Скачать или смотреть AWS Machine Learning Associate Exam Walkthrough 106 Q&A 61 to 80

AWS Machine Learning Associate Exam Walkthrough 106 Q&A 61-80 - October 14
VIEW RECORDING: https://fathom.video/share/2XdeJq2WAW...
Meeting Purpose

Review and explain AWS Machine Learning Associate Exam questions 61-80, focusing on key concepts and correct answers.

Key Takeaways

SageMaker Clarify is crucial for model explainability and regulatory compliance in ML deployments
Distributed training and proper instance placement significantly improve training performance for large datasets
Understanding data drift and consistent normalization is vital for maintaining model performance in production
EventBridge offers minimal operational overhead

Topics

Model Explainability and Compliance

SageMaker Clarify is the go-to solution for model explainability
Provides feature importance, prediction-level explanations, and bias detection
Critical for industries like financial services where explainable AI is mandatory

Addressing Class Imbalance

Class weights are preferred for severe imbalances (95% non-defective, 5% defective)
Preserves all original data
Formula: weight = 1 / (num_classes * class_frequency)

Secure Training with Sensitive Data

AWS Nitro Enclaves provide isolated compute environments
Ensures data remains inaccessible in plain text, even to AWS personnel
Ideal for healthcare and other sensitive data applications

Cost Optimization for ML Training

SageMaker Savings Plan with 1-year term and upfront payment offers discounts
Best for predictable workloads (35 hours/week for 55 weeks)
More cost-effective than on-demand or spot instances for regular, scheduled jobs

Efficient Data Formats for Image Training

Augmented Manifest format optimized for SageMaker image training
Supports efficient data loading without conversion
JSON structure includes image references and labels

Evaluation Metrics for Fraud Detection

Recall is the priority metric for fraud detection models
Focuses on minimizing false negatives (undetected fraud)
Formula: Recall = True Positives / (True Positives + False Negatives)

High Availability ML Deployments

Cross-region replication with multi-region endpoints ensures true high availability
Use Route 53 health checks for automatic failover between regions
Protects against regional failures and provides lowest latency

Optimizing Training for Long Text Sequences

Distributed training across multiple instances is preferred for long sequences
Parallelizes computation without truncating data
SageMaker supports distributed training out-of-the-box

Real-time Anomaly Detection for IoT Data

Kinesis Data Streams + Lambda + SageMaker Endpoint combination ideal for variable-rate streaming data
Handles high-throughput and sudden spikes with automatic scaling
Provides end-to-end real-time processing and low-latency inference

Gradual Model Deployment Strategies

Multi-variant endpoints support hosting multiple model versions with weighted traffic distribution
Enables canary deployments (90% old, 10% new) with real-time performance monitoring
Supports automatic rollback capabilities

ML Workflow Orchestration

SageMaker Pipelines is purpose-built for end-to-end ML workflow orchestration
Integrates data validation, training, evaluation, and conditional deployment
Supports MLOps best practices with built-in steps and model registry integration

Collaborative Filtering for Recommendations

Factorization Machines algorithm excels with sparse, high-dimensional data
Efficiently captures feature interactions for millions of users and items
Built-in SageMaker algorithm, effective for implicit feedback data

Handling Data Drift in Production

Data drift occurs when production data statistics differ from training data
Common cause of degraded model performance in production
Requires model retraining on updated data distribution

Normalization in Production Inference

Reuse the same min-max normalization statistics from training in production
Maintains consistent feature scaling between training and inference
Prevents distribution shift and preserves learned feature representations

Accessing Large Training Datasets

Mount FSx for NetApp ONTAP file system as a volume to SageMaker
Enables direct access to large datasets (6TB) without data copying
Provides low-latency, high-throughput access within the same VPC

Efficient ML Pipeline Triggering

Use EventBridge rules with S3 event patterns to trigger ML pipelines
Provides native integration with S3 events and direct pipeline invocation
Minimal operational overhead

Addressing Model Overfitting

Reduce max_depth hyperparameter in XGBoost to prevent overfitting
Creates simpler, less complex trees that generalize better to unseen data
Improves performance on new transactions in fraud detection scenarios