PMLE StudyProfessional Machine Learning Engineer

Learning goals

• Choose between BigQuery ML and AutoML based on the business problem
• Meet requirements with the least effort by choosing among pretrained APIs, foundation models and Model Garden
• Optimize cost, latency and availability of Gemini apps

Key points

• BigQuery ML builds models with SQL (CREATE MODEL, ML.PREDICT, ML.EVALUATE) without moving data: ARIMA_PLUS (time series), K-means (clustering), logistic regression and boosted trees, remote models (Gemini) and more.
• The TRANSFORM clause applies preprocessing consistently at training and prediction. Check features with ML.FEATURE_IMPORTANCE and the like.
• AutoML builds high-quality models without code from labeled tabular, image, text and video data.
• First check whether pretrained APIs (Vision, Document AI, Translation, Speech) or foundation models suffice; if not, AutoML; custom training last.
• Optimizing Gemini cost and latency: choosing lighter models (such as Flash), context caching, batch inference, limiting output tokens, provisioned throughput.

Workplace uses

• AccountingExtract invoice fields with Document AI, and build demand forecasts with BigQuery ML's ARIMA_PLUS using only SQL.

Exam objectives (from the exam guide)

• BigQuery ML・Agent Platform AutoML でのモデルの作成（分類・回帰・予測・クラスタリング）、BigQuery ML での特徴量エンジニアリングと予測、BigQuery での Gemini のファインチューニング
• Model Garden からのモデルの評価と選択、業界別の API（Document AI・Vision・Translation）、Gemini・Imagen・Veo・MaaS のチューニング、Gemini のアプリのコスト・レイテンシ・可用性の最適化

Learning resources

• Professional ML Engineer exam guide (English, PDF) ↗
• PMLE exam guide / study guide ↗

Learning goals

• Choose preprocessing tools that fit the scale and complexity of the data
• Choose and configure secure, collaborative notebook environments
• Track and compare experiments, evaluations and lineage

Key points

• Small data: in-memory pandas. Large SQL transforms: BigQuery. Large streaming and batch transforms: Dataflow (Apache Beam). Existing Spark assets: Dataproc (Serverless for Apache Spark).
• Feature Store centrally manages features, uses the same values in training and serving, and serves them online with low latency.
• Detect, mask and tokenize PII with Sensitive Data Protection (formerly Cloud DLP).
• Workbench is a managed Jupyter-based notebook; Colab Enterprise is a managed, collaborative Colab. Protect them with service accounts, VPC Service Controls, CMEK and more.
• Experiments records and compares parameters and metrics; ML Metadata tracks artifact lineage (which data and code produced which model).
• Evaluating generative AI: automatic metrics (BLEU, ROUGE), LLM-as-a-judge (another LLM scores against criteria), and human evaluation. Use the Gen AI evaluation service.

Workplace uses

• Data analyticsThe analytics team prototypes in Colab Enterprise, compares runs with Experiments, and moves the best settings into a pipeline.

Exam objectives (from the exam guide)

• データの探索と前処理（表・テキスト・画像、BigQuery・Dataflow・Spark・メモリ内の Python の使い分け、Feature Store、PII の保護）
• ノートブックでのプロトタイピング（Agent Platform Workbench、Colab Enterprise、セキュリティと共同作業、PyTorch・scikit-learn・JAX、Model Garden のモデル）
• 実験の追跡（Experiments、Pipelines、Kubeflow Pipelines、予測と生成 AI の評価・LLM-as-a-judge、ML Metadata による系統の追跡）

Learning resources

• Professional ML Engineer exam guide (English, PDF) ↗
• PMLE exam guide / study guide ↗

Learning goals

• Choose model types and products that fit the requirements
• Choose training methods and isolate the causes of failures
• Choose hardware and distributed training strategies

Key points

• Time series: ARIMA_PLUS and the like. Complex nonlinear patterns: DNNs. Language and generation tasks: LLMs.
• If interpretability is required, use linear models or decision trees, or Explainable AI (feature attributions).
• Custom training runs any framework in containers (prebuilt or custom). Tune hyperparameters with Vizier.
• Countering overfitting: regularization, dropout, early stopping, more data. If training loss won't fall, check the learning rate and data.
• TPUs excel at large matrix operations (DNNs and LLMs with large batches). GPUs are flexible. CPUs suit small models and preprocessing.
• Data parallelism: each worker runs the same model on different data. Model parallelism: splits a model too large for one device.
• Consider tuning foundation models when prompts or RAG aren't enough (format, style, specialized tasks): supervised fine-tuning, and parameter-efficient methods such as LoRA.

Workplace uses

• ManufacturingPrototype image defect detection with AutoML; since accuracy falls short, move to custom training on GPUs and tune the learning rate with Vizier.

Exam objectives (from the exam guide)

• コスト・複雑さ・レイテンシ・スケーラビリティを考えたモデルの構築（ARIMA・DNN・LLM、AutoML・BigQuery ML・Pipelines、デプロイの戦略、解釈可能性の要件）
• モデルの学習（Cloud Storage・BigQuery のデータの整理、取り込み、カスタム学習・GKE の Kubeflow・AutoML・Tabular Workflows、学習の失敗の対処、ハイパーパラメータのチューニング、基盤モデルのファインチューニング）
• 学習のハードウェア（CPU・GPU・TPU、データ並列・モデル並列の分散学習）

Learning resources

• Professional ML Engineer exam guide (English, PDF) ↗
• PMLE exam guide / study guide ↗

Learning goals

• Choose the inference style and serving platform that fit the requirements
• Design model version management and safe rollouts
• Scale serving to match throughput and latency

Key points

• Batch inference processes large data together (no latency need, cheap); online inference is real-time and low-latency.
• Prebuilt containers support TensorFlow, PyTorch, XGBoost and scikit-learn; use custom containers for others or custom processing.
• Manage model versions and aliases in Model Registry.
• Use traffic splitting for A/B tests and canaries (sending a little traffic to the new version first).
• Private endpoints (Private Service Connect, etc.) allow low-latency calls only from inside the VPC.
• Autoscaling (min and max replicas), GPU machine types, and lightening models with quantization or distillation. Edge means on-device inference (offline, low latency).

Workplace uses

• E-commerceServe product recommendations from an online endpoint, fetch the latest browsing features from Feature Store, and canary a new model on 10% of traffic.

Exam objectives (from the exam guide)

• バッチ推論とオンライン推論（Agent Platform・Model Garden・Cloud Run・GKE）、事前構築済み・カスタムのコンテナ、Model Registry、A/B テスト・カナリア、推論の前処理・後処理
• オンライン配信のスケーリング（Feature Store、パブリック・プライベートのエンドポイント、CPU・GPU・TPU・エッジ、スループットに応じたスケーリング、本番向けのチューニング）

Learning resources

• Professional ML Engineer exam guide (English, PDF) ↗
• PMLE exam guide / study guide ↗

Learning goals

• Choose a pipeline orchestration method
• Design to prevent training-serving skew
• Design retraining policies and CI/CD/CT

Key points

• Agent Platform Pipelines (formerly Vertex AI Pipelines) runs Kubeflow Pipelines and TFX definitions serverlessly and records artifacts and lineage.
• Managed Service for Apache Airflow (formerly Cloud Composer) orchestrates broad workflows including non-ML work. Ray handles distributed Python processing.
• Build data validation (schema and distribution checks) and model validation (meeting baseline metrics) into the pipeline.
• Use the same preprocessing code for training and serving (TRANSFORM clause, tf.Transform, Feature Store) to prevent skew.
• Retraining policies: triggered on a schedule, by volume of new data, or by detecting performance drops or drift.
• CI tests code, CD deploys models, CT retrains continuously. Automate builds and tests with Cloud Build.

Workplace uses

• FinanceRetrain a fraud model on a weekly schedule and on drift detection, and deploy automatically only when it meets the evaluation bar.

Exam objectives (from the exam guide)

• エンドツーエンドのパイプライン（データとモデルの検証、Agent Platform Pipelines・Managed Service for Apache Airflow・Ray、学習と配信の前処理の一貫性）
• 再学習の自動化（再学習のポリシー、CI/CD/CT、Cloud Build）

Learning resources

• Professional ML Engineer exam guide (English, PDF) ↗
• PMLE exam guide / study guide ↗

Learning goals

• Choose tools for AI-specific security risks and countermeasures
• Detect skew and drift with Model Monitoring
• Evaluate generative AI apps continuously

Key points

• Model Armor inspects prompts and responses, detecting and blocking prompt injection, jailbreaks, sensitive data and harmful content.
• Regex detects fixed patterns (such as card numbers); safety filters block harmful categories.
• Training-serving skew: input distributions differ between training and production. Data drift: production input distributions change over time. Concept drift: the relationship between inputs and answers changes.
• Feature attribution drift: each feature's contribution to predictions changes.
• Model Monitoring alerts when thresholds are exceeded and can trigger retraining.
• Explainable AI shows the reasons for predictions with feature attributions (Sampled Shapley, Integrated Gradients, etc.).

Workplace uses

• HRFor a hiring-support model, regularly check prediction differences by attribute and consider retraining on Model Monitoring alerts.

Exam objectives (from the exam guide)

• AI のリスク（データの流出、悪意のあるプロンプト、LLM への機密情報の共有、Regex・セーフティフィルタ・Model Armor）、責任ある AI（バイアスの監視）、Explainable AI
• Model Monitoring による継続的な評価、学習と配信のスキュー・データドリフト・コンセプトドリフト・特徴量アトリビューションのドリフト、生成 AI の監視・テスト・評価

Learning resources

• Professional ML Engineer exam guide (English, PDF) ↗
• PMLE exam guide / study guide ↗