I85PREP
⌘ K
CAIP (AIP-210) study materials
CAIP (AIP-210) study resources

CAIP (AIP-210) practice questions

One question at a time. Build your CAIP (AIP-210) understanding.

First 50 of 200 questions. Each one lists the choices, the correct answer, and an explanation. The checker at the top of this page saves progress in this browser when it is available.

Continue in the practice tool above to load more questions, 50 at a time.

Question 1AI and ML concepts, technologies, and tools

In the AWS Certified AI Practitioner (AIP-210) domain covering AI/ML concepts, which statement best distinguishes supervised learning from unsupervised learning?

  1. Supervised learning requires reinforcement signals, while unsupervised learning uses reward functions.
  2. Supervised learning can only solve regression tasks, while unsupervised learning handles classification.
  3. Supervised learning uses labeled training data, while unsupervised learning finds structure in unlabeled data. Correct answer.
  4. Supervised learning always uses neural networks, while unsupervised learning uses only decision trees.

Explanation

Supervised learning trains on input-output pairs where the target label is known, allowing the model to learn a mapping that generalizes to new inputs. Unsupervised learning receives no target labels and instead discovers patterns such as clusters, associations, or reduced representations. The other distinctions describe algorithms or task types rather than the core labeling difference.

Question 2AI and ML concepts, technologies, and tools

A team wants to predict a continuous dollar amount for each customer. Which ML task formulation is most appropriate?

  1. Clustering
  2. Regression Correct answer.
  3. Dimensionality reduction
  4. Multi-class classification

Explanation

Regression models estimate a continuous numeric target such as a dollar amount, which is exactly the prediction goal here. Classification predicts discrete class labels, clustering groups unlabeled records, and dimensionality reduction compresses features rather than producing a continuous outcome. Choosing regression aligns the loss function and evaluation metrics with the numeric target.

Question 3AI and ML concepts, technologies, and tools

Which metric is most appropriate for evaluating a binary classifier when the positive class is rare and the cost of missing a positive case is high?

  1. Mean absolute error
  2. R-squared
  3. Accuracy
  4. Recall Correct answer.

Explanation

Recall measures the fraction of actual positives that the model correctly identifies, so it directly reflects how many costly positive cases are missed. Accuracy can look high on imbalanced data while missing most positives, and R-squared and mean absolute error are regression metrics. When false negatives carry high cost, recall is the natural primary metric.

Question 4AI and ML concepts, technologies, and tools

What is the primary purpose of splitting data into training, validation, and test sets?

  1. To reduce the total amount of data the model must process
  2. To estimate generalization performance and tune hyperparameters without leaking test information Correct answer.
  3. To guarantee the model achieves higher accuracy on the test set
  4. To remove the need for any regularization during training

Explanation

The training set fits model parameters, the validation set guides hyperparameter and model selection, and the held-out test set provides an unbiased estimate of generalization. Keeping the test set unseen prevents optimistic bias from information leakage. The split does not shrink data for speed or guarantee accuracy, and it does not replace regularization.

Question 5AI and ML concepts, technologies, and tools

A model performs very well on training data but poorly on validation data. Which condition does this describe?

  1. Overfitting Correct answer.
  2. Class imbalance
  3. Data leakage
  4. Underfitting

Explanation

Overfitting occurs when a model captures noise and idiosyncrasies of the training set, producing a large gap between training and validation performance. Underfitting instead shows poor performance on both sets, while leakage and imbalance are data issues that may or may not cause this specific gap. The described pattern is the textbook signature of overfitting.

Question 6AI and ML concepts, technologies, and tools

Which technique directly reduces overfitting by penalizing large model coefficients during training?

  1. Increasing model depth without limit
  2. Removing the validation set
  3. L1/L2 regularization Correct answer.
  4. Data augmentation removal

Explanation

L1 and L2 regularization add a penalty on coefficient magnitude to the loss, discouraging the model from fitting noise and thereby reducing overfitting. Removing validation data or endlessly increasing depth would worsen generalization assessment or capacity. Data augmentation removal would reduce training variety rather than control complexity.

Question 7AI and ML concepts, technologies, and tools

In the context of Amazon SageMaker, what is the role of a built-in algorithm container?

  1. It stores raw training data in Amazon S3 for later labeling.
  2. It replaces the need for any IAM permissions in the account.
  3. It packages algorithm code and dependencies so training and inference jobs can run on managed infrastructure. Correct answer.
  4. It automatically converts a regression problem into a classification problem.

Explanation

SageMaker built-in algorithm containers bundle the algorithm implementation and runtime dependencies, enabling managed training and hosting without custom code. They run on provisioned infrastructure and still require appropriate IAM roles and S3 data access. They do not alter the problem type or store raw data themselves.

Question 8AI and ML concepts, technologies, and tools

A company wants a fully managed service to extract text and key-value pairs from scanned documents without building custom ML models. Which AWS AI service fits best?

  1. Amazon Rekognition
  2. Amazon Textract Correct answer.
  3. Amazon Polly
  4. Amazon Forecast

Explanation

Amazon Textract is designed to detect and extract printed text, handwriting, and structured data such as forms and tables from scanned documents. Rekognition focuses on image and video analysis, Polly converts text to speech, and Forecast produces time-series forecasts. Textract therefore matches the document-extraction requirement.

Question 9AI and ML concepts, technologies, and tools

Which AWS service provides pre-trained computer vision capabilities such as object and scene detection in images?

  1. Amazon Kendra
  2. Amazon Comprehend
  3. Amazon Rekognition Correct answer.
  4. Amazon Transcribe

Explanation

Amazon Rekognition offers pre-trained image and video analysis, including object, scene, and face detection, making it the fit for computer vision tasks. Comprehend handles natural language processing, Transcribe converts speech to text, and Kendra provides intelligent enterprise search. The vision requirement maps directly to Rekognition.

Question 10AI and ML concepts, technologies, and tools

A retailer wants to group customers by purchasing behavior without predefined segments. Which approach applies?

  1. Linear regression on purchase amounts
  2. Clustering such as k-means Correct answer.
  3. Reinforcement learning with reward shaping
  4. Supervised classification with labeled segments

Explanation

Clustering is an unsupervised technique that partitions records into groups based on similarity, which suits discovering customer segments without predefined labels. Supervised classification requires known segment labels, regression predicts numeric values, and reinforcement learning learns from reward signals through interaction. K-means is a standard choice for this segmentation task.

Question 11AI and ML concepts, technologies, and tools

Which statement correctly describes the bias-variance tradeoff?

  1. Bias and variance always decrease together as model complexity grows.
  2. High bias reflects overly simple models, while high variance reflects models sensitive to training data fluctuations. Correct answer.
  3. High bias models fit training data too closely, while high variance models are too simple.
  4. Variance can be eliminated entirely by adding more features.

Explanation

Bias arises from simplifying assumptions that cause systematic error, characteristic of underfitting models, while variance reflects sensitivity to sampling noise, characteristic of overfitting models. Increasing complexity typically lowers bias but raises variance, so they do not always decrease together. Adding features generally increases variance rather than eliminating it.

Question 12AI and ML concepts, technologies, and tools

What does feature engineering aim to accomplish in an ML pipeline?

  1. Create or transform input variables so patterns are easier for models to learn Correct answer.
  2. Replace the need for model evaluation
  3. Reduce the number of AWS regions available to the account
  4. Guarantee that the model will never overfit

Explanation

Feature engineering derives, scales, encodes, or combines raw inputs into representations that expose useful signal to the learning algorithm. Better features can improve accuracy and interpretability but do not guarantee freedom from overfitting or remove the need for evaluation. It is unrelated to regional service availability.

Question 13AI and ML concepts, technologies, and tools

A dataset contains a feature recorded after the outcome occurred, and it strongly predicts the target. What problem should be suspected?

  1. High cardinality encoding
  2. Target leakage Correct answer.
  3. Underfitting
  4. Model drift

Explanation

Target leakage occurs when features contain information that would not be available at prediction time, such as post-outcome data, producing unrealistically strong performance. Underfitting is poor fit, cardinality encoding concerns categorical handling, and drift describes changing data distributions over time. The post-outcome feature is the classic leakage signal.

Question 14AI and ML concepts, technologies, and tools

Which AWS service is purpose-built for building, training, and deploying ML models with managed infrastructure and notebooks?

  1. Amazon SageMaker Correct answer.
  2. AWS Glue
  3. Amazon CloudFront
  4. AWS Lambda

Explanation

Amazon SageMaker provides an integrated environment for preparing data, training, tuning, and deploying models on managed infrastructure, including notebook instances. Lambda is serverless compute, CloudFront is a content delivery network, and Glue is a data integration service. SageMaker is the ML-specific platform among these options.

Question 15AI and ML concepts, technologies, and tools

In reinforcement learning, what is the role of the reward signal?

  1. It labels each training example with the correct class.
  2. It defines the objective the agent seeks to maximize through its actions. Correct answer.
  3. It normalizes input features before training.
  4. It selects which AWS region hosts the model.

Explanation

In reinforcement learning an agent takes actions in an environment and receives rewards that indicate how desirable those outcomes are, and it learns a policy to maximize cumulative reward. Reward is not a class label, a feature normalization step, or an infrastructure setting. This distinguishes RL from supervised and unsupervised paradigms.

Question 16AI and ML concepts, technologies, and tools

A team needs to detect sentiment in customer reviews at scale with minimal ML expertise. Which AWS service is designed for this?

  1. Amazon Comprehend Correct answer.
  2. Amazon SageMaker Ground Truth
  3. Amazon EMR
  4. AWS DataSync

Explanation

Amazon Comprehend is a pre-trained NLP service that provides sentiment analysis, entity detection, and related text insights through an API, requiring no model building. Ground Truth supports data labeling, EMR handles big data processing, and DataSync moves data between storage systems. Comprehend directly delivers sentiment detection.

Question 17AI and ML concepts, technologies, and tools

Which practice helps mitigate bias in a deployed ML model that produces unfair outcomes for a subgroup?

  1. Auditing training data and outcomes across subgroups and applying mitigation techniques Correct answer.
  2. Removing all documentation about model limitations
  3. Deploying the model to more regions without analysis
  4. Increasing the learning rate until outputs change

Explanation

Detecting and reducing unfair outcomes requires measuring performance across subgroups, examining training data for representation issues, and applying mitigation such as reweighting or resampling. Removing documentation hides the problem, and tuning the learning rate or expanding deployment does not address group-level disparity. Structured auditing is the responsible approach.

Question 18AI and ML concepts, technologies, and tools

What is the main benefit of using automatic model tuning (hyperparameter optimization) in SageMaker?

  1. It searches hyperparameter combinations to find a configuration that optimizes a chosen objective metric Correct answer.
  2. It automatically writes the business requirements document
  3. It guarantees the model will be fair across all subgroups
  4. It eliminates the need for any training data

Explanation

SageMaker automatic model tuning runs multiple training jobs over specified hyperparameter ranges and selects the configuration that optimizes an objective metric such as validation accuracy. It still requires training data and does not by itself guarantee fairness or produce business documentation. Its purpose is efficient search over hyperparameters.

Question 19AI and ML concepts, technologies, and tools

Which of the following best describes a foundation model?

  1. A rule-based system that requires no training data
  2. A model that can only perform image classification
  3. A large model pre-trained on broad data that can be adapted to many downstream tasks Correct answer.
  4. A small model trained only on tabular data for a single task

Explanation

Foundation models are large models pre-trained on broad, general data and then adapted through prompting or fine-tuning to many downstream tasks. They are not limited to tabular data, rule-based logic, or a single task such as image classification. This breadth of adaptation is their defining characteristic.

Question 20AI and ML concepts, technologies, and tools

A company wants to use Amazon Bedrock to build a generative AI application. What does Bedrock primarily provide?

  1. A content delivery network for static assets
  2. A relational database engine for transactional workloads
  3. Managed access to foundation models through an API for building generative AI applications Correct answer.
  4. A container registry for storing Docker images

Explanation

Amazon Bedrock offers managed access to foundation models from multiple providers via a unified API, enabling generative AI applications without managing model infrastructure. It is not a relational database, a container registry, or a CDN. Its role is to simplify building on foundation models.

Question 21AI and ML concepts, technologies, and tools

What distinguishes few-shot prompting from zero-shot prompting when using a large language model?

  1. Few-shot prompting only works with image models.
  2. Zero-shot prompting always produces more accurate results than few-shot prompting.
  3. Few-shot prompting requires retraining the model weights, while zero-shot does not.
  4. Few-shot prompting includes a small number of example input-output pairs in the prompt, while zero-shot provides no examples. Correct answer.

Explanation

Few-shot prompting supplies a handful of worked examples within the prompt to guide the model's output format and reasoning, whereas zero-shot prompting asks the model to perform the task with instructions only. Neither approach retrains weights, and few-shot often improves accuracy rather than being universally inferior. It also applies to language tasks, not only images.

Question 22AI and ML concepts, technologies, and tools

Which statement accurately describes the difference between batch inference and real-time inference?

  1. Batch inference processes large datasets on a schedule, while real-time inference responds to individual requests with low latency. Correct answer.
  2. Batch inference always has lower latency than real-time inference.
  3. Real-time inference cannot use models trained in SageMaker.
  4. Batch inference requires no stored data.

Explanation

Batch inference runs predictions over large stored datasets, often on a schedule, trading latency for throughput, while real-time inference serves individual requests through a persistent endpoint with low latency. Batch does not have lower latency, real-time endpoints do serve SageMaker-trained models, and batch inference inherently requires stored input data.

Question 23AI and ML concepts, technologies, and tools

Which AWS service would help a team label a large image dataset with human reviewers to create training data?

  1. Amazon Athena
  2. Amazon Macie
  3. Amazon SageMaker Ground Truth Correct answer.
  4. AWS Backup

Explanation

SageMaker Ground Truth provides managed data-labeling workflows that combine human annotators with automated labeling to build high-quality training datasets. Athena is a query service, Macie discovers sensitive data, and Backup handles data protection. Ground Truth is purpose-built for annotation tasks.

Question 24AI and ML concepts, technologies, and tools

A model's input feature distribution shifts substantially after deployment, degrading accuracy. What is this phenomenon called and what is a reasonable response?

  1. Data drift; monitor input distributions and retrain with recent data Correct answer.
  2. Target leakage; add post-outcome features
  3. Overfitting; remove the validation set
  4. Underfitting; reduce model capacity

Explanation

Data drift, or covariate shift, occurs when the distribution of inputs changes after deployment, which can degrade model accuracy even if the underlying relationship is unchanged. Monitoring input distributions and retraining on recent data are standard mitigations. Overfitting, leakage, and underfitting describe different failure modes that do not match a post-deployment input shift.

Question 25AI and ML concepts, technologies, and tools

Which statement best explains why a confusion matrix is more informative than overall accuracy for a multi-class classifier?

  1. It replaces the need for a test dataset.
  2. It computes the training loss for each epoch.
  3. It guarantees the model is free of bias.
  4. It shows per-class correct and incorrect predictions, revealing which classes are confused. Correct answer.

Explanation

A confusion matrix tabulates predicted versus actual classes, exposing per-class errors and systematic confusions that a single accuracy number can hide, especially under imbalance. It does not compute epoch losses, guarantee fairness, or remove the need for held-out evaluation data. The per-class breakdown is its key advantage.

Question 26AI and ML concepts, technologies, and tools

A team wants to train a model on tabular data using Amazon SageMaker without writing training code. Which SageMaker capability fits best?

  1. AWS Glue DataBrew recipe jobs
  2. Built-in algorithms accessed through the SageMaker Python SDK estimators Correct answer.
  3. Amazon Comprehend custom classification
  4. Amazon Personalize campaigns

Explanation

SageMaker ships built-in algorithm images that can be launched through estimators, so the team supplies data and hyperparameters rather than training code. The other services either perform NLP, data preparation, or recommendation workloads rather than general model training.

Question 27AI and ML concepts, technologies, and tools

Which statement best describes the purpose of a validation dataset during model development?

  1. It stores misclassified examples for later manual review
  2. It provides unbiased estimates of final generalization after all tuning is complete
  3. It is used to tune hyperparameters and compare model variants before final testing Correct answer.
  4. It replaces the training set when the dataset is too small to split

Explanation

A validation set guides model selection and hyperparameter choices because it is not used to fit weights. The test set is held back to estimate generalization once tuning is finished, keeping that estimate free of selection bias.

Question 28AI and ML concepts, technologies, and tools

A company needs to store and query feature values for both training and real-time inference with low-latency lookups. Which AWS capability is built for this?

  1. AWS Glue crawlers
  2. Amazon SageMaker Feature Store Correct answer.
  3. Amazon S3 Glacier Deep Archive
  4. Amazon EMR on EKS

Explanation

SageMaker Feature Store maintains online and offline stores so the same feature definitions serve training jobs and low-latency inference. The other options address archival storage, schema discovery, or big-data processing rather than serving features.

Question 29AI and ML concepts, technologies, and tools

A model trained to predict loan defaults performs well overall but poorly for a small demographic group. Which evaluation approach best exposes this gap?

  1. Increasing the number of training epochs until loss plateaus
  2. Computing accuracy on a random holdout split only
  3. Disaggregating metrics by subgroup and comparing recall and false positive rates Correct answer.
  4. Reporting a single aggregate AUC across the full dataset

Explanation

Aggregate metrics can hide subgroup failures because a large majority group dominates the average. Breaking metrics out by subgroup shows whether error rates differ materially across populations.

Question 30AI and ML concepts, technologies, and tools

Which technique reduces the dimensionality of a high-dimensional feature set while preserving as much variance as possible?

  1. Principal component analysis Correct answer.
  2. One-hot encoding
  3. SMOTE oversampling
  4. Gradient boosting

Explanation

PCA projects data onto orthogonal components ordered by explained variance, letting a team keep fewer dimensions with minimal information loss. The other options address encoding, class imbalance, or predictive modeling rather than dimensionality reduction.

Question 31AI and ML concepts, technologies, and tools

A fraud team faces a dataset where only 0.5% of transactions are fraudulent, and a model predicting 'never fraud' achieves 99.5% accuracy. Which metric should drive model selection?

  1. Mean squared error
  2. Overall accuracy
  3. Area under the precision-recall curve Correct answer.
  4. R-squared

Explanation

With extreme class imbalance, accuracy is dominated by the majority class and rewards trivial predictors. Precision-recall curves focus on performance on the rare positive class and are the standard choice for rare-event detection.

Question 32AI and ML concepts, technologies, and tools

What is the primary purpose of a learning rate scheduler during neural network training?

  1. To adjust the step size used for weight updates over the course of training Correct answer.
  2. To select which features enter the model
  3. To determine the number of layers in the network
  4. To split data into training and test partitions

Explanation

The learning rate controls the magnitude of each gradient step, and schedulers reduce or cycle it so training converges more stably. Feature selection, architecture depth, and data splitting are separate design decisions.

Question 33AI and ML concepts, technologies, and tools

A team wants an embedding representation of product descriptions to power semantic search. Which approach produces this directly?

  1. Applying TF-IDF and keeping only the top 100 terms
  2. Encoding text with a sentence-transformer model and storing the vectors in a vector index Correct answer.
  3. Running k-means clustering on raw character counts
  4. Sorting products alphabetically by title

Explanation

Sentence-transformer models map text into dense vectors whose distances reflect semantic similarity, which is what semantic search needs. The alternatives produce sparse counts, unsupervised clusters of characters, or no representation at all.

Question 34AI and ML concepts, technologies, and tools

Which statement best explains the purpose of cross-validation?

  1. It estimates performance more reliably by rotating which fold is held out Correct answer.
  2. It increases the size of the training data beyond what was collected
  3. It guarantees the model will generalize to production traffic
  4. It removes the need for a separate test set in every situation

Explanation

K-fold cross-validation trains and evaluates the model K times with different held-out folds, producing a mean and spread that are more stable than a single split. It still cannot promise production performance or create new data.

Question 35AI and ML concepts, technologies, and tools

A regression model shows low error on training data but high error on validation data. Which pair of remedies is most appropriate?

  1. Apply regularization and gather more training data Correct answer.
  2. Increase model complexity and train longer
  3. Lower the learning rate and add more epochs
  4. Add more features and remove regularization

Explanation

The described pattern is overfitting: the model memorizes training noise and fails to generalize. Regularization constrains model capacity, and additional data gives the learner more signal to generalize from.

Question 36AI and ML concepts, technologies, and tools

Which task is an example of unsupervised learning?

  1. Classifying emails as spam or not spam
  2. Predicting house prices from square footage
  3. Estimating customer churn probability
  4. Grouping news articles into topics without labels Correct answer.

Explanation

Topic grouping without predefined labels is clustering, which learns structure from unlabeled data. The other tasks all have target labels and are supervised problems.

Question 37AI and ML concepts, technologies, and tools

A team wants to detect anomalies in sensor readings when almost no labeled anomalies exist. Which approach is most suitable?

  1. Multiclass classification over sensor identifiers
  2. Unsupervised density estimation that flags readings far from the learned normal distribution Correct answer.
  3. Linear regression on a timestamp feature
  4. Supervised classification trained on the few labeled anomalies only

Explanation

When anomalies are rare and unlabeled, modeling the distribution of normal data and scoring distance from it avoids dependence on scarce labels. Training a classifier on a handful of positives usually overfits and cannot represent unseen anomaly types.

Question 38AI and ML concepts, technologies, and tools

What is the role of the tokenizer in a large language model pipeline?

  1. It schedules learning rate decay across epochs
  2. It stores model weights on disk between training runs
  3. It converts raw text into integer token IDs the model can process Correct answer.
  4. It computes the final loss for backpropagation

Explanation

A tokenizer splits text into subword units and maps them to vocabulary indices, which is the numeric input the model consumes. Weights, loss computation, and optimizer schedules belong to other components.

Question 39AI and ML concepts, technologies, and tools

A team fine-tunes a large language model on proprietary documents but the model starts repeating training text verbatim in responses. Which mitigation is most direct?

  1. Removing the tokenizer vocabulary
  2. Increasing the number of training epochs
  3. Raising the sampling temperature to zero
  4. Reducing memorization with deduplication, lower learning rate, and early stopping Correct answer.

Explanation

Verbatim regurgitation signals memorization of training examples, which deduplication, gentler updates, and stopping earlier all reduce. More epochs or greedy decoding make copying more likely rather than less.

Question 40AI and ML concepts, technologies, and tools

Which statement correctly distinguishes precision from recall?

  1. Precision measures the share of predicted positives that are truly positive; recall measures the share of actual positives that were found Correct answer.
  2. Precision applies only to regression and recall only to classification
  3. Precision measures the share of actual positives found; recall measures the share of predicted positives that are correct
  4. Precision and recall are identical for binary classifiers

Explanation

Precision is true positives divided by all predicted positives, so it penalizes false alarms. Recall is true positives divided by all actual positives, so it penalizes missed cases.

Question 41AI and ML concepts, technologies, and tools

A recommendation system shows users only items similar to what they already bought, and their exposure narrows over time. Which concept describes this?

  1. Feedback loop amplification Correct answer.
  2. Data leakage
  3. Transfer learning
  4. Gradient clipping

Explanation

When model outputs shape the data the model later learns from, small preferences get reinforced and diversity collapses, a self-reinforcing feedback loop. Transfer learning, leakage, and gradient clipping describe unrelated phenomena.

Question 42AI and ML concepts, technologies, and tools

Which AWS service provides a managed environment to run Jupyter notebooks and is commonly used to explore data before training on SageMaker?

  1. Amazon SageMaker Studio Correct answer.
  2. AWS Batch
  3. Amazon Athena
  4. AWS Step Functions

Explanation

SageMaker Studio is the integrated notebook-based IDE for data exploration, training, and deployment within SageMaker. The other services handle orchestration, SQL querying, or batch compute jobs.

Question 43AI and ML concepts, technologies, and tools

A data scientist wants to explain which input features drove a specific prediction from a trained model. Which approach answers this?

  1. Local feature attribution such as SHAP values for that instance Correct answer.
  2. Confusion matrix analysis on the test set
  3. Global feature importance averaged over the training set
  4. Hyperparameter sweep results

Explanation

Instance-level attribution methods like SHAP assign contributions to each feature for a single prediction. Global importance, confusion matrices, and tuning results describe aggregate behavior rather than one decision.

Question 44AI and ML concepts, technologies, and tools

Which statement best describes how a decision tree ensemble such as a random forest reduces variance compared with a single deep tree?

  1. It guarantees zero training error
  2. It uses one tree with a lower maximum depth
  3. It averages many decorrelated trees, so individual tree errors partially cancel Correct answer.
  4. It removes all features with low correlation to the target before fitting

Explanation

Random forests grow trees on bootstrapped samples with random feature subsets, decorrelating them so averaging reduces variance. A single shallow tree reduces variance by limiting depth but is not an ensemble, and no ensemble guarantees zero error.

Question 45AI and ML concepts, technologies, and tools

A team must serve a model that returns predictions within tens of milliseconds for an interactive web application. Which deployment pattern fits?

  1. Offline scoring in a data warehouse
  2. Manual notebook execution on demand
  3. Real-time endpoint with autoscaling Correct answer.
  4. Batch transform on a nightly schedule

Explanation

Real-time endpoints keep the model loaded and respond synchronously with low latency, and autoscaling absorbs traffic variation. Batch, warehouse, and notebook approaches add scheduling or startup delays incompatible with interactive use.

Question 46AI and ML concepts, technologies, and tools

Which practice most directly supports reproducibility of an ML experiment?

  1. Deleting intermediate artifacts after each run to save space
  2. Training only on the most recent data snapshot
  3. Relying on the default hyperparameters of the library
  4. Recording data versions, code commits, hyperparameters, and random seeds for each run Correct answer.

Explanation

Reproducibility requires capturing every input that affects a result, including data version, code state, hyperparameters, and seeds. Defaults, deleted artifacts, and moving data snapshots all make reruns unrepeatable.

Question 47AI and ML concepts, technologies, and tools

A team trains a model on data from one region and deploys it in another region with different customer behavior. Accuracy drops sharply. Which concept best explains the failure?

  1. Class imbalance in the training set
  2. Covariate shift between training and serving distributions Correct answer.
  3. Insufficient model interpretability
  4. Overfitting to the training labels

Explanation

The input distribution at serving time differs from training time, which is covariate shift, so learned relationships no longer map cleanly. Overfitting, imbalance, and interpretability concern different problems.

Question 48AI and ML concepts, technologies, and tools

Which statement best explains why a foundation model can be adapted to many downstream tasks?

  1. It only works on the exact data format used during pretraining
  2. It was pretrained on broad data and learned general representations that transfer with limited task-specific tuning Correct answer.
  3. It contains a separate hand-coded module for each downstream task
  4. It is retrained from scratch for every new task

Explanation

Broad pretraining yields general-purpose representations that can be adapted through prompting or light fine-tuning. The alternatives contradict the transferable nature of foundation models.

Question 49AI and ML concepts, technologies, and tools

A team wants to compare two models on a metric while accounting for uncertainty in a small test set. Which method is appropriate?

  1. Increase the learning rate and retrain
  2. Report a confidence interval via bootstrap resampling of the test set Correct answer.
  3. Choose the model with the larger file size
  4. Report only the single point estimate of the metric

Explanation

Bootstrapping resamples the test set to estimate the sampling distribution of a metric, yielding a confidence interval that shows whether a difference is meaningful. A bare point estimate hides that uncertainty, and the other options are irrelevant to evaluation.

Question 50AI and ML concepts, technologies, and tools

Which AWS service would a team use to orchestrate a multi-step ML workflow with conditional branches and retries across services?

  1. Amazon Kendra
  2. Amazon SageMaker Ground Truth
  3. AWS Step Functions Correct answer.
  4. AWS Glue DataBrew

Explanation

Step Functions coordinates distributed services as a state machine with branching, error handling, and retries, which suits ML pipelines. The other services address labeling, enterprise search, or data preparation.