Prepare and test your skills

Prepare and test your skills

Choosing between pre-trained APIs, AutoML, and custom training is how organizations get business value from Google Cloud AI at the right cost, speed, and accuracy. Pre-trained APIs are fastest when a general model is good enough, AutoML builds a tailored model from company data without deep coding, and custom training gives full control for specialized needs.
A decision tree branching on whether a pre-trained API is good enough, whether labeled data exists, and how much control is needed to pick pre-trained API, AutoML, or custom training.
Strategic use of pre-trained APIs means calling ready-to-use models instead of building models from scratch. Data flows from the application to the API, and the AI result flows back from the API to the application, while Google manages the model, training data, and infrastructure. Model Garden supports model selection with Google models like the Gemini family and Imagen, open models like Gemma and Llama, and third-party models from partners like Anthropic.
Common Google Cloud APIs cover standard jobs without custom training:
This approach reduces time-to-market from months to days and reduces the need for data science experts.
AutoML in Vertex AI is a managed training option that builds a custom model from an organization's own labeled data without extensive coding. Data flows from the organization's labeled dataset into the AutoML training service, and the trained model flows back to the business application to serve predictions. Because the model learns from company images, text, or tabular records, it captures patterns that a general model trained on public data misses.
AutoML improves accuracy for traditional AI tasks where categories are unique to the business, such as classification of defective products, regression to predict price or revenue, time series forecasting of sales and demand, and clustering for customer segmentation. The tradeoff is that AutoML depends on having sufficient, representative training data, and the team must still define the business goal and measure success in accuracy, revenue lift, or cost reduction.
The choice depends on whether an existing model is good enough, whether training data is available, and how much control the team needs. First check if a pre-trained API meets the requirement; if it does, use it to save time and cost.
| Option | Data needed | Expertise and control |
|---|---|---|
| Pre-trained API | No training data needed | Lowest effort, no control over model internals, best when general accuracy is enough |
| AutoML | Requires sufficient labeled proprietary data | Low expertise needed, managed tuning, best for higher accuracy without custom code |
| Custom training | Requires training data plus ML engineering capacity | Highest control over algorithm, layers, learning rate, hyperparameters, and explainability |
Custom training fits when the team needs detailed control and explainability, while a specialized product like Anti Money Laundering AI shows differentiation: it trains only on the customer's own banking data to generate monthly risk scores with explainability for investigation.
Exam tip: Use a pre-trained API when it meets the classification or detection need; use AutoML in Vertex AI when it does not meet the need but you have enough labeled data and want a custom model without extensive coding.
BigQuery ML is a capability of BigQuery that lets analysts build and run machine learning models directly inside the data warehouse using standard SQL. A model created with BigQuery ML is stored in a BigQuery dataset like a table or view, so training data never has to leave the warehouse for a separate Python environment.
Data locality in BigQuery ML means bringing machine learning to the data instead of moving data to the model. Training data flows from BigQuery tables into the training statement, and prediction results flow from the stored model back into BigQuery tables. This design reduces complexity, avoids data transfer costs and latency, and speeds up development because large datasets do not need reformatting.
Lowering the technical barrier means letting SQL practitioners build models with familiar tools. Users work through the Google Cloud console, the query editor, the bq command-line tool, the BigQuery REST API (opens in a new tab), Colab Enterprise notebooks, or external tools like Jupyter notebooks and business intelligence platforms. No Python or Java programming is required, so the people who already understand the data can build the models.
The SQL workflow in BigQuery ML is the ordered path from stored tables through training to prediction without leaving BigQuery. The order matters because evaluation and prediction both depend on the stored model produced by training.
A left-to-right process flow showing six ordered BigQuery ML steps from dataset creation through CREATE MODEL, refine, ML.EVALUATE, ML.PREDICT, and explain, all inside BigQuery.
CREATE MODEL statement to train a model on a query result.TRANSFORM functions and hyperparameter tuning when needed.The three core statements are CREATE MODEL to train, ML.EVALUATE to check quality, and ML.PREDICT to generate predictions. BigQuery ML supports both batch predictions over large datasets and online predictions on individual rows.
Regression and classification models in BigQuery ML predict a label from labeled examples. Linear regression predicts a continuous number such as price, while logistic regression assigns each row to one of up to 50 categories. Other tabular types include boosted trees based on XGBoost, random forest, deep neural network (DNN), Wide and Deep for recommender and ranking, and AutoML tabular classifiers.
Other built-in models handle different question shapes. K-means clustering segments data such as customer groups without labels, matrix factorization builds product recommendations from behavior and ratings, principal component analysis reduces dimensionality, time series models like ARIMA_PLUS forecast and detect anomalies, and contribution analysis measures how dimensions like store location affected revenue.
Evaluation measures how a model will perform on new data. ML.EVALUATE evaluates a stored model, while ML.METRICS scores any table that already has actual and predicted columns.
| Function | Requires stored model | Input |
|---|---|---|
| ML.EVALUATE | Yes | No input returns training-time metrics; specified input tests on new data |
| ML.METRICS | No | Any table or query with actual and predicted columns |
Exam tip: Use ML.EVALUATE with no input to see training-time metrics, use ML.EVALUATE with input data to test on new data, and use ML.METRICS when there is no stored model, only actual and predicted values.
Selecting the correct Google Cloud pre-trained API (opens in a new tab) means matching the job — audio direction, image task, or text task — to the service built for that job. All of these APIs are pre-trained, so developers add the capability with an API call without training a custom model.
Audio APIs are chosen by conversion direction. Speech-to-Text converts spoken audio into written text for input scenarios, while Text-to-Speech converts written text into natural-sounding speech for output scenarios.
| API | Conversion direction | Choose when |
|---|---|---|
| Speech-to-Text | Audio to text | Input is spoken audio and output must be written text, such as voicemail to message or captions |
| Text-to-Speech | Text to audio | Input is written text and output must be spoken audio, such as phone assistants, notifications, and accessibility readers |
Audio flows from the spoken source to the Speech-to-Text API and text flows back as the result. Text flows into the Text-to-Speech API and audio flows out as the result. When a solution needs both directions, the APIs compose into a voice loop where Speech-to-Text converts the request to text for processing, then Text-to-Speech converts the answer back to speech. Speech-to-Text can also run at scale from BigQuery through a remote model and the ML.TRANSCRIBE function, with inference occurring in Vertex AI Agent Platform.
The Cloud Vision API is the pre-trained API that analyzes images without custom training. An application sends an image from a local file or Cloud Storage with a requested feature type, and the API returns annotations describing what it found.
Image labeling and text reading cover organization and digitization. LABEL_DETECTION examines the whole image and returns objects, scenes, or activities for photo libraries and product catalogs. Optical character recognition (OCR) converts visible text to machine-readable text: TEXT_DETECTION is optimized for sparse text in a larger image such as a sign or shelf label, while DOCUMENT_TEXT_DETECTION is for dense documents, PDF or TIFF files, and handwriting and takes precedence when both are requested.
Object and safety detection covers location and compliance. OBJECT_LOCALIZATION detects multiple objects and where each sits in the image for shelf counting and inventory, FACE_DETECTION finds faces, LOGO_DETECTION finds company logos, LANDMARK_DETECTION finds geographic landmarks, and WEB_DETECTION finds similar images and topical entities on the web. SAFE_SEARCH_DETECTION screens images for unsafe content, and the application then decides to allow, block, blur, or send the image for human review.
Text APIs split by function. The Cloud Natural Language API analyzes text structure and meaning, while the Cloud Translation API translates text between languages. When text is sent to the Cloud Translation API, Google holds it briefly in memory only to perform the translation, does not store it, does not use it to train models, does not share it, and does not claim ownership of it.
Translation endpoints and glossaries control location and terminology. Cloud Translation - Basic API offers only global endpoints, AutoML Translation offers only global endpoints, and Cloud Translation - Advanced API offers both global endpoints and regionalized endpoints configurable for data location. A business that must keep translation data inside a region therefore needs Cloud Translation - Advanced API with a regionalized endpoint. Glossaries store customer-specific terminology encrypted in the customer's own project and are protected by fine-grained Identity and Access Management permissions. Translation can be a feature inside a larger application, but the API cannot be repackaged and resold as a standalone translation service.
Exam tip: Speech-to-Text converts audio to text, while Text-to-Speech converts text to audio — check the starting input to pick the correct API.
AutoML-driven business value is the measurable gain an organization gets when it trains custom models on its own proprietary data and uses them in real work. Proprietary data such as purchases, product images, sales history, and support tickets teaches the model the patterns of that specific business, so predictions reflect real customers and operations better than a generic model alone.
Proprietary data makes predictions more relevant because the model learns from the organization's past outcomes instead of public examples. Training data flows from warehouse and storage systems to the model, and predictions flow back to teams and applications that act on them. BigQuery ML keeps this path short by training inside the warehouse, so data stays where it is governed.
Data suited for custom training is historical data where past examples show the correct outcome:
| Use case | Example training data | What the model predicts |
|---|---|---|
| Classification | Emails marked spam or not spam | Category for a new email or product |
| Regression | Houses with features and sale price | Continuous number such as future price |
| Forecasting | Past sales by day or season | Future sales and demand |
| Clustering | Customer purchase records | Customer segments with similar behavior |
Reusing governed data lowers cost and speeds development. Knowledge Catalog (opens in a new tab) unifies distributed data without moving it, organizes data into lakes and zones, and controls access through groups and catalog roles. BigQuery sharing exchanges curated assets as listings, data clean rooms allow controlled collaboration, Agent Platform Feature Store centralizes features for online and offline use, and BigQuery object tables govern unstructured data.
Managed services automate building so teams without deep programming expertise can still train custom models. BigQuery ML lets SQL users train with GoogleSQL, Agent Platform Pipelines and Kubeflow Pipelines orchestrate transformation, training, and deployment, Model Registry stores versions for reuse, and Model Garden provides pre-trained Google, open, and third-party models that can be adapted with learning transfer using less data and compute.
Business value is measured by connecting each model to a business goal and tracking return on investment. Predictive AI is judged on accuracy and precision, computer vision AI on accuracy, and generative AI on adoption, satisfaction, and content quality. Industry metrics then tie those outputs to retail revenue lift, healthcare patient outcomes, or finance fraud reduction.
Value improves through an iterative cycle: first monitor indicators and costs to find deviations, then make data-driven adjustments to models and infrastructure, then refine objectives as needs evolve, and finally review results with stakeholders to plan future work. Machine learning operations (MLOps) scales this reliably with version control in Git, continuous integration and delivery with Cloud Build and Artifact Registry, observability with Cloud Monitoring and Cloud Logging, scheduled retraining with Agent Platform Pipelines, and autoscaled serving with load balancing. FinOps practices assign cost owners, label resources by project and use case, and analyze billing data exported to BigQuery to prove ROI.
Exam tip: Predictive and computer vision models are judged on accuracy, plus precision for predictive, while generative models are judged on adoption, satisfaction, and content quality.
Custom models on Vertex AI (opens in a new tab) are machine learning models trained or adapted on an organization's own data to solve its specific problems. A generic pre-trained solution handles common tasks the same way for everyone, while a custom model learns unique patterns, language, and customer behavior that competitors using only generic models cannot copy.
The Vertex AI Model Development Lifecycle is the ordered workflow from preparing data to serving predictions, with managed infrastructure at each stage. Data moves forward from one managed service to the next, which lowers complexity.
Data preparation turns raw data into features, where a feature is an attribute used as input for training or inference. Vertex AI Feature Store manages those features, with online serving retrieving the latest values for real-time inference and offline serving exporting high volumes of historical data for training. A managed dataset is typically split with about 80% for training, plus separate validation and test sets. Training can use pre-built containers for TensorFlow and PyTorch or custom jobs, while Vertex AI Experiments tracks steps, inputs, and outputs and Vertex AI TensorBoard visualizes runs. Vertex AI Vizier tunes hyperparameters, Model Evaluation generates metrics such as F1 score, and Explainable AI shows feature importance.
Deployment in Vertex AI is placing a trained model into production for inference on unseen data. Vertex AI Model Registry is the central repository where teams organize versions and deploy a chosen version directly or by alias to an endpoint. The registry creates a boundary between development, where versions change often, and production, where a stable version serves traffic.
| Inference mode | Input pattern | Result pattern |
|---|---|---|
| Batch inference | Group of requests processed together | Results output together in one file |
| Online inference | Individual data points sent as they arrive | Results returned in real time |
Operational integration connects the endpoint to products through APIs and SDKs for Python, Go, Java, and Node.js. For generative experiences, grounding techniques connect custom knowledge to responses: information retrieval finds relevant documents, and retrieval-augmented generation adds that content to the prompt so answers reflect current business data. Vertex AI Agent Engine then deploys and scales AI agents in production.
Generic pre-trained solutions accelerate time to value for standard needs such as optical character recognition, Speech-to-Text, and translation. Vertex AI Model Garden favors reuse by offering procured models from Marketplace. Choose a generic solution when the need is standard, the data is not proprietary, and speed matters more than uniqueness.
| Option | Data used | Strategic effect |
|---|---|---|
| Generic pre-trained API | Broad public data | Fast to launch but the same for every competitor |
| Custom model on Vertex AI | Organization's own image, text, tabular, and video data | Harder to build but creates a capability competitors do not have |
| Fine-tuned foundation model | Broad base plus company examples | Balances strong base quality with business-specific voice and knowledge |
Cost, performance, and scalability determine whether differentiation lasts. Fully managed deployment lets Vertex AI instantiate resources after the team selects machine type, while self-managed serving on Google Kubernetes Engine gives fine-grained control. Larger models can learn more complex patterns but can have higher latency and cost, with some models charged by tokens and others by node hours. Google foundation models already have project-specific endpoints, but a tuned model must be deployed to an endpoint before use.
Exam tip: Choose a generic pretrained API when the task is common across businesses, and choose a custom or fine-tuned Vertex AI model when advantage comes from learning unique business data and delivering a distinct user experience.
TensorFlow is end-to-end open source software for building and training machine learning models, while Cloud TPU is Google's proprietary hardware that accelerates those models. TensorFlow provides the tools from preparation to deployment, and Cloud TPUs provide the specialized math performance that runs TensorFlow jobs faster than general-purpose processors.
Explanation of TensorFlow shows it supports the full model lifecycle in one platform, from developing a model to training it on large datasets to deploying it for use. Google open sourced TensorFlow in 2015, and it has become one of the most popular ML frameworks. What TensorFlow provides across environments is a common set of tools that run on many different systems.
Data movement is central because training speed depends on how quickly data flows from storage to the training process running on accelerators. As GPUs and Cloud TPUs make the math faster, slow data delivery can become the bottleneck. TensorFlow Enterprise (opens in a new tab) is Google Cloud's enterprise version for businesses that want enterprise-grade support and performance from the creators of TensorFlow, with faster and easier access from cloud storage to training.
Tensor Processing Units (TPUs) are Google's custom application-specific integrated circuits designed to accelerate ML and AI workloads. A TensorFlow job sends heavy math operations to the TPU, and the TPU returns results much faster than general-purpose compute. Three hardware capabilities drive this acceleration:
Because of this design, Cloud TPUs fit TensorFlow models dominated by matrix computations, with no custom operations in the main loop, that train for weeks or months, use large effective batch sizes, or use ultra-large embeddings for ranking and recommendation. Cloud TPUs fit poorly for linear algebra with frequent branching, workloads requiring high-precision arithmetic, or neural networks with custom operations in the main training loop.
Provisioning TPUs is how Google Cloud controls availability and scaling through managed services. Compute Engine creates and manages individual TPU VMs or slices with full lifecycle management, while Google Kubernetes Engine (GKE) orchestrates TPU nodes and node pools inside Kubernetes clusters for large-scale AI workloads. Cloud TPU API (opens in a new tab) provisioning is legacy and receives only bug fixes, so new resources should use Compute Engine or GKE.
| Option | What it manages | Choose when |
|---|---|---|
| Compute Engine | Individual TPU VMs or slices | The team needs direct control over TPU VM lifecycle |
| Google Kubernetes Engine (GKE) | TPU nodes and node pools | The workload needs Kubernetes orchestration across many AI jobs |
Compute Engine also offers GPU-based accelerator-optimized families. The A series targets high-performance computing and AI, with later A4 and A4X types for pre-training foundation models on large clusters, while the G series targets graphics, simulation, and inference, with G2 using Ada Lovelace and 24 GB memory for inference and G4 using Blackwell and 96 GB per GPU for simulation and local large language model fine-tuning. AI Hypercomputer provides densely connected infrastructure with GKE and Slurm integrations for network-bound training. Project setup for TPU VMs follows a fixed order: create a project with billing, enable the Compute Engine API, grant user roles for instances and service accounts, create a least-privilege service account, and grant it Storage Admin and Logs Writer before attaching it to the TPU VM.
Exam tip: TensorFlow is open source software for building and training models, while Cloud TPU is Google's proprietary hardware that accelerates those models.
CREATE MODEL trains, ML.EVALUATE checks quality, and ML.PREDICT generates predictions without moving data.TEXT_DETECTION for sparse text and DOCUMENT_TEXT_DETECTION for dense documents and handwriting.Use a pre-trained API when an existing general model meets the classification or detection need because it requires no training data and reduces time-to-market from months to days. Use AutoML in Vertex AI when the pre-trained option does not meet the need but you have sufficient labeled proprietary data and want a custom model without extensive coding, and use custom training when you need the highest control over algorithm, layers, learning rate, hyperparameters and explainability.
BigQuery ML lets analysts build and run models directly inside the data warehouse using standard SQL, with the model stored in a BigQuery dataset like a table or view so training data never leaves the warehouse. The ordered workflow is to create the model with CREATE MODEL to train, then check quality with ML.EVALUATE and generate predictions on new data with ML.PREDICT, supporting both batch predictions and online predictions.
The Speech-to-Text API converts spoken audio into written text for input scenarios such as voicemail to message or captions, while the Text-to-Speech API converts written text into natural-sounding speech for output scenarios such as phone assistants, notifications and accessibility readers. Check the starting input to pick the correct API, and the two can compose into a voice loop where Speech-to-Text converts the request to text and Text-to-Speech converts the answer back to speech.
TensorFlow is end-to-end open source software for building and training machine learning models across the full lifecycle from development to training to deployment. Cloud Tensor Processing Unit (TPU) is Google proprietary hardware that accelerates those models with Matrix Multiply Units for matrix computations, high-bandwidth memory for larger models, and slices that scale a single job up to thousands of chips.
Last updated