GCP Machine Learning Engineer Mind Map: Section 1

Every topic from Section 1 of the exam on one page, with the pairs the exam swaps to trick you. It's about 13% of your score.

1.1 Low-code ML: BigQuery ML and AutoML

Building models without writing training code

Agent Platform (formerly Vertex AI): Gemini Enterprise Agent Platform is the new name for Vertex AI: AutoML, training, Model Garden, endpoints and pipelines all live here now.

BigQuery ML: Create, train and use models with SQL, directly where the data already lives in BigQuery.
Agent Platform AutoML: Upload labeled data and let Google search for the best model automatically, for tabular or image tasks.
Tabular Workflows: Managed, configurable pipelines for tabular AutoML with more control over steps and architecture search.
AutoML data requirements: Labeled data with train, validation and test splits; enough examples per class.
Optimization objective: Tell AutoML what to optimize, such as AUC ROC, AUC PR or log loss for classification, RMSE or MAE for regression.
Training budget: Set a compute budget in node hours; more budget lets AutoML search longer.
AutoML forecasting: Agent Platform AutoML can train forecasting models on tabular time-series data.
Exporting AutoML models to the edge: Some AutoML image models can be exported to run on devices.

Classification: Predict a category, like churn yes/no or product type. In BigQuery ML: logistic regression, boosted trees, DNN.
Regression: Predict a number, like price or demand. In BigQuery ML: linear regression, boosted trees, DNN.
Forecasting: Predict future values of a time series. In BigQuery ML: ARIMA_PLUS.
Forecasting with extra signals: ARIMA_PLUS_XREG adds external variables, like promotions or weather, to a time-series forecast.
Pretrained forecasting (AI.FORECAST): Forecast in BigQuery with a pretrained time-series foundation model, without training your own.
Clustering: Group similar records without labels. In BigQuery ML: k-means.
Matrix factorization: Recommends items from user–item interactions, like ratings or purchases. In BigQuery ML: MATRIX_FACTORIZATION.
Anomaly detection: Flags unusual records or time points with ML.DETECT_ANOMALIES on models like ARIMA_PLUS, k-means or autoencoders.
Dimensionality reduction: PCA compresses many features into fewer components that keep most of the information.

TRANSFORM clause: Defines feature preprocessing inside CREATE MODEL, so the same transformations apply automatically at prediction time.
Feature engineering and selection: Create features with SQL and built-in functions (bucketizing, scaling, crosses) and keep only useful ones.
ML.PREDICT and ML.FORECAST: Generate predictions or forecasts with SQL on new data.
ML.EVALUATE: Returns evaluation metrics for a trained model.
Fine-tuning Gemini in BigQuery: Create a remote model over Gemini and tune it with training examples stored in a BigQuery table, all in SQL.
Data splitting: DATA_SPLIT_METHOD controls how training data is split for evaluation: automatic, random, custom column or sequential.
Hyperparameter tuning in BigQuery ML: Set NUM_TRIALS and search spaces in CREATE MODEL to tune automatically.
Explaining predictions: ML.EXPLAIN_PREDICT gives per-prediction feature attributions; ML.GLOBAL_EXPLAIN and ML.FEATURE_IMPORTANCE show overall importance.
Remote models: A BigQuery ML model that points to a model hosted elsewhere, like Gemini, so you call it from SQL with functions such as ML.GENERATE_TEXT or ML.GENERATE_EMBEDDING.
Imported models: Bring trained TensorFlow, ONNX or XGBoost models into BigQuery ML to predict with SQL.
Online serving for BigQuery ML models: Register BigQuery ML models in Model Registry and deploy them to an endpoint for real-time predictions.

Using Google’s prebuilt AI before building your own

Model Garden: Catalog of Google, partner and open models on Agent Platform, to compare, test and deploy.
Models as a service: Call partner or open models through a managed API, paying per use, with no infrastructure.
Self-deployed open models: Deploy an open model like Gemma or Llama to your own endpoint and hardware.
Gemini: Google’s multimodal foundation models for text, code, images, audio and video understanding.
Gemini image models and Veo: Gemini image models (formerly Imagen, shut down in August 2026) generate and edit images; Veo generates video.
Open models like Gemma: Google’s open-weight models you can tune and host yourself.
Prompt prototyping in the console: Test prompts, parameters and models interactively before writing code (formerly Vertex AI Studio).
Model cards: Each Model Garden model documents intended use, limitations, license and evaluation.

Document AI: Processors that extract text, forms, tables and fields from documents, with specialized ones for invoices, IDs and more.
Vision API: Image analysis: labels, objects, text (OCR), faces and explicit content detection.
Translation API: Translate text and documents; the advanced tier adds glossaries, batch and custom models.
Natural Language API: Entities, sentiment, syntax and content classification for text.
Speech-to-Text and Text-to-Speech: Transcribe audio to text, and generate natural speech from text.
Video Intelligence API: Detects labels, shots, objects and text in video.
Document AI processors: OCR processor for text, Form Parser for key-value pairs and tables, specialized processors (invoice, ID), and custom extractors you train.
API before custom model: If a prebuilt API solves the problem, use it before training anything.

Tuning for a use case: Supervised fine-tuning teaches a model a format, style or task with labeled examples.
Model choice for cost and latency: Use the smallest model that meets quality: Flash-class for speed and cost, Pro-class for hard reasoning.
Context caching: Store a long, repeated context once and reuse it across requests at lower cost.
Batch inference: Send many requests as one job for lower cost when results aren’t needed immediately.
Provisioned Throughput: Reserve capacity for predictable performance and availability at high volume.
Global endpoint: Lets Google route Gemini requests to available capacity across regions for higher availability.
Dynamic shared quota and 429s: Pay-as-you-go Gemini capacity is shared; under heavy demand you can get 429 errors, handled with retries and backoff.
Grounding: Connect Gemini to Google Search or your own data so answers rely on real sources.
Structured output: Constrain Gemini to return JSON matching a schema.
Safety settings: Configure Gemini’s blocking thresholds per harm category for your application.