Free MLA-C01 practice test: 30 questions with answers

JumpStart provides pretrained models and solution templates that can be fine-tuned or deployed with little code. Edge Manager is for devices, Kinesis Video handles streams, and Batch runs general compute jobs.

SageMaker Autopilot automatically explores preprocessing and model candidates for tabular data and produces a ranked set of models with notebooks. Neo compiles models for hardware, Serverless Inference hosts models, and Ground Truth Plus manages labelling.

Automatic model tuning runs many training jobs over a hyperparameter range and optimises an objective metric. Clarify analyses bias, Model Registry tracks versions, and Edge Manager manages devices.

Distributed training splits work across multiple GPUs or instances, using data parallelism or model parallelism, to cut wall-clock time. Shrinking the data or removing accelerators does not solve the scaling problem, and skipping validation hides overfitting.

Regularization, early stopping, and more representative data reduce the gap between training and validation performance. Removing regularization, raising learning rates, and evaluating on training data all worsen or hide overfitting.

Model Registry groups model versions with metadata, metrics, and approval status, which supports governed promotion to production. The other services store data or metrics but do not manage model approval workflows.

A held-out test set that was excluded from training gives an unbiased estimate of performance. Training-run metrics and early-epoch values are optimistic or incomplete.

Quantization reduces numerical precision and pruning removes weights, both shrinking the model, and compilation adapts it to the target hardware. Adding layers increases size, and the other options do not reduce model footprint.

Many small objects create overhead per read, and partitioned columnar formats like Parquet let jobs read only the needed columns and partitions. Compression alone or more instances do not fix the small-file problem.

Data Wrangler provides a visual interface to import, transform, and analyse tabular data with previews, and it exports the flow to a processing job or pipeline. Model Monitor watches deployed models, Ground Truth labels data, and Clarify analyses bias and explainability.

SageMaker Feature Store keeps an online store for low-latency inference and an offline store for training, so both use the same feature definitions. Athena views, Edge Manager, and crawlers do not serve features to inference endpoints.

Ground Truth manages labelling jobs with workforces, annotation tooling, and consensus or quality controls. Rekognition trains models rather than providing labelling workforces, and Comprehend performs text analysis.

Fitting preprocessing on training data only prevents test statistics from leaking into training, and a time-based split keeps the future out of training for time-ordered data. Fitting on the full dataset, using the target as an input, and random splits of time series each leak information.

The AWS Glue Data Catalog stores table definitions and schemas, and crawlers populate it from S3 data. Athena and EMR use the catalog to query the same datasets. The other services do not provide that metadata layer.

Resampling or class weights help the model learn the minority class, and metrics such as precision, recall, and F1 expose poor performance on it. Accuracy can look high while the model misses every positive case, and deleting data discards information.

Detecting and redacting or tokenizing PII before training keeps sensitive values out of the model and its artifacts. Public access, larger instances, and disabled encryption each increase exposure or do nothing for privacy.

Comparing live inputs with the training baseline reveals data drift, and tracking quality against labelled outcomes reveals performance drift. Ignoring inputs, disabling logs, and retraining only once remove the signals needed to notice drift.

Model Monitor captures endpoint data, compares it with a baseline, and raises alerts on data quality and drift. Ground Truth labels data, JumpStart provides models, and Neo compiles models.

Registering versions with an approval gate controls what can be deployed, and CloudTrail records the API activity that makes the change. The other approaches allow unreviewed or untracked deployments.

Running in a VPC with gateway and interface endpoints keeps S3 and API traffic on the AWS network. A public NAT route or a public bucket sends the traffic over the internet, and disabling networking stops the job from reading data.

CloudTrail logs management API calls with the identity, time, and parameters, which supports auditing endpoint changes. Macie, Budgets, and Polly do not record SageMaker API activity.

ModelLatency and invocation metrics show whether the model itself is slower under the new load, which guides scaling decisions. The other metrics do not describe endpoint behaviour.

Managed Spot Training uses spare capacity at a discount, and checkpoints let the job resume after interruptions. On-demand without checkpoints pays full price and loses progress, while dedicated hosts and idle notebooks increase cost.

Batch transform processes large datasets offline from S3 and writes results back, with no endpoint to keep running. A real-time endpoint or WebSocket API is designed for low-latency requests, and notebooks are for interactive work.

Auto scaling adjusts the number of instances to match invocation load, which balances latency and cost across the day. Fixed peak capacity wastes money at night, and manual or notebook-based serving does not meet production needs.

SageMaker Pipelines defines ML workflows as connected steps, including conditional steps that gate deployment on metrics, and it records each execution. Clarify analyses bias, Edge Manager manages devices, and DataBrew prepares data.

Edge Manager packages models for edge devices and manages their deployment and monitoring. Serverless Inference runs in the cloud, Feature Store serves features, and Athena queries data.

Serverless Inference provisions capacity on demand and scales to zero between bursts, so idle periods cost nothing. A dedicated always-on endpoint pays for idle time, and manual or colocated servers add operational work.

A canary sends a small share of traffic to the new version, so problems affect few users and rollback is quick. Cutting over all at once, deleting the old endpoint first, or testing only in a notebook each removes that safety margin.

Batch transform reads inputs from S3, runs inference on the whole dataset, and writes outputs back, without keeping an endpoint running. The other options add cost or rely on manual steps.