Google Cloud Professional Machine Learning Engineer - Fact Sheet¶

Quick Reference¶

Exam Code: Professional Machine Learning Engineer Duration: 120 minutes (2 hours) Questions: 50-60 questions Passing Score: ~70% (not officially published) Cost: $200 USD Validity: 2 years Difficulty: ⭐⭐⭐⭐⭐ (Advanced-level certification) Prerequisites: Recommended 3+ years ML/data science experience, including 1+ year with GCP ML tools Exam Format: Multiple choice and multiple select questions based on real-world scenarios

Exam Domains¶

ML Problem Framing¶

Defining ML Use Cases¶

Business to ML Translation: - Identify when ML is appropriate vs rule-based systems - Define clear success criteria and KPIs - Understand constraints (latency, cost, accuracy requirements) - 📖 ML Problem Framing - Google's problem framing course - 📖 ML Use Cases - Common ML application patterns - 📖 Choosing ML Models - Model selection guide

ML Task Types¶

Supervised Learning: - Classification - Binary, multi-class, multi-label - Regression - Continuous value prediction - 📖 Classification Models - Classification on Agent Platform (formerly Vertex AI) - 📖 Regression Models - Regression on Agent Platform

Unsupervised Learning: - Clustering - Grouping similar data points - Dimensionality Reduction - PCA, t-SNE, autoencoders - Anomaly Detection - Outlier identification - 📖 Clustering with AutoML - Clustering solutions

Other ML Tasks: - Recommendation Systems - Collaborative filtering, content-based - Time Series Forecasting - Temporal pattern prediction - Computer Vision - Image classification, object detection, segmentation - Natural Language Processing - Text classification, entity extraction, sentiment analysis - 📖 Recommendation AI - Recommendation systems - 📖 Time Series Forecasting - Forecasting solutions

Success Metrics¶

Classification Metrics: - Accuracy - Overall correctness - Precision - True positives / (True positives + False positives) - Recall/Sensitivity - True positives / (True positives + False negatives) - F1 Score - Harmonic mean of precision and recall - AUC-ROC - Area under ROC curve - AUC-PR - Area under precision-recall curve - 📖 Classification Metrics - Understanding classification metrics

Regression Metrics: - MAE (Mean Absolute Error) - Average absolute differences - RMSE (Root Mean Squared Error) - Square root of mean squared errors - R² (R-Squared) - Proportion of variance explained - MAPE (Mean Absolute Percentage Error) - Percentage error - 📖 Regression Metrics - Understanding regression metrics

Ranking Metrics: - NDCG (Normalized Discounted Cumulative Gain) - Ranking quality - MAP (Mean Average Precision) - Precision across multiple queries - MRR (Mean Reciprocal Rank) - Average of reciprocal ranks

Agent Platform¶

Agent Platform Overview¶

Unified ML Platform: - End-to-end ML workflow management - Integrates training, deployment, and monitoring - Supports custom and AutoML models - 📖 Agent Platform Overview - Platform introduction - 📖 Agent Platform Workbench - Managed Jupyter notebooks - 📖 Agent Platform Pipelines - ML workflow orchestration - 📖 Agent Platform Model Registry - Model versioning and management

Agent Platform Workbench¶

Development Environment: - Managed Notebooks - Fully managed JupyterLab instances - User-Managed Notebooks - More control, custom configurations - Pre-installed ML frameworks (TensorFlow, PyTorch, scikit-learn) - Integration with BigQuery, Cloud Storage, and other GCP services - 📖 Workbench User Guide - Notebook management - 📖 Workbench Executor - Schedule notebook execution - 📖 Workbench Instances - Instance creation and configuration

Agent Platform Managed Training¶

Training Options: - AutoML - No-code model training - Custom Training - Pre-built containers or custom containers - Hyperparameter Tuning - Automated hyperparameter optimization - Distributed Training - Multi-node, multi-GPU training - 📖 Training Overview - Training options comparison - 📖 Custom Training - Custom training jobs - 📖 Distributed Training - Multi-worker training - 📖 Hyperparameter Tuning - Optimization strategies - 📖 Training with GPUs - GPU acceleration - 📖 Training with TPUs - TPU acceleration

Agent Platform Inference¶

Model Deployment: - Online Prediction - Real-time, low-latency predictions - Batch Prediction - Offline, high-throughput predictions - 📖 Prediction Overview - Deployment options - 📖 Deploy Models - Model deployment guide - 📖 Online Prediction - Real-time serving - 📖 Batch Prediction - Batch inference - 📖 Private Endpoints - VPC-based serving - 📖 Model Monitoring - Production model monitoring

Model Monitoring on Agent Platform¶

Production Monitoring: - Training-Serving Skew Detection - Distribution drift from training - Prediction Drift Detection - Changes in prediction distribution - Feature Attribution - Input feature importance - 📖 Model Monitoring Setup - Monitoring configuration - 📖 Skew Detection - Training-serving skew - 📖 Drift Detection - Prediction drift monitoring - 📖 Alerting - Monitoring alerts setup

AutoML¶

AutoML Tables¶

Structured Data ML: - Automated feature engineering - Automated model selection and ensemble - No-code model training - Supports regression, classification, forecasting - 📖 AutoML Tabular - Structured data solutions - 📖 Training AutoML Models - AutoML training guide - 📖 Feature Engineering - Automatic feature transformations - 📖 Model Interpretability - Understanding AutoML predictions

AutoML Vision¶

Computer Vision: - Image Classification - Single-label and multi-label - Object Detection - Bounding box detection - Image Segmentation - Pixel-level classification - 📖 AutoML Vision - Vision capabilities - 📖 Image Classification - Classification tutorial - 📖 Object Detection - Detection tutorial - 📖 Edge Deployment - Deploy to edge devices

AutoML Natural Language¶

NLP Tasks: - Text Classification - Sentiment analysis, topic classification - Entity Extraction - Named entity recognition - Text Entity Extraction - Custom entity types - 📖 AutoML Natural Language - NLP solutions - 📖 Text Classification - Classification guide - 📖 Entity Extraction - NER guide - 📖 Sentiment Analysis - Sentiment classification

AutoML Video Intelligence¶

Video Analysis: - Video Classification - Video-level labels - Video Object Tracking - Track objects across frames - Action Recognition - Temporal action detection - 📖 AutoML Video Intelligence - Video analysis - 📖 Video Classification - Classification tutorial - 📖 Object Tracking - Tracking guide

Data Engineering for ML¶

Data Ingestion¶

Data Sources: - BigQuery - Data warehouse for structured data - Cloud Storage - Object storage for files - Pub/Sub - Real-time streaming data - Cloud SQL/Spanner - Relational databases - Bigtable - NoSQL for time-series data - 📖 Data Ingestion Patterns - Best practices - 📖 Dataflow - Stream and batch processing - 📖 Data Fusion - Visual ETL tool - 📖 Transfer Service - Data migration

Data Preprocessing¶

Data Cleaning: - Handle missing values (imputation, deletion) - Outlier detection and treatment - Data validation and quality checks - 📖 TensorFlow Data Validation - TFDV for data validation - 📖 Data Preprocessing - Preprocessing strategies

Data Transformation: - Normalization and standardization - Encoding categorical variables (one-hot, label, target encoding) - Feature scaling - Text preprocessing (tokenization, stemming, lemmatization) - 📖 TensorFlow Transform - TFX Transform for preprocessing - 📖 Feature Engineering - Engineering best practices - 📖 Dataflow Templates - Pre-built preprocessing pipelines

Feature Engineering¶

Feature Creation: - Polynomial features - Interaction features - Time-based features (day of week, hour, seasonality) - Domain-specific feature engineering - 📖 Feature Engineering Best Practices - Feature representation - 📖 Feature Crosses - Combining features

Feature Selection: - Filter methods (correlation, mutual information) - Wrapper methods (forward/backward selection) - Embedded methods (L1/L2 regularization, tree-based importance) - 📖 Feature Selection Techniques - Selection strategies

Agent Platform Feature Store¶

Centralized Feature Management: - Store, serve, and share ML features - Online and offline serving - Point-in-time correctness for training - Feature versioning and lineage - 📖 Feature Store Overview - Feature store architecture - 📖 Creating Features - Feature store setup - 📖 Serving Features - Online feature serving - 📖 Feature Store Best Practices - Usage patterns

BigQuery for ML¶

BigQuery ML: - Create ML models using SQL - Supports linear regression, logistic regression, k-means, time series, deep learning - AutoML integration for advanced models - 📖 BigQuery ML Overview - BQML introduction - 📖 Creating Models - Model creation - 📖 Model Types - Available algorithms - 📖 Hyperparameter Tuning - BQML tuning - 📖 Model Evaluation - Evaluation functions - 📖 Predictions - Making predictions - 📖 Export Models - Export for serving

Data Validation¶

TensorFlow Data Validation (TFDV): - Generate statistics from training data - Infer schema and detect anomalies - Detect drift and skew - 📖 TFDV Guide - Data validation guide - 📖 Schema Generation - Automatic schema inference - 📖 Anomaly Detection - Data quality checks

Data Splitting¶

Train/Validation/Test Split: - Typical split: 70% train, 15% validation, 15% test - Time-based splitting for time series - Stratified splitting for imbalanced data - Cross-validation for small datasets - 📖 Data Splitting Best Practices - Splitting strategies

TensorFlow and Deep Learning¶

TensorFlow Ecosystem¶

Core Components: - TensorFlow 2.x - Eager execution, Keras integration - Keras - High-level neural network API - TensorFlow Datasets - Pre-processed datasets - TensorFlow Hub - Reusable model components - 📖 TensorFlow Guide - Complete TensorFlow documentation - 📖 Keras Guide - Keras API guide - 📖 TensorFlow Datasets - Dataset catalog - 📖 TensorFlow Hub - Pre-trained models

TensorFlow Extended (TFX)¶

Production ML Pipelines: - ExampleGen - Data ingestion - StatisticsGen - Statistics generation - SchemaGen - Schema inference - ExampleValidator - Anomaly detection - Transform - Feature engineering - Trainer - Model training - Tuner - Hyperparameter tuning - Evaluator - Model evaluation - Pusher - Model deployment - 📖 TFX Overview - TFX architecture - 📖 TFX Components - Component documentation - 📖 TFX on Cloud - TFX with Agent Platform - 📖 TFX Pipeline Tutorial - End-to-end example

Neural Network Architectures¶

Common Architectures: - Feedforward Neural Networks - Basic dense networks - Convolutional Neural Networks (CNNs) - Computer vision - Recurrent Neural Networks (RNNs) - Sequential data - Long Short-Term Memory (LSTM) - Long sequences - Transformers - NLP and beyond - 📖 Neural Network Guide - NN fundamentals - 📖 CNN Architectures - CNN tutorial - 📖 RNN Guide - RNN with Keras - 📖 Transformer Models - Transformer tutorial

Transfer Learning¶

Pre-trained Models: - Use pre-trained models from TensorFlow Hub - Fine-tune on domain-specific data - Feature extraction vs fine-tuning - 📖 Transfer Learning Guide - Transfer learning tutorial - 📖 TensorFlow Hub Integration - Using pre-trained models - 📖 Fine-tuning Strategies - Fine-tuning best practices

Training Optimization¶

Optimization Techniques: - Gradient descent variants (SGD, Adam, RMSprop) - Learning rate scheduling - Batch normalization - Dropout for regularization - Early stopping - 📖 Optimization Guide - Keras optimizers - 📖 Learning Rate Schedules - LR scheduling - 📖 Regularization - Preventing overfitting - 📖 Callbacks - Training callbacks

Distributed Training Strategies¶

TensorFlow Distribution Strategies: - MirroredStrategy - Single machine, multiple GPUs - MultiWorkerMirroredStrategy - Multiple machines - TPUStrategy - Cloud TPU training - ParameterServerStrategy - Large-scale distributed training - 📖 Distributed Training - Distribution strategies - 📖 Multi-GPU Training - Multi-GPU guide - 📖 TPU Training - TPU usage guide - 📖 Parameter Server - Parameter server strategy

Model Building and Training¶

Classical ML Algorithms¶

Scikit-learn Integration: - Linear/Logistic Regression - Decision Trees and Random Forests - Gradient Boosting (XGBoost, LightGBM) - Support Vector Machines - K-Nearest Neighbors - 📖 Scikit-learn on Agent Platform - Pre-built containers - 📖 Custom Prediction Routines - Deploy scikit-learn models

Hyperparameter Tuning¶

Tuning Strategies: - Grid search - Random search - Bayesian optimization - Hyperband algorithm - 📖 Vertex AI Hyperparameter Tuning - Tuning service - 📖 Tuning Algorithms - Algorithm comparison - 📖 Parallel Trials - Parallel tuning - 📖 Early Stopping - Stop underperforming trials

Model Evaluation¶

Evaluation Techniques: - Hold-out validation - K-fold cross-validation - Stratified K-fold - Time series cross-validation - 📖 Model Evaluation - Evaluation strategies - 📖 What-If Tool - Interactive evaluation - 📖 TensorBoard Integration - Visualization

Handling Imbalanced Data¶

Techniques: - Class weighting - Oversampling (SMOTE) - Undersampling - Ensemble methods - Anomaly detection approaches - 📖 Imbalanced Data Guide - Handling imbalance - 📖 Class Weights - TensorFlow class weights

Model Interpretability¶

Explainability Techniques: - Feature importance - SHAP (SHapley Additive exPlanations) - LIME (Local Interpretable Model-agnostic Explanations) - Integrated Gradients - 📖 Vertex Explainable AI - Explainability overview - 📖 Feature Attributions - Attribution methods - 📖 Explanation Methods - Configuration guide - 📖 What-If Tool - Interactive explanations

ML Pipeline Automation and MLOps¶

Kubeflow Pipelines¶

ML Workflow Orchestration: - Define, orchestrate, and schedule ML workflows - Component-based pipeline architecture - Integration with Agent Platform Pipelines - 📖 Kubeflow Pipelines - KFP documentation - 📖 Agent Platform Pipelines - Managed Kubeflow - 📖 Pipeline Components - Building components - 📖 Pre-built Components - Google Cloud components - 📖 Pipeline Scheduling - Scheduled execution

MLOps Best Practices¶

CI/CD for ML: - Version control for code, data, and models - Automated testing (unit tests, integration tests) - Continuous training (CT) - Continuous deployment (CD) - 📖 MLOps Architecture - MLOps patterns - 📖 MLOps Level 0-2 - Maturity levels - 📖 Cloud Build for ML - CI/CD with Cloud Build

Model Versioning¶

Version Management: - Model Registry for version tracking - A/B testing between model versions - Canary deployments - Shadow mode deployment - 📖 Model Registry - Version management - 📖 Model Versioning - Version control - 📖 Model Aliases - Version aliases

Workflow Orchestration¶

Orchestration Options: - Agent Platform Pipelines - Managed Kubeflow Pipelines - Cloud Composer - Managed Apache Airflow - Cloud Scheduler - Cron-based scheduling - 📖 Cloud Composer - Airflow for ML workflows - 📖 Composer for ML - ML orchestration patterns - 📖 Cloud Scheduler - Job scheduling

Continuous Training¶

Automated Retraining: - Trigger retraining based on data drift - Schedule-based retraining - Performance degradation triggers - New data availability triggers - 📖 Continuous Training - CT architecture - 📖 Retraining Pipelines - Pipeline creation

Model Deployment Strategies¶

Deployment Patterns: - Blue/Green Deployment - Switch between two environments - Canary Deployment - Gradual rollout to subset - A/B Testing - Compare model versions - Shadow Mode - Run new model without serving - 📖 Deployment Strategies - Deployment patterns - 📖 Traffic Split - Split traffic between models - 📖 Private Endpoints - Secure deployment

Container Management¶

Containerization: - Docker containers for reproducibility - Pre-built training and prediction containers - Custom containers for specific frameworks - 📖 Container Registry - Image storage - 📖 Artifact Registry - Next-gen artifact storage - 📖 Pre-built Containers - Available containers - 📖 Custom Containers - Building custom containers

Model Monitoring and Optimization¶

Production Model Monitoring¶

Monitoring Components: - Prediction request/response logging - Input feature distribution monitoring - Prediction output distribution monitoring - Model performance metrics - 📖 Model Monitoring on Agent Platform - Monitoring overview - 📖 Monitoring Configuration - Setup guide - 📖 Monitoring Alerts - Alert configuration - 📖 Cloud Logging Integration - Prediction logging

Data Drift Detection¶

Drift Types: - Covariate Shift - Input distribution changes - Concept Drift - Relationship between X and Y changes - Label Drift - Output distribution changes - 📖 Drift Detection - Drift monitoring - 📖 Skew Detection - Training-serving skew - 📖 Statistical Distance - Distance metrics

Model Performance Optimization¶

Optimization Strategies: - Model compression (pruning, quantization) - Knowledge distillation - Hardware acceleration (GPUs, TPUs) - Caching predictions - Batch predictions for throughput - 📖 Model Optimization - TensorFlow optimization - 📖 Quantization - Model quantization - 📖 Pruning - Weight pruning

Prediction Latency Optimization¶

Latency Reduction: - Model simplification - Feature reduction - Prediction caching - Hardware acceleration - Edge deployment - 📖 Prediction Performance - Performance tuning - 📖 GPU Prediction - GPU serving - 📖 Batch Prediction - High-throughput inference

Cost Optimization¶

Cost Reduction Strategies: - Batch prediction instead of online - Auto-scaling endpoints - Spot VMs for training - Model compression - Feature Store for feature reuse - 📖 Training Cost Optimization - Training cost reduction - 📖 Prediction Cost Optimization - Serving cost reduction - 📖 Preemptible VMs - Spot instance training

A/B Testing¶

Experimentation: - Traffic splitting between model versions - Statistical significance testing - Multi-armed bandit approaches - 📖 Traffic Splitting - A/B testing setup - 📖 Model Comparison - Compare model versions

Responsible AI¶

Fairness and Bias¶

Bias Detection: - Identify protected attributes - Measure disparate impact - Demographic parity - Equalized odds - 📖 Responsible AI - Google's principles - 📖 Fairness Indicators - Bias detection tools - 📖 What-If Tool - Interactive fairness analysis - 📖 Fairness Best Practices - Implementing fairness

Explainability¶

Model Transparency: - Vertex AI Explainable AI - Feature attribution methods - Counterfactual explanations - Model cards for documentation - 📖 Explainable AI - Explainability overview - 📖 Sampled Shapley - Shapley values - 📖 Integrated Gradients - IG method - 📖 XRAI - Image explainability

Privacy¶

Privacy-Preserving ML: - Differential privacy - Federated learning - Secure multi-party computation - Data anonymization and pseudonymization - 📖 Privacy Best Practices - Privacy techniques - 📖 Data Loss Prevention - Sensitive data protection - 📖 Confidential Computing - Encrypted computation

Model Cards¶

Documentation: - Model description and use cases - Training data details - Performance metrics - Limitations and biases - Ethical considerations - 📖 Model Cards - Model documentation framework - 📖 Model Card Toolkit - Automated model cards

AI Ethics¶

Ethical Considerations: - Transparency and accountability - Privacy and security - Inclusiveness and fairness - Reliability and safety - 📖 AI Principles - Google's AI principles - 📖 PAIR - People + AI Research

Pre-trained AI Services¶

Vision AI¶

Computer Vision APIs: - Cloud Vision API for image analysis - AutoML Vision for custom models - 📖 Cloud Vision API - Pre-trained vision models - 📖 Vision API Features - Available features - 📖 Object Detection - Object localization

Natural Language AI¶

NLP APIs: - Natural Language API for text analysis - AutoML Natural Language for custom models - Translation API for language translation - 📖 Natural Language API - Text analysis - 📖 Entity Recognition - NER API - 📖 Sentiment Analysis - Sentiment API - 📖 Translation API - Language translation

Speech AI¶

Speech APIs: - Speech-to-Text API - Text-to-Speech API - 📖 Speech-to-Text - Audio transcription - 📖 Text-to-Speech - Speech synthesis - 📖 Speech Adaptation - Custom vocabulary

Video Intelligence AI¶

Video Analysis: - Video Intelligence API for video analysis - AutoML Video for custom models - 📖 Video Intelligence API - Video analysis - 📖 Video Features - Available features

Document AI¶

Document Processing: - Document AI for document understanding - Form parsing and extraction - 📖 Document AI - Document processing - 📖 Document Processors - Available processors - 📖 Form Parser - Form extraction

Integration with GCP Services¶

BigQuery Integration¶

ML with BigQuery: - BigQuery ML for SQL-based ML - Export training data from BigQuery - Store predictions in BigQuery - 📖 BigQuery ML - BQML overview - 📖 Agent Platform and BigQuery - Integration guide - 📖 Export to BigQuery - Batch predictions to BQ

Cloud Storage Integration¶

Data Management: - Store training data in Cloud Storage - Model artifacts storage - Batch prediction input/output - 📖 Cloud Storage - Object storage - 📖 Storage for ML - Storage best practices

Pub/Sub Integration¶

Real-time ML: - Stream predictions with Pub/Sub - Trigger retraining on new data - Event-driven ML workflows - 📖 Pub/Sub - Messaging service - 📖 Streaming Predictions - Real-time inference

Dataflow Integration¶

Data Processing: - Preprocess data at scale - Feature engineering pipelines - TensorFlow Transform integration - 📖 Dataflow - Apache Beam pipelines - 📖 Dataflow for ML - ML data processing

Cloud Functions Integration¶

Serverless ML: - Lightweight prediction endpoints - Trigger pipelines on events - Custom prediction logic - 📖 Cloud Functions - Serverless functions - 📖 Functions for ML - Serverless serving

Security and Governance¶

IAM for ML¶

Access Control: - Agent Platform-specific IAM roles - Service accounts for training/serving - Workload Identity for GKE - 📖 Vertex AI IAM - Access control - 📖 IAM Roles - Available roles - 📖 Service Accounts - Custom service accounts

VPC and Networking¶

Network Security: - VPC peering for private access - Private endpoints for serving - VPC Service Controls - 📖 VPC Peering - Private connectivity - 📖 Private Endpoints - Secure serving - 📖 VPC Service Controls - Data perimeter

Encryption¶

Data Protection: - Encryption at rest and in transit - Customer-managed encryption keys (CMEK) - 📖 Encryption - CMEK for Agent Platform - 📖 Key Management - Cloud KMS

Audit Logging¶

Compliance: - Cloud Audit Logs for ML operations - Model lineage tracking - Data lineage tracking - 📖 Audit Logs - Agent Platform audit logs - 📖 Model Lineage - ML Metadata

Exam Scenarios and Solutions¶

Scenario 1: Real-time Fraud Detection¶

Requirements: Detect fraudulent transactions in real-time, < 100ms latency, high accuracy

Solution: - Train model using AutoML Tables or custom TensorFlow - Deploy to Agent Platform Inference with GPU-enabled machine - Use Feature Store for real-time feature serving - Monitor for data drift with Model Monitoring on Agent Platform - A/B test new models before full rollout

Key Decision: Online prediction for low latency, Feature Store for fast feature retrieval

Scenario 2: Image Classification at Scale¶

Requirements: Classify millions of images daily, custom categories, cost-effective

Solution: - Start with AutoML Vision for quick baseline - If more control needed, use TensorFlow with transfer learning - Train on Agent Platform with GPU/TPU acceleration - Deploy to Agent Platform for batch prediction - Store results in BigQuery for analysis

Key Decision: Batch prediction for cost efficiency, AutoML for faster development

Scenario 3: NLP Text Classification¶

Requirements: Classify customer support tickets, multi-class classification, explainable

Solution: - Use AutoML Natural Language for rapid development - Enable Explainable AI for feature attributions - Deploy to online prediction endpoint - Implement Model Monitoring for drift detection - Store predictions in BigQuery for analytics

Key Decision: AutoML NLP for quick start, Explainable AI for transparency

Scenario 4: Time Series Forecasting¶

Requirements: Forecast demand for inventory planning, weekly forecasts, seasonal patterns

Solution: - Use AutoML Forecasting on Agent Platform - Prepare data with seasonal and holiday features - Train with multiple time series (one per product) - Evaluate using MAPE and RMSE - Deploy for batch predictions, schedule weekly

Key Decision: AutoML Forecasting for automatic seasonality handling

Scenario 5: Continuous Model Retraining¶

Requirements: Retrain models weekly with new data, automated pipeline, minimal manual intervention

Solution: - Build Vertex AI Pipeline with Kubeflow components - Include data validation (TFDV), training, evaluation, deployment - Schedule pipeline with Cloud Scheduler - Implement automatic deployment if metrics exceed threshold - Monitor with Model Monitoring on Agent Platform

Key Decision: Agent Platform Pipelines for orchestration, automated deployment gates

Scenario 6: Model Serving with Traffic Split¶

Requirements: Deploy new model version safely, gradual rollout, rollback capability

Solution: - Deploy new model version to Agent Platform - Configure traffic split (90% old, 10% new) - Monitor both versions for performance - Gradually increase traffic to new version - Keep old version for quick rollback

Key Decision: Traffic split for canary deployment, monitoring for safe rollout

Scenario 7: Privacy-Preserving ML¶

Requirements: Train on sensitive data, comply with GDPR, minimize data exposure

Solution: - Use differential privacy techniques - Implement federated learning if applicable - Use Data Loss Prevention API to detect PII - Anonymize/pseudonymize data before training - Store only aggregated results

Key Decision: DLP for PII detection, differential privacy for training

Scenario 8: Bias Detection in Models¶

Requirements: Ensure fairness across demographic groups, detect and mitigate bias

Solution: - Use Fairness Indicators to measure bias - Analyze model performance across subgroups - Implement re-weighting or re-sampling strategies - Use What-If Tool for interactive analysis - Document findings in Model Card

Key Decision: Fairness Indicators for measurement, What-If Tool for exploration

Scenario 9: Hybrid ML Architecture¶

Requirements: Train on-premises, deploy to cloud, data cannot leave premises

Solution: - Train models on-premises with TensorFlow - Export trained models - Deploy to Agent Platform Inference - Use VPN/Interconnect for secure connectivity - Serve predictions via private endpoint

Key Decision: Export trained models, private endpoints for secure serving

Scenario 10: High-Throughput Batch Predictions¶

Requirements: Score millions of records daily, cost-effective, results to BigQuery

Solution: - Use Agent Platform Batch Inference - Source data from BigQuery or Cloud Storage - Configure large batch size for efficiency - Output predictions directly to BigQuery - Schedule daily with Cloud Scheduler

Key Decision: Batch prediction for high throughput, BigQuery output for analytics

Exam Tips and Strategy¶

Keywords to Watch¶

Question Patterns: - "Real-time/low latency" → Online prediction, Feature Store, caching - "Batch/high throughput" → Batch prediction, BigQuery ML - "Cost-effective" → AutoML, batch processing, preemptible VMs, BigQuery ML - "No code/quick start" → AutoML solutions - "Custom model" → TensorFlow, PyTorch, custom training - "Explainability" → Vertex Explainable AI, What-If Tool - "Fairness/bias" → Fairness Indicators, What-If Tool - "Automated pipeline" → Agent Platform Pipelines, TFX, Kubeflow - "Data drift" → Model Monitoring, TFDV - "Scalable" → Distributed training, Agent Platform, managed services

Service Selection Decision Trees¶

ML Service Decision:

Need quick solution?
├─ YES → Have labeled data?
│  ├─ YES → AutoML (Tables/Vision/NLP)
│  └─ NO → Pre-trained APIs
└─ NO → Custom model needed?
   ├─ YES → TensorFlow/PyTorch on Vertex AI
   └─ NO → BigQuery ML (for tabular data)

Deployment Decision:

Latency requirements?
├─ < 1 second → Online prediction endpoint
├─ Minutes/hours → Batch prediction
└─ SQL-based → BigQuery ML prediction functions

Pipeline Orchestration:

Need orchestration?
├─ ML-specific → Vertex AI Pipelines (Kubeflow)
├─ General workflow → Cloud Composer (Airflow)
└─ Simple scheduling → Cloud Scheduler

Time Management¶

• 120 minutes ÷ 50 questions = 2.4 minutes per question
• First pass: Answer confident questions (60 minutes)
• Second pass: Tackle scenario-based questions (45 minutes)
• Final pass: Review flagged questions (15 minutes)

Common Traps¶

• ❌ Choosing custom solutions when AutoML would suffice
• ❌ Selecting online prediction when batch would be more cost-effective
• ❌ Ignoring data drift and model monitoring
• ❌ Over-complicating deployment (start simple)
• ❌ Not considering Responsible AI requirements
• ❌ Forgetting about BigQuery ML for SQL-based workflows
• ❌ Choosing wrong metric for the problem type
• ❌ Not implementing proper train/validation/test splits

Study Checklist¶

Core ML Knowledge: - [ ] Understand supervised vs unsupervised learning - [ ] Know when to use classification vs regression - [ ] Understand evaluation metrics for different problem types - [ ] Know how to handle imbalanced data - [ ] Understand overfitting and regularization techniques - [ ] Know feature engineering best practices - [ ] Understand cross-validation and data splitting

Agent Platform: - [ ] Can create and manage Agent Platform Workbench instances - [ ] Know how to train custom models on Agent Platform - [ ] Understand AutoML capabilities and limitations - [ ] Can deploy models to online and batch prediction - [ ] Know how to configure Model Monitoring - [ ] Understand Feature Store architecture and use cases - [ ] Can build and run Agent Platform Pipelines

TensorFlow and Deep Learning: - [ ] Understand TensorFlow 2.x and Keras APIs - [ ] Know TFX components and pipeline architecture - [ ] Can implement transfer learning - [ ] Understand distributed training strategies - [ ] Know optimization techniques (learning rate, regularization) - [ ] Can use TensorFlow Hub for pre-trained models

MLOps: - [ ] Understand CI/CD for ML pipelines - [ ] Know model versioning strategies - [ ] Can implement A/B testing for models - [ ] Understand retraining triggers and strategies - [ ] Know deployment patterns (canary, blue/green) - [ ] Can orchestrate workflows with Kubeflow/Composer

Data Engineering: - [ ] Know data ingestion options (BigQuery, Cloud Storage, Pub/Sub) - [ ] Understand data preprocessing techniques - [ ] Can use BigQuery for ML workloads - [ ] Know Dataflow for data processing - [ ] Understand data validation with TFDV

Responsible AI: - [ ] Can detect and mitigate bias in models - [ ] Know explainability techniques (SHAP, LIME, IG) - [ ] Understand fairness metrics - [ ] Can implement privacy-preserving ML - [ ] Know how to create Model Cards

Preparation: - [ ] Hands-on experience with Agent Platform (build at least 3 projects) - [ ] Complete official Google Cloud Skills Boost labs - [ ] Practice with sample case studies - [ ] Review TensorFlow and Keras documentation - [ ] Build end-to-end ML pipelines - [ ] Study official exam guide thoroughly

Pro Tip: The Professional ML Engineer exam emphasizes practical scenarios over theoretical knowledge. Focus on understanding when to use which tool/service, how to build production ML systems, and MLOps best practices. Always consider the full ML lifecycle: data → training → deployment → monitoring → retraining.

Documentation Count: This fact sheet contains 120+ embedded documentation links to official Google Cloud and TensorFlow documentation.

Good luck! This certification demonstrates expert-level machine learning engineering skills on Google Cloud Platform.