CDL
Deep Dive
Key AI/ML concepts, types of learning, CDL exam terminology, and the spectrum of Google Cloud solutions.
These four technologies are used on the CDL exam with specific meanings. Understanding the differences is essential.
A broad field of computer science that seeks to create systems capable of performing tasks that normally require human intelligence: reasoning, learning, perception, language understanding.
Examples on GCP
Recommendation systems, speech recognition, machine translation, playing chess.
A subset of AI where systems learn patterns from data without being explicitly programmed for every case. The model improves with more data and feedback.
Examples on GCP
Spam detection, churn prediction, fraud detection, dynamic pricing.
A subset of ML using artificial neural networks with many layers (deep). Especially good for unstructured data: images, audio, text.
Examples on GCP
Image recognition (Vision API), speech transcription (Speech-to-Text), neural translation.
A subset of Deep Learning that can generate new content (text, images, code, audio) that was not explicitly in the training data. Based on foundation models like Gemini.
Examples on GCP
Gemini for text/code, Imagen for images, Veo for video, Chirp for voice. Summarizing documents, writing code, creating images.
Supervised
The model learns from labeled data — (input, correct output) pairs. It learns to map inputs to outputs.
Use cases
Classification (spam/not spam), regression (house price), disease detection.
Unsupervised
The model learns patterns from data WITHOUT labels. It discovers hidden structure in the data.
Use cases
Customer segmentation, anomaly detection, image compression.
Reinforcement
An agent learns by trial and error: it takes actions, receives rewards/penalties, and optimizes its policy.
Use cases
Games (AlphaGo), robots, dynamic recommendation systems, trading.
Google Cloud offers AI at three levels of abstraction. Choose based on your use case and technical experience.
Google's ready-to-use models. No training data of your own. You just call the API.
Train custom models with your own data, without writing ML code.
Full control: you write your own ML code with any framework.
| Term | Definition for the CDL |
|---|---|
| Foundation model | A large-scale AI model trained on enormous amounts of data that can adapt to many tasks. E.g.: Gemini, Imagen, Gemma. |
| Fine-tuning | Adjusting a foundation model with domain-specific data to improve its performance on specific tasks. |
| Hallucination | When an LLM generates factually incorrect information with confidence. Reduced with grounding and RAG. |
| RAG | Retrieval-Augmented Generation: grounding AI responses with your own documents to reduce hallucinations. |
| Grounding | Connecting the model to up-to-date data sources (web, your own documents) so its responses are factual. |
| Embeddings | Numerical representations (vectors) of text/images that capture their semantic meaning. The foundation of vector search. |
| Temperature | A parameter that controls the model's randomness. 0 = deterministic. 1+ = more creative/random. |
| Prompt engineering | The technique of writing effective instructions to get better responses from an LLM. Zero-shot, few-shot, chain-of-thought. |
Pre-trained APIs let you use Google's AI capabilities without your own training data or ML code. They are the lowest-effort level in the GCP AI spectrum.
| API | What it does | Typical use case |
|---|---|---|
| Vision AI | Detects objects, text (OCR), faces, labels, and explicit content in images. Landmark recognition. | Image moderation, extracting text from scanned documents, photo classification. |
| Natural Language API | Sentiment analysis, entity extraction, content classification, and syntax analysis in text. | Analyzing customer reviews, categorizing support tickets, extracting information from contracts. |
| Cloud Translation API | Translates text across 100+ languages. Automatic source language detection. | Multi-language support in apps, content translation, global apps. |
| Speech-to-Text | Converts spoken audio to text. Supports 125+ languages, specialized vocabulary (medical, telephony). | Automatic captions, support call transcription, voice commands. |
| Text-to-Speech | Converts text to audio with natural voices (WaveNet). Supports 220+ voices in 40+ languages. | Voice assistants, screen readers, accessibility apps. |
| Document AI | Extracts structured data from complex documents (invoices, contracts, forms). Goes beyond simple OCR. | Automating invoice processing, customer onboarding with KYC, extracting data from contracts. |
| Contact Center as a Service (CCAI Platform) | A unified, AI-powered contact center platform: virtual agents (Conversational Agents, formerly Dialogflow CX), real-time assistance for human agents, and conversation analytics. Featured AI product on GCP. | AI-powered customer service centers: deflect calls to virtual agents, assist human agents in real time, analyze sentiment and call quality. |
To choose between pre-trained APIs, AutoML, or Custom Training, evaluate four factors:
| Factor | Pre-trained APIs | AutoML | Custom Training |
|---|---|---|---|
| Speed | Minutes (just call the API) | Hours-days (training) | Days-weeks |
| Effort | Minimal: no data, no ML code | Medium: your own data, no code | High: code, architecture, tuning |
| Differentiation | Low: everyone uses the same model | Medium: model with your data | High: unique model for your case |
| Expertise required | Developer (API calls only) | Data analyst | Data Scientist / ML Engineer |
BigQuery ML — ML without leaving BigQuery
BigQuery ML lets you create and run ML models directly in BigQuery using standard SQL. There's no need to export data or learn an ML framework. Ideal for data analysts who already know SQL and want regression, classification, time series, or recommendation models on data that's already in BigQuery.
Did you get this topic?
Test what you just learned
A company wants to detect defective products on its production line using images from cameras. They don't have an ML team but have thousands of labeled images (defective/not defective). Which Google Cloud solution is most suitable?
Put what you learned to the test with original CDL exam questions.
Start CDL mock examUnlimited mock exams with the Pro plan
Log in to track your progress.