Lesson 2: GCP AI/ML Ecosystem Overview

1. GCP ML Landscape Overview

GCP ML Capability Spectrum:

LOW CODE ◄────────────────────────────────────► HIGH CONTROL
  │                        │                           │
  ▼                        ▼                           ▼
Pre-trained APIs      Vertex AI AutoML        Custom Training
(Vision, NLP,         (no code needed,        (full control,
Translation)          you bring data)         you bring code)
  │                        │                           │
No ML expertise       Some domain              ML expertise
needed                expertise               required

BigQuery ML ────── SQL interface for ML on warehouse data

2. Vertex AI — Unified ML Platform

Vertex AI is GCP's unified platform for the entire ML lifecycle. Understanding its components is essential for the exam.

Component	Purpose
Vertex AI Workbench	Managed Jupyter notebooks for data scientists
Vertex AI Training	Custom training jobs (CPUs, GPUs, TPUs)
Vertex AI AutoML	No-code model training (Tabular, Image, Text, Video)
Vertex AI Endpoints	Deploy models for online prediction
Vertex AI Batch Prediction	Asynchronous batch scoring
Vertex AI Feature Store	Serve features consistently across training/serving
Vertex AI Pipelines	Kubeflow Pipelines-based ML workflow orchestration
Vertex AI Experiments	Track runs, compare metrics
Vertex AI Model Registry	Version control for models
Vertex AI Model Monitoring	Detect feature skew and prediction drift

3. AutoML vs. Custom Training

Criteria	AutoML	Custom Training
ML expertise needed	Minimal	Required
Training time	Hours (automated)	Variable (you control)
Model interpretability	Limited	Full control
Cost	Higher per model	Pay per compute used
Best for	Quick prototypes, standard tasks	Custom architectures, research
Supported data types	Tabular, Image, Text, Video	Any (you write the code)

Exam tip: Questions with "team doesn't have ML expertise" or "fastest time to deployment" → AutoML. Questions with "custom neural architecture" or "full control over training loop" → Custom Training.

4. BigQuery ML

BigQuery ML lets you train and serve ML models using SQL — no need to export data from BigQuery.

Model Type	SQL Keyword	Use Case
Linear Regression	LINEAR_REG	Price prediction
Logistic Regression	LOGISTIC_REG	Classification
K-Means Clustering	KMEANS	Customer segmentation
XGBoost	BOOSTED_TREE_CLASSIFIER/REGRESSOR	Tabular classification/regression
Deep Neural Network	DNN_CLASSIFIER/DNN_REGRESSOR	Complex patterns
Matrix Factorization	MATRIX_FACTORIZATION	Recommendations
Imported TF models	TENSORFLOW	Custom TF models

5. Pre-trained AI APIs

API	Capabilities	Use Case
Cloud Vision API	Labels, OCR, faces, logos, safe search	Image analysis without training
Cloud Natural Language API	Entities, sentiment, syntax, categories	Text analytics
Cloud Translation API	100+ language pairs	Multi-language content
Cloud Speech-to-Text	Transcription, speaker diarization	Audio processing
Cloud Text-to-Speech	WaveNet voices, SSML	Voice UI, accessibility
Document AI	Form parsing, invoice extraction	Document automation
Recommendations AI	Real-time product recommendations	E-commerce personalization

6. Service Selection Decision Tree

WHICH GCP ML SERVICE?

Do you have LABELED DATA?
│
├── NO → Pre-trained API sufficient for your task (Vision, NLP)?
│         YES → Use Pre-trained API
│         NO  → Vertex AI Custom Training (unsupervised)
│
└── YES → Is your data already IN BigQuery?
          │
          ├── YES → BigQuery ML (SQL-based, fast, no export)
          │
          └── NO → Need rapid prototyping, no ML team?
                    │
                    ├── YES → Vertex AI AutoML
                    │
                    └── NO  → Vertex AI Custom Training

7. Practice Questions

Q1: A data analytics team has petabytes of customer transaction data in BigQuery. They want to build a churn prediction model using their existing SQL skills without data exports. Which approach is BEST?

• A) Export to Cloud Storage, then use Vertex AI Custom Training
• B) Use Cloud Natural Language API
• C) Use BigQuery ML with CREATE MODEL LOGISTIC_REGRESSION ✓
• D) Use Vertex AI AutoML Tabular

Explanation: BigQuery ML allows training classification models directly on BigQuery data using SQL, leveraging existing data infrastructure and skills without exporting data. This is the fastest path when data is already in BigQuery.

Q2: A small startup needs to add sentiment analysis to customer reviews. They have no ML team and no labeled sentiment data. Which solution requires the LEAST effort?

• A) Vertex AI AutoML Text Sentiment
• B) Train a custom BERT model on Vertex AI
• C) Cloud Natural Language API sentiment analysis ✓
• D) BigQuery ML DNN classifier

Explanation: Cloud Natural Language API is a pre-trained, fully managed service that requires no training data, no ML expertise, and no infrastructure setup. Just call the API. AutoML requires labeled sentiment examples; custom BERT requires significantly more expertise.

Q3: Which Vertex AI component should a team use to ensure that feature values used during model training are identical to those served at prediction time?

• A) Vertex AI Experiments
• B) Vertex AI Feature Store ✓
• C) Vertex AI Model Registry
• D) Vertex AI Pipelines

Explanation: Vertex AI Feature Store provides a centralized repository for storing, serving, and sharing ML features. It ensures training-serving consistency by using the same feature definitions and values for both training and online/batch prediction, preventing training-serving skew.