Free MLA-C02 practice test: 30 questions with answers

SageMaker Autopilot automatically explores candidate pipelines, ranks them on a leaderboard, and can generate notebooks that explain the process. Edge Manager manages models on devices, Pipelines orchestrates workflows, and Serverless Inference hosts models.

Automatic model tuning supports Bayesian optimization, which uses results from earlier trials to choose promising hyperparameters. A sequential manual grid is inefficient, sampling data does not tune hyperparameters, and instance size does not search the hyperparameter space.

Fine-tuning a pretrained model from JumpStart is transfer learning, which reuses learned representations and needs less labeled data. Feature hashing, clustering and batch inference are different techniques.

Rising validation loss with falling training loss indicates overfitting, and early stopping halts training when validation performance stops improving. Raising the learning rate, removing validation data, and training longer all make the problem worse or hide it.

K-fold cross-validation rotates the validation fold across k partitions and averages the results, which gives a more stable performance estimate on small data. The other techniques transform features or create data rather than estimating performance.

Precision-recall AUC and F1 both focus on the positive class and so remain informative when positives are rare. Accuracy alone can look high by predicting the majority class, and mean squared error is a regression metric.

SageMaker Debugger captures tensors during training and provides built-in rules, including vanishing gradients, that can be checked without code changes. Clarify analyzes bias and explainability, Ground Truth labels data, and Feature Store manages features.

The built-in XGBoost algorithm is a gradient-boosted tree method that performs well on tabular data and needs no custom code. Convolutional networks are designed for images, reinforcement learning requires an environment and reward, and Monte Carlo simulation is not a predictive model.

SageMaker Data Wrangler provides a visual, no-code interface for data preparation, including deduplication and missing-value handling, and it can export the flow to a processing job. Model Monitor watches deployed models, Ground Truth manages labeling, and Feature Store stores features.

AWS Glue provides a serverless data catalog through crawlers and serverless Spark ETL jobs. Snowball Edge is for offline data transfer, Kinesis Video Streams ingests video, and Cloud9 is an IDE.

SMOTE creates synthetic minority-class samples by interpolating between existing minority examples. PCA reduces dimensionality, one-hot encoding represents categories, and min-max scaling normalizes ranges, none of which address imbalance.

Feature Store offers an online store for low-latency lookups during inference and an offline store in S3 for training datasets, keeping the features consistent. Clarify analyzes bias, Model Registry tracks model versions, and Neo compiles models.

Columnar, compressed formats such as Parquet combined with partitioning let Athena read only the columns and partitions a query needs, which cuts scan costs. Uncompressed JSON scans the most, archive storage is not queryable with standard Athena, and workgroups do not change scanned bytes.

A private workforce keeps labeling within authorized individuals in the organization, which is appropriate for confidential data. Public and external workforces expose the data to people outside the organization's control.

SSE-KMS encrypts objects at rest under a key the organization controls, and a bucket policy that denies non-TLS requests protects data in transit. Public ACLs expose the data, and disabling versioning removes recovery options without improving protection.

Mode imputation fills missing categorical values with the most frequent category and keeps the column categorical. A mean of encoded integers has no meaning for categories, dropping rows discards data unnecessarily, and a constant that duplicates a real category introduces a misleading signal.

Asynchronous inference queues requests, supports large payloads and long processing times, and scales to zero when idle. Real-time endpoints need a persistent low-latency response, serverless inference has tighter payload and duration limits, and batch transform processes an entire dataset offline.

SageMaker Pipelines includes built-in step types such as TrainingStep and ProcessingStep, among others for evaluation, conditions and model registration. EmailStep and PrintStep are not pipeline step types.

Model Registry stores model versions with metadata and an approval status, and deployment pipelines can require approved models before deploying. The other components do not manage model approval states.

SageMaker blue/green deployment policies shift traffic in canary or linear steps and can roll back when CloudWatch alarms trigger. Deleting and recreating causes downtime, changing S3 artifacts alone does not redirect the endpoint, and batch transform does not serve live traffic.

Multi-model endpoints share instances across many models that are loaded on demand, and serverless inference avoids paying for idle capacity on infrequently used models. Oversized dedicated endpoints and always-on GPUs increase cost, and the inference recommender is correct as a sizing aid but the question asks for the two cost-sharing features.

A multi-model endpoint hosts many models behind one endpoint, loading each from S3 on demand, which suits rarely used models. Batch transform processes datasets offline, one-model-per-instance endpoints are costly, and edge packages run on devices.

SageMaker Neo compiles trained models into optimized code for specific hardware targets, including edge devices. Clarify explains bias, Model Monitor watches production data, and Ground Truth labels data.

Model Monitor compares live data with a baseline created from training data and raises alerts on violations such as drift, through scheduled monitoring jobs. The other options do not monitor production data quality.

Training jobs accept a volume KMS key identifier that encrypts the attached storage volumes with the specified customer managed key. Bucket policies do not encrypt attached volumes, lifecycle rules delete data, and instance size does not change encryption.

SageMaker publishes ModelLatency to CloudWatch, and an alarm with an SNS action notifies responders automatically. Manual checks and cost reports are not timely, and disabling logs removes visibility.

A scoped execution role grants only the specific S3 actions and resource prefix plus the CloudWatch Logs actions required. Administrator access, root credentials and public buckets all violate least privilege.

An S3 gateway endpoint lets resources in private subnets reach S3 over the AWS network through the route table, without an internet gateway or NAT. NAT routes traffic to the internet, a VPN goes to the corporate network, and disabling access blocks the service.

Model quality monitoring compares captured predictions with ground truth labels that arrive later, which tracks accuracy over time. Data quality monitoring does not use labels, and notebook logs do not measure model accuracy.

SageMaker Clarify computes SHAP-based feature attributions for individual predictions and for models overall. Edge Manager, the Autopilot leaderboard and the Feature Store do not produce per-prediction explanations.