AWS Machine Learning Specialty Practice

Athena is a serverless query service for data in S3. EC2 provides compute, Route 53 is DNS, and IAM manages access.

AWS Glue is an ETL service that prepares and catalogs data. S3 stores data, Route 53 resolves DNS, and CloudFront delivers content.

Feature stores centralize features for reuse across training and inference. Transaction databases, DNS, and CDNs are different.

S3 is object storage commonly used for datasets and model artifacts. DynamoDB and RDS are databases, and Route 53 is DNS.

Data leakage occurs when test information reaches training, causing overoptimistic evaluation. Storage, deployment, and compression are not leakage.

Normalization scales features to comparable ranges. Removing rows, encryption, and DNS are different operations.

Categorical variables represent categories such as color or country. Numeric measurements, indexes, and metrics are different.

A separate test set evaluates generalization on unseen data. Splitting is not about storage cost, security, or avoiding S3.

Data Wrangler provides a visual interface for data preparation. DNS, firewalls, and load balancers are different.

S3 stores raw data, Glue prepares and catalogs, feature stores share features, and Data Wrangler provides visual preparation.

Linear Learner supports regression and classification. K-Means clusters, PCA reduces dimensions, and Object Detection works on images.

XGBoost is widely used for tabular classification and regression. K-Means clusters, PCA reduces dimensions, and BlazingText handles text.

Hyperparameters are set before training, while model parameters are learned from data. Test records and artifacts are different.

An epoch is one full pass over the training dataset. Tables, deployments, and security groups are different.

Overfitting means high training accuracy but poor performance on new data. Simple models, deployment, and encryption are different.

Early stopping ends training when validation metrics no longer improve, reducing overfitting. It is not deletion, encryption, or restart.

Transfer learning reuses knowledge from a pretrained model. Copying, moving endpoints, and changing instances are different.

A confusion matrix summarizes classification results. Loss, cost, and latency are different metrics.

Loss functions measure how far predictions are from targets. Buckets, firewall rules, and DNS records are not metrics.

Linear Learner supports regression, XGBoost handles classification, an epoch is a training pass, and overfitting means poor generalization.

IAM controls users, roles, and permissions. S3 stores data, CloudFront delivers content, and Glue prepares data.

KMS manages encryption keys. Route 53 is DNS, WAF protects web traffic, and CloudWatch monitors metrics.

A VPC is an isolated virtual network. Databases, models, and load balancers run inside or alongside it.

VPC endpoints provide private connectivity to AWS services. They do not improve accuracy, store data, or create reports.

TLS encrypts data while it travels. Buckets, DNS records, and alarms do not encrypt traffic.

Notebook instances assume IAM roles to access resources. DNS records, endpoints, and alarms do not grant permissions.

CloudTrail logs API activity for auditing. Athena queries, Glue prepares data, and SageMaker builds models.

Least privilege limits permissions to what is necessary. Full access, encryption, and VPC creation are not least privilege.

Private endpoints are reachable only within a VPC. Public websites, model artifacts, and alarms are different.

IAM grants permissions, KMS manages keys, VPC provides network isolation, and CloudTrail audits API calls.

SageMaker endpoints host models and serve real-time predictions. Datasets, DNS records, and security groups are different.

Batch transform processes a dataset and writes predictions. It is not real-time serving, training, or encryption.

Inference pipelines chain preprocessing and prediction containers. Buckets, training jobs, and DNS zones are different.

Autoscaling changes endpoint capacity to match traffic. It does not improve accuracy, encrypt data, or create reports.

Endpoint configurations define model, instance type, and capacity. Training data, DNS, and subnets are configured elsewhere.

SageMaker supports real-time, serverless, and batch inference. It does not resolve DNS or store data instead of S3.

Model artifacts are trained model files. Hyperparameters are settings, test data is input, and alarms are monitoring tools.

Production variants group model and instance settings for routing traffic. They are not datasets, VPC endpoints, or DNS records.

SageMaker routes traffic percentages to variants for A/B testing. Deleting models, single instances, or disabling monitoring are not A/B testing.

Real-time endpoints serve live requests, batch transform processes batches, pipelines chain containers, and autoscaling adjusts capacity.

CloudWatch collects metrics and alarms. Athena queries, Glue prepares data, and Route 53 resolves DNS.

Model Monitor detects drift and quality issues in inputs and predictions. DNS, cost, and latency are other concerns.

Data drift occurs when production data distribution changes from training data. Deleting models, stopping endpoints, and VPC changes are different.

SageMaker Experiments tracks runs, parameters, and metrics. Data warehouses, DNS, and load balancers are different.

SageMaker Pipelines automates ML workflow steps. Storage, alarms, and security groups are different.

Lineage records how artifacts and models relate. Encryption, deletion, and VPC creation are not lineage.

Clarify detects bias and explains predictions. Route 53, Glue, and CloudFront are not bias detection tools.

Alarms monitor metrics and trigger actions. Buckets, training jobs, and DNS records are different.

Retraining with current data addresses drift. Deleting endpoints, disabling monitoring, or removing roles would not fix the model.

CloudWatch provides metrics, Model Monitor detects drift, Experiments track runs, and Pipelines automate workflows.

Managed Spot Training uses Spot instances for lower-cost training. On-demand only, latency increases, or disabled encryption are not Spot training.

Distributed training spreads workload across instances for faster training. One CPU, deleting data, or one model are not distributed training.

Autopilot automates model building and tuning. DNS, firewalls, and load balancers are different.

Automatic model tuning explores hyperparameters to improve metrics. Deletion, encryption, and reports are different.

Instance types define compute capacity. Models, datasets, and DNS records are not instance types.

Right-sizing, serverless, and batch inference reduce cost. Excess instances, disabled monitoring, or more storage increase cost.

Quotas limit resource usage. Metrics, datasets, and DNS records are different.

Cost optimization balances spend and performance. Buying everything, disabling security, or deleting models are not optimization.

Compression reduces model size for faster, cheaper inference. Encryption, increasing size, or adding instances are not compression.

Spot training lowers cost, distributed training speeds up, Autopilot builds models, and auto tuning searches hyperparameters.