Amazon MLA-C01: Computer Vision, NLP and Specialized ML — Study Guide

By Laurynas Brazys · 3 August 2026 · #amazon #mla-c01 #certification #exam-prep #study-guide

Part of the AWS Machine Learning Engineer Associate MLA-C01 — Study Guide. Practice with verified answers in the Amazon exam hub, or take timed practice tests on ExamRoll.io.

Core concept

Computer vision and NLP solutions in AWS hinge on three layered concerns: data ingestion and security, feature and label preparation, and model lifecycle (training, registry, deployment, monitoring). For image classification and object detection, training workflows center on labeled asset storage in Amazon S3 with tight encryption and network controls, annotation formats such as Pascal VOC / COCO for object detection and RecordIO or TFRecord for high-throughput training. For text classification and named entity recognition (NER), inputs are commonly tokenized sequences (WordPiece/BPE for transformer models) or line-delimited JSON with labels; preprocessing must preserve token-to-character offsets when using Rekognition Textract outputs or human-labeled spans so that NER label alignment is consistent. Feature engineering for tabular fraud models focuses on time-window aggregation, categorical cardinality reduction, and consistent encoding between training and serving; using a centralized transformation (Feature Store or Data Wrangler flow) prevents skew between off-line training and online inference.

Model-building choices map to specialized AWS services: Amazon SageMaker offers built-in algorithms (XGBoost, Linear Learner, Random Cut Forest) and managed containers for frameworks (MXNet, TensorFlow, PyTorch) plus SageMaker Clarify for bias and explainability. Amazon Rekognition covers pre-built image APIs for labeling, face/celebrity detection, and custom labels training via Rekognition Custom Labels when low operational overhead is required. For text extraction and structured data from documents, Amazon Textract provides OCR and form/table extraction; for language-level tasks like sentiment, entity recognition or topic modeling, Amazon Comprehend provides managed APIs and Comprehend Medical for clinical NER. Critical to production is tying these pieces into a consistent pipeline so that preprocessing, model inputs, and monitoring are reproducible and auditable.

Key services and configuration

The core AWS services and relevant API/configuration parameters to know are Amazon SageMaker (TrainingJob, ModelPackage, ModelPackageGroup, ModelRegistry), SageMaker Pipelines (RegisterModel, CallbackStep, ConditionStep), SageMaker Model Monitor and Clarify (BaselineConfig, DataConfig, ModelConfig, bias_config), Amazon Rekognition, Amazon Comprehend, Amazon Textract, AWS Glue, and Lake Formation. For secure isolated storage, configure S3 with server-side encryption using SSE-KMS (S3.PutBucketEncryption with SSEKMSKeyId), attach bucket policies limiting access to the SageMaker execution role, and enable a Gateway VPC endpoint for S3 for private network transfers. When launching training jobs, set the VpcConfig parameter (SecurityGroupIds and Subnets) in CreateTrainingJob so instances run in the customer VPC, and enable NetworkIsolation and InstanceMetadataServiceConfiguration to limit access.

To centrally manage models with minimal operational overhead, use the SageMaker Model Registry by creating a ModelPackageGroup and adding ModelPackage versions with ModelApprovalStatus set to ‘PendingManualApproval’ via CreateModelPackage. Automate a manual-approval gate by adding a CallbackStep or a dedicated “ManualApproval” step in SageMaker Pipelines; the pipeline waits for an external signal, and you then call UpdateModelPackage to set ModelApprovalStatus=‘Approved’ for deployment. For on-demand bias or drift assessments of deployed real-time endpoints, use SageMaker Clarify for bias metrics and SageMaker Model Monitor for data & prediction drift: capture inference payloads by enabling DataCaptureConfig in CreateEndpoint (EnableCapture, CaptureOptions, DestinationS3Uri), and then start a Clarify processing job via CreateProcessingJob with the Clarify SDK parameters DataConfig, ModelConfig and bias_config to compute drift metrics immediately.

Relevant services include:

Amazon SageMaker (TrainingJob, CreateModelPackage, UpdateModelPackage, CreateProcessingJob, CreateMonitoringSchedule)
SageMaker Pipelines (RegisterModel, CallbackStep, ConditionStep)
SageMaker Clarify (DataConfig, ModelConfig, bias_config, explainability_config)
Amazon Rekognition (StartProjectVersion, DetectLabels, CreateProjectVersion)
Amazon Comprehend (DetectEntities, StartEntitiesDetectionJob)
Amazon Textract (StartDocumentTextDetection, AnalyzeDocument)
AWS Glue and AWS Lake Formation (Crawlers, Jobs, Data Catalog)
Amazon Lookout for Metrics and Amazon QuickSight for anomaly visualization

Design patterns and trade-offs

For tabular classification like fraud detection, the practical design pattern is to centralize raw sources into a secure S3 data lake using AWS Glue or DMS for relational tables, catalog them in Glue Data Catalog, and enforce access via Lake Formation. Transformations are expressed once in SageMaker Data Wrangler flows or as Glue ETL jobs and persisted in SageMaker Feature Store so the same transformations run at training and at inference time. Choosing the Feature Store adds operational steps up front but reduces skew and debugging time; direct in-job transformations are lower initial overhead but make reproducibility and online feature computation harder. For training algorithms, XGBoost (SageMaker-provided container) is a strong default for fraud detection because of its tabular performance and support for weight columns; configure hyperparameters using HyperParameters in CreateTrainingJob and pass a weight column or set scale_pos_weight to counter class imbalance, which avoids more complex resampling pipelines.

For computer vision, if you need rapid iteration and labeled data is limited, Rekognition Custom Labels provides the fastest path with minimal ops; for maximum control and custom architectures, use SageMaker training with MXNet/PyTorch, and store datasets in RecordIO or ImageFolder on S3. For object detection, prefer training with frameworks that output COCO format and leverage SageMaker distributed training by specifying InstanceType=ml.p3.2xlarge or higher and setting the MPI or horovod configuration in the TrainingJob’s AlgorithmSpecification and TrainingInputMode. Trade-offs include throughput versus cost: batch jobs using managed spot instances reduce cost but add latency/termination risk; pipeline caching in SageMaker Pipelines reduces repeated step startup when inputs are unchanged, which minimizes unnecessary compute start time.

Common pitfalls and decision criteria

A frequent pitfall is failing to centralize preprocessing artifacts. If tokenization, label mapping for NER, or categorical encoders are applied differently in training and inference, you introduce prediction errors that are hard to trace. Use Feature Store or persist the preprocessing artifact (tokenizer, vocab.json, label_map) with the model package in the Model Registry so the deployed container retrieves and applies exact transforms. Another common failure is insufficient capture for monitoring: without enabling DataCaptureConfig on the endpoint and a proper ground-truth labeling workflow, Model Monitor cannot compute drift or quality metrics and clarifying drops in F1 becomes guesswork. For class imbalance, the least operationally intensive solution is to use algorithm-supported weighting (for example XGBoost’s scale_pos_weight hyperparameter or providing a weights column in training data) rather than blind upsampling/downsampling, which can introduce sampling bias.

Practical Problem: Use-Case Scenario

Named company: MeridianPay. MeridianPay must build a real-time fraud detection pipeline combining transaction logs in S3 and customer profiles in an on-prem MySQL database. Requirements include secure isolated storage, a central model registry with manual approval, minimal startup latency across consecutive training runs, automated anomaly detection and visualization after aggregation, support for mixed categorical and numerical features with minimal operations, handling class imbalance, and a production model that can be monitored for drift and bias on demand.

Aggregate and secure data: use AWS DMS to continuously replicate on-prem MySQL tables into an encrypted S3 landing zone, run AWS Glue crawlers, and register resulting tables in AWS Glue Data Catalog. Enforce access via AWS Lake Formation and S3 bucket policies; enable a Gateway VPC Endpoint for S3 and configure SageMaker execution role with least-privilege IAM. Use SSE-KMS on S3 with a customer-managed KMS key and set BucketEncryption with the KmsMasterKeyId.
Transform and store features: author transformations in SageMaker Data Wrangler or a Glue ETL job, persist derived features in Amazon SageMaker Feature Store online store for low-latency lookup at inference and offline store for training. Store tokenizer/encoder artifacts alongside model artifacts. Use FeatureGroup APIs to create feature groups and configure RecordIdentifierFeatureName and EventTimeFeatureName to enable point-in-time correctness.
Train with minimal operational overhead: use the SageMaker XGBoost built-in container by creating a CreateTrainingJob with AlgorithmSpecification pointing to the XGBoost container URI, specify VpcConfig to run in the VPC, pass HyperParameters including “objective”:“binary:logistic” and set “scale_pos_weight” to the ratio of negative to positive examples to address class imbalance. To reduce repeated startup latency, implement SageMaker Pipelines and enable caching for the data-prep steps so that consecutive pipeline runs skip unchanged steps; use persistent feature store lookups rather than reprocessing whenever possible.
Model registry and manual approval: register trained models into a ModelPackageGroup via CreateModelPackage with ModelApprovalStatus=‘PendingManualApproval’. Integrate a SageMaker Pipelines CallbackStep that pauses after RegisterModel; reviewers then call UpdateModelPackage to set ModelApprovalStatus=‘Approved’. Use this approved package for CreateModel and CreateEndpoint configurations.
Anomaly detection and visualization: after aggregation, use Amazon Lookout for Metrics to perform automatic anomaly detection on time-series transaction aggregates and configure integrations to S3/Glue. For visualization, connect Lookout results and data in QuickSight dashboards or use SageMaker Studio notebooks to visualize feature drift. For on-demand model bias/drift assessment of deployed endpoints, enable DataCaptureConfig in CreateEndpoint and run an on-demand SageMaker Clarify processing job (CreateProcessingJob) with DataConfig and ModelConfig pointing to captured data; use Model Monitor/CreateMonitoringSchedule to schedule recurring checks and to trigger a one-off run with StartMonitoringSchedule.

Rationale: this approach centralizes data and transformations, uses managed services where they reduce operational burden (Glue/DMS/Lookout/Comprehend/Textract for their domains), leverages SageMaker Model Registry and Pipelines for versioning and manual approval with minimal custom infra, resolves class imbalance through algorithm parameters to avoid heavy resampling, and provides both scheduled and on-demand bias/drift assessments via Model Monitor and Clarify while keeping data encrypted and network-isolated.

← Cost Optimization for ML Workloads · All domains

Practice these questions → · Timed practice on ExamRoll.io →