CDL AI/ML solutions covers approximately 15-20% of the exam. Knowing which AI solution to choose and understanding Vertex AI and pre-trained APIs are the key skills tested.
What are AI, ML, and Deep Learning?
These three concepts form a nesting hierarchy — like Russian dolls: Deep Learning is inside ML, which is inside AI.
AI (Artificial Intelligence) is the broadest concept: making machines think and reason like humans. Chess computers, auto-translation systems, and voice assistants are all forms of AI.
ML (Machine Learning) is one way to implement AI. Instead of humans programming explicit rules, the system learns patterns from large amounts of example data. A spam filter, for instance, learns from millions of spam/legitimate email examples to judge new emails on its own.
Deep Learning is a subset of ML using artificial neural networks inspired by the human brain. It excels at complex pattern recognition: image recognition, natural language processing, speech recognition. Modern LLMs (Large Language Models) like ChatGPT are deep learning technology.
| Concept | Core Definition | Example | |---------|----------------|---------| | AI | Machines think like humans | Auto-translation, chess AI | | ML | Learn patterns from data | Spam filters, recommendation algorithms | | Deep Learning | Neural network-based ML | Image recognition, ChatGPT |
Business value of ML: Prediction (demand forecasting, churn prediction), Automation (auto-classify tasks), Personalization (tailored product recommendations), Anomaly Detection (fraud detection), Natural Language Understanding (auto-classify customer queries).
Google Cloud AI Solution Spectrum
Google Cloud AI/ML solutions come in three tiers based on required ML expertise. The most important CDL skill is judging which tier fits each situation — like choosing how to cook based on skill level: meal kit (pre-trained APIs), following a recipe (AutoML), or creating from scratch (custom models).
!Google Cloud AI solution spectrum from pre-trained APIs to custom models
Level 1: Pre-trained APIs — No ML Expertise Required
Google has already trained models on billions of data points and exposes them as APIs. Just a few lines of code to add powerful AI to your app. No ML knowledge required.
| API | Capabilities | Use Case Examples | |-----|-------------|------------------| | Vision API | Object detection, facial recognition, OCR, explicit content detection | Auto-tag product images, ID OCR | | Natural Language API | Sentiment analysis, entity recognition, syntax analysis | Customer review sentiment, news classification | | Speech-to-Text | Speech → text (120+ languages) | Auto-transcribe call centers, voice commands | | Text-to-Speech | Text → speech | Audiobooks, voice navigation | | Translation API | Auto-translate 100+ languages | Multilingual content, real-time chat translation | | Video Intelligence API | Scene detection, object tracking, explicit content detection | Auto-classify video content | | Document AI | Document understanding and data extraction | Auto-process invoices, contract analysis |
Use when: generic AI capabilities are sufficient, no domain customization needed, fast implementation is important.
Level 2: AutoML — Some Data, Minimal Code
Train a custom model with your own data automatically. For example, the generic Vision API recognizes common objects, but if you need to classify your company's specific product types, AutoML can build a custom image classifier from your labeled photos.
No ML coding required. AutoML automatically searches for optimal neural network architecture and hyperparameters. Only hundreds to thousands of labeled training examples needed.
Use when: pre-trained APIs don't fit your domain, you need a custom model without ML engineers.
Level 3: Custom Models — Maximum Flexibility
Build models from scratch using ML frameworks like TensorFlow, PyTorch, or scikit-learn. Maximum control but requires ML expertise and substantial data.
Use when: solving a highly specialized problem, pre-trained APIs and AutoML are insufficient, you have an ML research and development team.
AI Solution Selection Flow:
Vertex AI — Unified ML Platform
Vertex AI is Google Cloud's fully managed ML platform — handle every stage of the ML pipeline in one place. Supports everything from AutoML to custom TensorFlow/PyTorch models.
Key capabilities: Workbench: Jupyter notebook ML development environment Managed Datasets: training data management and versioning AutoML: auto-train custom models without code Training: distributed training for custom TensorFlow/PyTorch models Model Registry: model version management and metadata tracking Prediction (Endpoints): deploy models as REST APIs Pipelines: ML workflow orchestration and automation Explainable AI: explain why a model made a prediction
Vertex AI unified the previously separate AutoML and AI Platform services.
BigQuery ML — Machine Learning with SQL
BigQuery ML occupies a unique position: data analysts (not ML engineers) can train and run ML models directly inside BigQuery using only familiar SQL — no Python, no data movement.
Biggest advantage: no data movement. Perform ML where the data already lives in BigQuery. Reduces data engineering work and lets analysts run ML themselves.
Supported models: linear regression, logistic regression, k-means clustering, deep neural networks, XGBoost, time series forecasting.
Responsible AI
AI is powerful but can cause discrimination or bias if misused. Google emphasizes these principles (also tested on CDL):
Fairness: prevent AI from being biased against specific groups Transparency: understand and explain AI decision processes (Vertex AI Explainable AI supports this) Safety: design AI to not behave in unintended ways Privacy: protect personal data during training and inference Accountability: clearly define responsibility for AI outcomes
Exam Key Points
"Machines think like humans" -- AI / "Learn patterns from data" -- ML / "Neural network-based" -- Deep Learning
"Pre-trained model via API, no ML expertise needed" -- Pre-trained APIs
"Image analysis (objects, faces, OCR)" -- Vision API
"Text sentiment/entity analysis" -- Natural Language API
"Convert speech to text" -- Speech-to-Text API
"Translate 100+ languages" -- Translation API
"Auto-train custom model with own data, no ML coding" -- AutoML
"Build with TensorFlow, maximum flexibility" -- Custom Model
"Unified ML platform: training → deployment → management" -- Vertex AI
"ML with SQL in BigQuery, no data movement" -- BigQuery ML
"Prevent AI bias, fairness/transparency/safety" -- Responsible AI
Expertise level: Pre-trained APIs (least) → AutoML → Custom Model (most)