need to choose document chunking strategies. Which TWO strategies are considered best practices? (Choose two.)

Select 2 answers
A.Semantic chunking (e.g., sentence or paragraph boundaries)
B.Fixed-size chunking with no overlap
C.Hierarchical chunking (sections, subsections)
D.Single chunk per document
E.Random character-length chunks
AnswersA, C

Semantic chunking maintains coherence within each chunk, improving embedding quality.

Why this answer

Semantic chunking preserves natural boundaries, and hierarchical chunking respects document structure, both improving retrieval quality.

Full explanation →
650
MCQeasy

An organization wants to train a machine learning model on sensitive patient data without exposing individual records. Which privacy-preserving technique allows the model to learn from data distributed across multiple hospitals without raw data leaving each site?

A.Homomorphic encryption
B.Federated learning
C.k-anonymity
D.Differential privacy
AnswerB

Federated learning trains on decentralized data and only shares model weights, preserving privacy.

Why this answer

Federated learning trains models locally at each site and only shares model updates, not raw data. Differential privacy adds noise, homomorphic encryption allows computation on encrypted data, and k-anonymity is a data anonymization technique.

Full explanation →
651
MCQeasy

A data scientist is preparing a dataset for a supervised learning model. The dataset contains missing values in 15% of the rows for a numeric feature. Which preprocessing technique should be applied to minimize bias?

A.Remove all rows with missing values.
B.Impute missing values with the mean of the feature.
C.Encode the missing values as a separate category.
D.Use a model that handles missing values natively.
AnswerB

Mean imputation preserves the dataset size and is appropriate for numeric features when missingness is random.

Why this answer

Imputing missing values with the mean of the feature is the correct approach because it preserves the overall distribution of the numeric feature and avoids introducing systematic bias when only 15% of rows are affected. Removing rows would reduce sample size and potentially introduce selection bias, while encoding missing values as a separate category is inappropriate for numeric features in regression or classification tasks. Mean imputation is a standard, low-bias technique for MCAR (Missing Completely At Random) data in supervised learning pipelines.

Exam trap

Cisco often tests the misconception that removing rows with missing values is always safe, but the trap here is that candidates overlook how even 15% missingness can bias the model if the missingness is not random, making imputation a more robust preprocessing step.

How to eliminate wrong answers

Option A is wrong because removing rows with missing values reduces the dataset size and can introduce selection bias if the missingness is not completely random (MCAR), leading to a non-representative training set. Option C is wrong because encoding missing values as a separate category is designed for categorical features, not numeric features; for numeric features, this would create an artificial ordinal relationship or require one-hot encoding that distorts the feature's numerical meaning. Option D is wrong because relying on a model that handles missing values natively (e.g., tree-based models) is not a preprocessing technique; it shifts the burden to the algorithm, which may still produce biased splits or require additional assumptions about missingness patterns.

Full explanation →
652
Multi-Selecteasy

Which TWO are characteristics of supervised learning?

Select 2 answers
A.Does not require target variable
B.Requires labeled data
C.Uses reinforcement signals
D.Learns to cluster data
E.Predicts continuous or categorical output
AnswersB, E

Supervised learning uses input-output pairs for training.

Why this answer

Supervised learning requires labeled data because the model learns a mapping from input features to a known target variable. The correct answer B is fundamental: without labeled examples, the algorithm cannot calculate a loss function to adjust its weights during training.

Exam trap

Cisco often tests the distinction between supervised and unsupervised learning by presenting 'clustering' or 'reinforcement signals' as plausible characteristics of supervised learning, trapping candidates who confuse task types.

Full explanation →
653
MCQmedium

An AI system trained on historical medical records shows that certain racial groups have higher predicted risk for a disease. The data reflects real-world differences in diagnosis rates due to unequal access to healthcare. Which type of bias is this?

A.Algorithmic bias
B.Selection bias
C.Historical bias
D.Confirmation bias
AnswerC

The data reflects historical disparities in healthcare access, which is a classic example of historical bias.

Why this answer

Historical bias occurs when the training data reflects past societal inequities. Selection bias would occur if the sample is not representative; confirmation bias stems from the model's own predictions; algorithmic bias arises from the model design.

Full explanation →
654
MCQhard

An AI model for skin cancer detection achieves high accuracy but performs poorly on dark skin tones. The team wants to evaluate whether the model is calibrated across skin tones. Which fairness metric should they use?

A.Equalised odds
B.Demographic parity
C.Individual fairness
D.Calibration
AnswerD

Calibration checks that for a given predicted probability, the actual outcome rate is the same across groups.

Why this answer

Calibration is the correct metric because it directly measures whether the predicted probabilities of skin cancer match the actual outcomes across different skin tones. A model can have high overall accuracy but be miscalibrated for a subgroup if its confidence scores are systematically over- or under-confident for that group. In this scenario, the team needs to check if the model's risk scores are equally reliable for dark skin tones as for light skin tones, which is exactly what calibration assesses.

Exam trap

Cisco often tests the distinction between fairness metrics by presenting a scenario where 'accuracy' is high but subgroup performance differs, and candidates mistakenly choose equalized odds or demographic parity instead of recognizing that the core issue is confidence score reliability, i.e., calibration.

How to eliminate wrong answers

Option A is wrong because equalized odds requires that the true positive rate and false positive rate are equal across groups, which is a measure of error rate fairness, not calibration. Option B is wrong because demographic parity requires that the proportion of positive predictions is the same across groups, which can be achieved even if the model is poorly calibrated. Option C is wrong because individual fairness requires that similar individuals receive similar predictions, which is a different concept from group-level calibration across skin tones.

Full explanation →
655
MCQmedium

A company uses an LLM to generate code. They want to ensure that the model does not accidentally output sensitive internal logic. Which practice should they implement?

A.Rate limiting API calls
B.Red teaming the model
C.Output filtering to block sensitive patterns
D.Federated learning
AnswerC

Output filtering scans the model's output for sensitive content and blocks it.

Why this answer

Output filtering is the correct practice because it directly inspects the model's generated text for patterns that match sensitive internal logic (e.g., API keys, source code snippets, or proprietary algorithms) and blocks or redacts them before the output is returned to the user. This is a reactive security control that operates at the application layer, ensuring that even if the LLM inadvertently generates sensitive content, it is never exposed. Rate limiting, red teaming, and federated learning address different concerns (availability, vulnerability discovery, and data privacy during training, respectively) and do not prevent the accidental leakage of internal logic in real-time outputs.

Exam trap

Cisco often tests the distinction between proactive security testing (red teaming) and reactive runtime controls (output filtering), leading candidates to confuse vulnerability discovery with real-time content protection.

How to eliminate wrong answers

Option A is wrong because rate limiting controls the frequency of API requests to prevent abuse or denial-of-service, but it does not inspect or filter the content of the LLM's responses, so sensitive internal logic could still be output. Option B is wrong because red teaming is a proactive testing methodology to identify vulnerabilities and weaknesses in the model, but it is not a runtime control that prevents sensitive outputs from being delivered to users. Option D is wrong because federated learning is a distributed training technique that keeps training data local to preserve privacy, but it does not affect the model's inference-time outputs and cannot filter generated content for sensitive patterns.

Full explanation →
656
MCQmedium

A team is considering whether to fine-tune a base LLM or use RAG for a question-answering system over a large, static corpus of scientific papers. The answer must be highly accurate and grounded in the papers. Which approach is BEST and why?

A.Fine-tuning because it adapts the model to the scientific domain
B.Fine-tuning because it is faster at inference time
C.RAG because it retrieves and grounds answers in the source documents
D.RAG because it does not require any labeled data
AnswerC

RAG retrieves relevant passages and passes them as context, ensuring the answer is directly supported by the source, reducing hallucination.

Why this answer

RAG retrieves exact passages and grounds answers in the source, which is critical for accuracy and citation in scientific domains. Fine-tuning may memorize but can hallucinate.

Full explanation →
657
MCQeasy

Which of the following best describes the difference between narrow AI and general AI?

A.Narrow AI is designed for a specific task; general AI aims to perform any cognitive task a human can.
B.Narrow AI relies on supervised learning; general AI uses unsupervised learning exclusively.
C.Narrow AI requires large datasets; general AI can learn from few examples.
D.Narrow AI can perform any intellectual task; general AI is limited to specific tasks.
AnswerA

Narrow AI excels at one domain (e.g., chess), whereas general AI would be versatile.

Why this answer

Narrow AI specializes in one task, while general AI would possess human-like cognitive abilities across domains.

Full explanation →
658
MCQeasy

A hospital wants to train a diagnostic model using patient data from multiple hospitals without sharing raw patient records. Which technique enables collaborative model training while keeping data decentralised?

A.Pseudonymisation
B.Differential privacy
C.Federated learning
D.Anonymisation
AnswerC

Federated learning trains a shared model by aggregating only model updates, never raw patient data.

Why this answer

Federated learning trains models across decentralised data without exchanging raw data. Differential privacy adds noise. Anonymisation removes identifiers.

Pseudonymisation replaces identifiers with pseudonyms.

Full explanation →
659
MCQmedium

A company is building a recommendation system that uses user embeddings stored in a vector database. The system must retrieve the top 10 most similar items for a given user query. Which vector database feature is MOST critical for this task?

A.Built-in data versioning
B.ACID transaction support
C.Approximate nearest neighbor (ANN) search
D.SQL query interface
AnswerC

ANN search is designed to efficiently find the closest vectors to a query vector, which is exactly what the recommendation system requires.

Why this answer

Approximate nearest neighbor (ANN) search is the most critical feature because it enables the vector database to efficiently find the top-10 most similar items to a user query embedding without scanning the entire dataset. Unlike exact nearest neighbor search, ANN algorithms (e.g., HNSW, IVF) trade a small amount of accuracy for massive performance gains, which is essential for real-time recommendation systems handling millions of high-dimensional vectors.

Exam trap

Cisco often tests the misconception that SQL or ACID features are needed for all database tasks, but in vector databases, the critical differentiator is the ANN search algorithm, not traditional relational or transactional capabilities.

How to eliminate wrong answers

Option A is wrong because built-in data versioning manages historical changes to data but does not directly impact the speed or accuracy of similarity search; it is irrelevant to the core retrieval task. Option B is wrong because ACID transaction support ensures data consistency and reliability during writes but does not optimize or accelerate vector similarity queries; it addresses transactional integrity, not search performance. Option D is wrong because a SQL query interface is designed for structured relational queries and lacks native support for high-dimensional vector similarity operations; using SQL for nearest neighbor search would require inefficient full-table scans or custom extensions, defeating the purpose of a vector database.

Full explanation →
660
MCQeasy

A data scientist is preparing a dataset for a classification model. The dataset contains a column "Age" with 10% missing values and a column "Income" with 30% missing values. Which imputation strategy is MOST appropriate to minimize bias?

A.Replace missing Age with the mean and missing Income with the median.
B.Delete all rows with missing values.
C.Replace missing Age with the mode and missing Income with a constant value.
D.Replace missing values with zeros.
AnswerA

Mean for symmetric Age, median for skewed Income minimizes bias.

Why this answer

Option A is correct because using mean imputation for Age (10% missing) and median imputation for Income (30% missing) minimizes bias. Mean is suitable for roughly symmetric distributions with low missingness, while median is robust to outliers and skewness, which is common in income data. This combination reduces distortion of central tendency and preserves data integrity better than uniform methods.

Exam trap

CompTIA often tests the misconception that a single imputation method (e.g., mean for all columns) is universally appropriate, when in fact the choice must consider the missingness rate and the distribution of each feature to minimize bias.

How to eliminate wrong answers

Option B is wrong because deleting all rows with missing values (listwise deletion) reduces sample size and can introduce selection bias, especially when missingness is not completely at random (MCAR). Option C is wrong because replacing Age with the mode is inappropriate for continuous variables like age, as it discards variance and can create artificial clusters; replacing Income with a constant value (e.g., 0) introduces systematic bias and distorts the distribution. Option D is wrong because replacing missing values with zeros for both Age and Income is arbitrary and unrealistic, leading to severe underestimation of central tendency and inflated model error.

Full explanation →
661
MCQmedium

A company wants to store unstructured text data for AI model training while enabling SQL-based queries for analytics. Which storage solution should they use as the primary data source?

A.A vector database like Pinecone
B.A data lake like Amazon S3
C.A data warehouse like Snowflake
D.A NoSQL database like DynamoDB
AnswerB

Data lakes store unstructured data in native format; SQL queries can be run on top via services like Athena or Presto.

Why this answer

Amazon S3 is a highly scalable object storage service that can store unstructured text data in its native format (e.g., CSV, JSON, Parquet) and supports SQL-based queries via services like Amazon Athena or S3 Select. This makes it ideal as a primary data source for AI model training while enabling analytics without requiring data transformation or loading into a separate system.

Exam trap

The trap here is that candidates often confuse a data warehouse (Snowflake) with a data lake (S3) for storing unstructured data, forgetting that data warehouses require structured schemas and are not designed for raw, schema-on-read storage.

How to eliminate wrong answers

Option A is wrong because vector databases like Pinecone are optimized for similarity search and embedding storage, not for SQL-based analytics or general unstructured text storage for training. Option C is wrong because data warehouses like Snowflake require structured, schema-on-write data and are not designed to store raw unstructured text files as the primary source. Option D is wrong because NoSQL databases like DynamoDB are key-value/document stores that enforce schema constraints and are not optimized for SQL queries on large volumes of unstructured text data.

Full explanation →
662
Multi-Selecthard

Which TWO of the following are effective defenses against adversarial evasion attacks on image classifiers?

Select 2 answers
A.Data augmentation
B.Gradient masking
C.Adversarial training
D.Input validation
E.Feature squeezing
AnswersB, C

Gradient masking obscures gradient information to prevent crafting adversarial examples.

Why this answer

Adversarial training and gradient masking (e.g., defensive distillation) are common defenses. Data augmentation helps generalization but not specifically against adversarial perturbations; feature squeezing reduces input complexity; input validation is generic.

Full explanation →
663
MCQhard

The exhibit shows a model configuration for a classification task with 10 classes. What is wrong with this setup?

A.The loss function should be categorical crossentropy, not mean squared error
B.The metric should be precision, not accuracy
C.The activation should be sigmoid in hidden layers
D.The optimizer should be SGD, not Adam
AnswerA

Correct: MSE is for regression; classification requires crossentropy loss.

Why this answer

In a multi-class classification task with 10 classes, the correct loss function is categorical crossentropy because it measures the dissimilarity between the true probability distribution and the predicted probability distribution. Mean squared error (MSE) is designed for regression tasks and penalizes errors in a way that is not suitable for classification probabilities, leading to poor gradient behavior and slower convergence.

Exam trap

Cisco often tests the misconception that MSE can be used as a generic loss function for any task, but in classification, crossentropy is specifically designed to handle probability distributions and one-hot encoding.

How to eliminate wrong answers

Option B is wrong because accuracy is the standard metric for multi-class classification tasks; precision is typically used for binary classification or when focusing on specific class performance, but it is not a general replacement for accuracy. Option C is wrong because sigmoid activation in hidden layers can cause vanishing gradients and is not optimal; ReLU or its variants are preferred for hidden layers to mitigate gradient issues. Option D is wrong because Adam is a widely used optimizer that adapts learning rates and often outperforms SGD in practice; there is no inherent problem with using Adam for this setup.

Full explanation →
664
MCQmedium

A team is using a cloud AI service with a pay-per-token pricing model. They want to minimize costs while maintaining response quality. Which strategy is MOST effective?

A.Switch to a smaller, less capable model
B.Increase the batch size for API calls
C.Use prompt caching for repeated query patterns
D.Reduce the model's max_tokens to a very low value
AnswerC

Caching avoids reprocessing identical prompts, saving token costs and reducing latency while preserving quality.

Why this answer

Prompt caching reduces costs by avoiding redundant token processing for repeated query patterns. The cloud AI service charges per token, so caching the prefix of frequent requests (e.g., system prompts or common context) means only the new, unique tokens are billed, directly lowering expenditure without sacrificing response quality.

Exam trap

Cisco often tests the misconception that reducing model size or output length is the only way to cut costs, but the correct strategy leverages architectural features like prompt caching to reduce token consumption without affecting quality.

How to eliminate wrong answers

Option A is wrong because switching to a smaller, less capable model typically reduces response quality, which contradicts the requirement to maintain quality. Option B is wrong because increasing batch size for API calls does not reduce per-token cost; it may improve throughput but still charges for all tokens processed. Option D is wrong because reducing max_tokens to a very low value can truncate responses, degrading quality, and does not address the cost of input tokens or repeated patterns.

Full explanation →
665
MCQmedium

A data science team needs to implement privacy-preserving ML for a healthcare model. They require that individual patient records cannot be distinguished in the training output. Which technique should be applied?

A.Differential privacy
B.Homomorphic encryption
C.Model pruning
D.Federated learning
AnswerA

Correct. Differential privacy provides formal guarantees against membership inference.

Why this answer

Differential privacy is the correct technique because it adds calibrated noise to the training process or query outputs, ensuring that the inclusion or exclusion of any single patient record does not significantly affect the model's output. This provides a formal mathematical guarantee that individual records cannot be distinguished, which directly meets the requirement for privacy-preserving ML in healthcare.

Exam trap

Cisco often tests the misconception that federated learning alone provides privacy, but without differential privacy, federated learning can still leak individual patient data through model inversion or membership inference attacks.

How to eliminate wrong answers

Option B is wrong because homomorphic encryption allows computations on encrypted data but does not prevent inference about individual records in the model output; it protects data in transit or at rest, not the distinguishability of training records. Option C is wrong because model pruning reduces model size by removing redundant parameters, which has no effect on privacy guarantees and does not prevent individual record identification. Option D is wrong because federated learning trains models across decentralized data without sharing raw data, but the model updates or final model can still leak information about individual records through gradient or membership inference attacks without additional differential privacy mechanisms.

Full explanation →
666
MCQmedium

An AI engineer is training a deep neural network for image recognition. The training loss decreases steadily for the first few epochs but then plateaus and starts to oscillate. Which adjustment is most likely to improve convergence?

A.Add more layers
B.Increase the learning rate
C.Increase the batch size
D.Reduce the learning rate
AnswerD

A lower learning rate can smooth convergence and reduce oscillation.

Why this answer

The plateau and oscillation of the training loss indicate that the optimizer is overshooting the minimum due to a learning rate that is too high. Reducing the learning rate allows the optimizer to take smaller, more precise steps, dampening oscillations and enabling convergence to a lower loss. This is a standard technique in gradient descent optimization, often implemented via learning rate schedules or adaptive methods like Adam.

Exam trap

Cisco often tests the misconception that increasing the learning rate speeds up convergence, when in fact it causes divergence or oscillation, and that adding layers always improves performance, ignoring the risk of overfitting and optimization difficulty.

How to eliminate wrong answers

Option A is wrong because adding more layers increases model complexity, which typically exacerbates overfitting and can worsen convergence issues when the loss is already oscillating. Option B is wrong because increasing the learning rate would make the oscillations larger and more erratic, moving the optimizer further from the minimum. Option C is wrong because increasing the batch size reduces the variance of gradient estimates but does not address the fundamental issue of an overly large step size causing oscillations; it may even slow convergence by requiring more epochs to process the same data.

Full explanation →
667
MCQhard

A large e-commerce company uses a recommendation system based on collaborative filtering. The system uses a matrix factorization model that is trained nightly on the entire user-item interaction history. Recently, the company launched a flash sale with thousands of new products. Users are reporting that the recommendations are not showing the new products, even for users who have purchased them during the sale. The data engineering team notices that the new products have very few interactions in the training data. The model's loss on the validation set has increased, and the recall@10 metric has dropped from 0.45 to 0.32. The team needs to improve the recommendation of new items without retraining the entire model from scratch every hour. Which approach should the team take?

A.Use a hybrid model that combines collaborative filtering with content-based features from product metadata
B.Retrain the model every hour to incorporate new interactions quickly
C.Remove the new products from the recommendation pool until they accumulate enough interactions
D.Increase the number of latent factors in the matrix factorization model
AnswerA

Content-based features allow the model to recommend new items based on their attributes, overcoming the cold-start problem.

Why this answer

Option A is correct because a hybrid model that combines collaborative filtering with content-based features (e.g., product metadata like category, price, or description) can recommend new products even with zero or very few user interactions. The content-based component leverages item attributes to compute similarity between new and existing items, enabling the system to surface new products without requiring extensive interaction history. This approach addresses the cold-start problem for new items while preserving the collaborative filtering signal for established items, and it does not require retraining the entire model from scratch every hour.

Exam trap

CompTIA often tests the misconception that simply retraining more frequently or increasing model complexity (e.g., more latent factors) can solve the cold-start problem, but the core issue is the lack of interaction data for new items, which requires a content-based or hybrid approach to leverage item metadata.

How to eliminate wrong answers

Option B is wrong because retraining the model every hour would be computationally expensive and operationally impractical for a large-scale system with thousands of new products; it also does not solve the fundamental cold-start issue since new items still have very few interactions in each hourly training window. Option C is wrong because removing new products from the recommendation pool defeats the business purpose of the flash sale, which is to promote and surface new items to users, and it would lead to a poor user experience and lost revenue. Option D is wrong because increasing the number of latent factors in matrix factorization does not address the lack of interaction data for new items; it may even exacerbate overfitting to sparse data and increase computational cost without improving cold-start recommendations.

Full explanation →
668
MCQeasy

A company implements a chatbot using a rule-based system. Users complain the chatbot cannot handle new queries. Which AI approach should be considered to improve flexibility?

A.Expert system
B.Natural language processing (NLP)
C.Robotic process automation
D.Machine learning
AnswerD

ML enables the system to learn patterns from data.

Why this answer

Machine learning (ML) enables a chatbot to learn from new data and adapt to unseen queries, unlike a static rule-based system. By training on historical conversations, an ML model can generalize patterns and handle novel inputs without requiring explicit rules for every scenario.

Exam trap

CompTIA often tests the misconception that NLP alone is sufficient for adaptive chatbots, but NLP is a component of understanding language, not a learning mechanism—machine learning is required for flexibility.

How to eliminate wrong answers

Option A is wrong because an expert system is also rule-based, relying on a fixed knowledge base and inference engine, which cannot adapt to new queries without manual rule updates. Option B is wrong because natural language processing (NLP) alone provides text understanding (e.g., tokenization, parsing) but does not inherently learn from new data; it must be combined with ML for adaptive behavior. Option C is wrong because robotic process automation (RPA) automates repetitive, rule-based tasks in structured environments and cannot handle the variability of new, unseen queries.

Full explanation →
669
MCQhard

An organization uses an LLM to generate financial reports. They want to ensure the model does not output sensitive customer data that it may have memorized during training. Which technique should be implemented in the AI pipeline to detect and block such outputs?

A.Input validation
B.Output filtering
C.Rate limiting
D.Federated learning
AnswerB

Output filtering inspects model responses for sensitive content and prevents leakage.

Why this answer

Output filtering is the correct technique because it operates after the LLM generates a response, scanning the output for sensitive data patterns (e.g., PII, financial account numbers) and blocking or redacting them before delivery. This directly addresses the risk of the model regurgitating memorized customer data from its training set, which input validation cannot catch since the sensitive data appears only in the output.

Exam trap

Cisco often tests the distinction between input controls (validation) and output controls (filtering), tricking candidates into choosing input validation because they focus on preventing data from entering the system rather than catching data that the model generates from memory.

How to eliminate wrong answers

Option A is wrong because input validation sanitizes data entering the model (e.g., user prompts), but it cannot prevent the model from generating memorized sensitive data in its output, which is a generative behavior. Option C is wrong because rate limiting controls the frequency of API requests to prevent abuse or denial-of-service, not the content of the model's responses. Option D is wrong because federated learning is a distributed training technique that keeps data local to preserve privacy during model training, but it does not inspect or block outputs at inference time.

Full explanation →
670
Multi-Selecteasy

An organization is deploying an image classification model to detect defects on a production line. Which TWO steps are essential during the model monitoring phase of the AI project lifecycle?

Select 2 answers
A.Track model performance metrics such as precision and recall on a held-out test set over time
B.Perform data cleaning and normalization on the production data before inference
C.Periodically retrain the model with newly labeled data
D.Automatically roll back to the previous model version if a metric drops below a threshold
E.Monitor for data drift between the training data distribution and incoming production data
AnswersA, E

Tracking metrics on live data (or periodic golden sets) identifies model drift.

Why this answer

Monitoring should track data drift (changes in input distribution) and model drift (performance degradation). Retraining is an action after monitoring, not a step of monitoring itself. Data preparation is completed before deployment.

Full explanation →
671
MCQhard

A company uses a machine learning model to recommend products to customers. The marketing team notices that the model is recommending high-profit items more frequently than low-profit items, even when customers are likely to prefer the latter. This behavior is causing customer dissatisfaction. Which approach would best align the model with customer preferences while maintaining profitability?

A.Train the model with a loss function that weights profit more heavily than customer satisfaction.
B.Use a multi-objective optimization framework to balance profit and customer satisfaction.
C.Adjust the model's hyperparameters to reduce the influence of profit features.
D.Remove profit data from the training set and only use customer preference data.
AnswerB

Multi-objective optimization explicitly considers trade-offs between competing objectives, leading to more balanced recommendations.

Why this answer

Option D is correct because multi-objective optimization allows the model to balance multiple goals (e.g., profit and customer satisfaction) explicitly. Option A is incorrect because it still prioritizes profit, which may not address satisfaction. Option B is incorrect because it completely removes profit, which may harm business goals.

Option C is incorrect because it only adjusts profit weights without true multi-objective framework.

Full explanation →
672
Multi-Selectmedium

A company is deploying an LLM-based system that can execute API calls on behalf of users. Which TWO measures should they implement to prevent excessive agency?

Select 2 answers
A.Implement strict output filtering
B.Restrict the LLM to read-only or low-risk actions
C.Apply rate limiting to API calls
D.Require human-in-the-loop approval for high-risk actions
E.Use input validation to sanitize user prompts
AnswersB, D

Limiting the scope of actions reduces the risk of unintended actions.

Why this answer

Option B is correct because restricting the LLM to read-only or low-risk actions directly limits the scope of actions the model can execute, preventing it from performing unauthorized or destructive operations. This is a fundamental principle of least privilege applied to LLM-based systems, ensuring that even if the model is compromised or misused, it cannot escalate its agency to high-impact actions.

Exam trap

Cisco often tests the distinction between input/output security measures (like filtering and validation) and agency control measures (like action restrictions and human oversight), leading candidates to confuse defensive coding with permission scoping.

Full explanation →
673
MCQhard

An AI developer observes that the training accuracy of a neural network is high, but the test accuracy is low. The model uses a ReLU activation function and Adam optimizer. Which approach is most likely to improve test accuracy?

A.Increase the learning rate
B.Add L2 regularization to the loss function
C.Switch to a stochastic gradient descent optimizer
D.Increase the number of epochs
AnswerB

L2 regularization penalizes large weights, preventing overfitting.

Why this answer

L2 regularization adds a penalty on large weights, reducing overfitting and improving test accuracy.

Full explanation →
674
Multi-Selectmedium

Which THREE techniques can help reduce overfitting in neural networks?

Select 3 answers
A.Increasing training data size
B.L2 regularization
C.Using a larger learning rate
D.Dropout
E.Increasing number of layers
AnswersA, B, D

More data helps the model generalize better.

Why this answer

Increasing the training data size helps reduce overfitting by providing the model with more examples to learn from, which reduces the variance and improves generalization. With more data, the model is less likely to memorize noise and instead learns the underlying patterns, making it more robust on unseen data.

Exam trap

Cisco often tests the misconception that increasing model complexity (e.g., more layers or larger learning rates) can help with overfitting, when in fact these changes typically worsen it by increasing variance or destabilizing training.

Full explanation →
675
MCQmedium

A team is evaluating an LLM-based chatbot that frequently hallucinates when answering questions about internal policies. Which testing approach would MOST effectively quantify this issue?

A.Evaluation frameworks for LLM output quality
B.Integration tests for API calls
C.Unit tests for the data pipeline
D.Regression testing of model accuracy over time
AnswerA

Evaluation frameworks specifically measure output quality metrics like faithfulness and hallucination rate.

Why this answer

Option A is correct because evaluation frameworks for LLM output quality, such as those using metrics like faithfulness, factuality, or ROUGE/BLEU scores, are specifically designed to detect and quantify hallucinations by comparing generated responses against a ground-truth knowledge base. This directly measures the rate at which the chatbot fabricates or misstates internal policy details, providing a quantitative baseline for improvement.

Exam trap

Cisco often tests the distinction between functional testing (e.g., API integration, data pipeline) and output quality evaluation, leading candidates to mistakenly choose integration or unit tests when the real issue is semantic accuracy of generated content.

How to eliminate wrong answers

Option B is wrong because integration tests for API calls verify that the chatbot's endpoints and external service interactions work correctly, but they do not assess the semantic accuracy or factual consistency of the generated text. Option C is wrong because unit tests for the data pipeline validate data ingestion, transformation, and storage logic, not the output quality of the LLM's responses. Option D is wrong because regression testing of model accuracy over time typically measures performance on a static benchmark (e.g., classification accuracy) rather than quantifying open-ended hallucination rates in a conversational context.

Full explanation →
Page 8

Page 9 of 14

Page 10

All pages

1234567891011121314

Practice AI0-001 by domain

Target a specific domain to shore up weak areas.

AI Infrastructure and TechnologiesAI SecurityAI Concepts and FoundationsAI Concepts and TechniquesMachine Learning and Deep LearningAI Models and Data EngineeringImplementing AI SolutionsAI Implementation and OperationsAI Security, Ethics and GovernanceAI Governance and Ethics
See all domains with question counts →
CompTIA AI+ AI0-001 AI0-001 Questions 601–675 | Page 9/14 | Courseiva