Ngân hàng đề — AWS Certified Machine Learning Engineer Associate

SageMaker Model Monitor to track data distribution and performance.

SageMaker Autopilot for automatic tuning and optimization of the model’s hyperparameters.

SageMaker Pipelines to automate model retraining when bias is detected.

SageMaker Data Wrangler for pre-training data processing.

Use Glue crawlers to infer the schema, store the schema in Glue Data Catalog, and use Athena to run queries directly on the CSV files in S3.

Use Redshift Spectrum to read the CSV files from S3 and manually create a schema for Redshift to query the data.

Use Glue ETL to convert CSV files into JSON format, create a schema in Glue Data Catalog, and then import data into Redshift for querying.

Use Glue ETL to transform the CSV data into a Parquet format, store the schema in the Glue Data Catalog, and use Athena to query the data.

Drop the dominating features to balance the dataset

Label encoding for categorical features

Feature normalization using Min-Max scaling

Use recursive feature elimination to remove unimportant features

By leveraging Ground Truth’s human-in-the-loop feedback mechanism to review and correct model outputs for bias and vulnerabilities.

By using Ground Truth only for labeling data, leaving bias detection to other AWS services like SageMaker Clarify.

By exclusively relying on third-party vendors to detect bias and vulnerabilities in model outputs.

By using Ground Truth's fully automated labeling system to eliminate all bias from the training data.

A policy attached to a user that allows them to assume roles in different AWS accounts.

A policy that defines permissions to access AWS resources from external services but only for AWS-managed resources like Lambda functions.

A policy that defines all actions permitted for an AWS account’s root user.

A policy that specifies which identities (users, services, or accounts) are allowed to assume a role, commonly used for cross-account access.

By enabling the same pipeline step to be reused across different pipelines without redefinition.

By storing steps in Amazon S3, which can be manually imported into each pipeline.

By requiring each pipeline to be redefined from scratch for different models.

By allowing steps to be manually copied and pasted across different pipeline definitions.

A process where human laborers label all the data to ensure the highest accuracy.

A combination of machine learning and human feedback where human workers only intervene in complex tasks, improving labeling efficiency.

A method of labeling data that eliminates the need for human feedback entirely through automation.

A system where human labelers work in parallel with machines to label data simultaneously.

Use SageMaker Experiments to organize model runs into trials under a single experiment, where each trial logs configurations and metrics, enabling easy comparison of different models and hyperparameters.

Log every hyperparameter and configuration change manually to a spreadsheet and compare the metrics offline after each run.

Create a single trial with one configuration and manually adjust hyperparameters each time for new tests.

Use SageMaker Ground Truth to label datasets and organize each training run as a separate trial component under different experiments.

Use Amazon S3 to store training data, manually track hyperparameter configurations, and SageMaker Pipelines to handle training and deployment.

Use SageMaker Pipelines to automate the workflow and SageMaker Experiments to track and compare multiple configurations.

Use SageMaker Feature Store to manage model versions and SageMaker Autopilot to automatically deploy the best model based on accuracy.

Use SageMaker Ground Truth to create a labeled dataset, SageMaker Autopilot to automate model training, and SageMaker Feature Store to track hyperparameters and metrics.

SageMaker Model Monitor automatically manages and optimizes the model splitting process across multiple GPUs.

SageMaker’s Model Parallelism Library automatically splits the model across multiple GPUs or machines, based on the model architecture and available memory.

SageMaker’s built-in algorithms dynamically adjust the batch size to ensure the model can be trained on multiple machines.

SageMaker handles model parallelism by dividing the dataset across multiple machines and averaging the model gradients.