Prepare and test your skills

Prepare and test your skills

Artificial intelligence (AI) and machine learning (ML) on Google Cloud are delivered as managed services that help teams govern data, reuse ML assets, automate work, and start from pre-trained models. The platform ranges from fully managed tools to fully configurable infrastructure, so teams choose based on how much control and management they want. Data flows from governed storage into notebooks and pipelines for training, then trained models flow into a registry and to prediction endpoints, with cost tracking alongside every stage.
Governing and reusing data for AI and ML means organizing distributed data with clear policies so models train on high-quality inputs without duplicated work. A governance framework defines policies, procedures, and roles that improve data quality, security, and utilization while reducing redundancy and cost. Knowledge Catalog is an intelligent data fabric that unifies distributed data without moving it into one place, which lowers distribution costs and promotes reuse. Within Knowledge Catalog, logical data lakes and zones organize data, Google Groups and roles control access, auto data quality streamlines quality checks, and a universal catalog manages metadata with custom entries and business context for easier discovery.
Reusable data products let teams share work instead of rebuilding the same inputs for every project. BigQuery sharing exchanges curated assets through data exchanges that hold listings, data clean rooms that protect sensitive data for partner projects, and permissions that control who can view or publish datasets. For structured data, Agent Platform Feature Store provides a central repository for organizing, storing, and serving features for online and offline use. Unstructured data is stored in Cloud Storage and governed through BigQuery object tables and signed URLs, while vector embeddings are managed with metadata in Vector Search indexes.
Starting from pre-trained models and APIs means reusing models trained on vast datasets instead of training complex models from scratch. Training from scratch requires significant compute and time, so reuse reduces cost and speeds up development. Model Garden offers pre-trained models for tasks such as image classification, object detection, and natural language processing:
Google Cloud also provides AI APIs that add intelligence to applications without building models, with each API focused on one content type:
Pre-trained models can be adapted with learning transfer, which applies knowledge from a pre-trained model to a new task with less data and compute time.
Managed notebooks for experimentation are shared, interactive environments where data scientists explore data, develop models, and iterate without managing hardware. Collaboration in notebooks speeds up development, code reviews, and knowledge transfer while reducing duplicated effort. Teams use on-demand infrastructure instead of buying expensive machines, which keeps experimentation cost-effective.
Agent Platform Workbench is a Jupyter notebook environment for the full data science workflow. From a notebook, users access data through BigQuery (opens in a new tab) and Cloud Storage integrations, automate recurring updates with scheduled executions, process data on a Managed Service for Apache Spark cluster, and run a notebook as a pipeline step. Colab Enterprise is a collaborative managed notebook environment with Google Cloud security and compliance. Users share notebooks with a single user, a Google group, or a Google Workspace domain with access controlled by Identity and Access Management (IAM), integrate with Agent Platform and BigQuery, and track changes with version control tools like Git.
Automating the ML lifecycle with MLOps means using pipelines and reusable assets to reduce manual work and cost for data scientists and engineers. Automation avoids duplication, enforces dependencies between steps, and manages data flow from transformation through training to deployment. Google Cloud offers orchestration tools that fit different pipeline needs:
Reliable pipelines depend on versioning, reusability, and monitoring. Versioning tracks pipeline and component versions for reproducibility, with Kubeflow templates stored in Artifact Registry. Reusability stores templates in Cloud Storage for sharing, while Feature Store reuses features, Agent Platform datasets reduce data duplication, Model Registry stores trained models for later pipelines or online prediction, and custom containers in Artifact Registry provide consistent training environments. Monitoring watches execution with Cloud Logging and Cloud Monitoring to find and fix issues.
BigQuery ML (opens in a new tab) lets users create, evaluate, and run machine learning models with GoogleSQL queries or the Google Cloud console. Models are stored in BigQuery datasets like tables and views, so SQL practitioners can build models without extensive programming or ML frameworks. Users work through the console, query editor, bq command-line tool, BigQuery REST API, Colab Enterprise notebooks, or external tools such as Jupyter notebooks and business intelligence platforms.
BigQuery ML brings ML to the data rather than moving data to ML. Because large datasets stay in the warehouse, fewer tools are needed and models reach production faster without reformatting data for Python frameworks. BigQuery ML also accesses Agent Platform models and Cloud AI APIs for tasks such as text generation and translation, with Gemini assistance available for BigQuery tasks.
Exam tip: BigQuery ML brings ML to the data so analysts can build models with SQL, while Agent Platform Pipelines and Kubeflow Pipelines orchestrate full training and deployment workflows.
Choosing a platform for AI/ML workloads means matching the level of control and management to the workload. Google Kubernetes Engine (GKE) is a Google-managed Kubernetes platform for containerized AI/ML work, Vertex AI is a fully managed end-to-end platform, and Cloud Run is a serverless platform for containerized inference. The choice depends on model size, latency needs, expertise, and portability.
| Service | What it is | Choose when |
|---|---|---|
| Google Kubernetes Engine (GKE) | Managed Kubernetes for containerized AI/ML | You need deep infrastructure control, train very large models, serve with minimal latency, use Spot VMs and Flex-start VMs for efficiency, or want portability and open standards |
| Vertex AI | Fully managed end-to-end platform | You want to accelerate development, offload infrastructure, and focus on MLOps and rapid time-to-value |
| Cloud Run | Serverless platform for containerized inference | You want inference that scales to zero, event-driven apps, or cost-effective serving of smaller models |
Google Kubernetes Engine (GKE) (opens in a new tab) supports inference, training, and unified development. For serving, it offers GPU and TPU flexibility, GKE Inference Gateway for model-aware routing and load balancing, benchmarked deployment profiles, and Autopilot for automated operations and right-sizing. For training, faster startup nodes cut GPU startup times by up to 80%, flex-start provisioning helps secure scarce GPUs and TPUs for short jobs, Kueue provides Kubernetes-native queueing and prioritization, and TPU multislice connects multiple slices over the Data Center Network (DCN) for large-scale training. For unified development, the Ray on GKE add-on abstracts Kubernetes so teams can scale preprocessing, distributed training, and serving with minimal code changes.
Controlling and optimizing AI and ML costs depends on visibility into where costs come from so they can be assigned to projects, teams, and activities. FinOps (opens in a new tab) principles give each project estimated costs, ongoing tracking, clear owners, and accountability for costs and benefits. Resources are organized by organizations, folders, and projects to mirror teams and workflows, with consistent labels for project, team, environment, model, dataset, and use case. Billing data exported to BigQuery supports granular analysis by label, project, or time period, while anomaly detection, cost allocation, and prebuilt Cloud Billing reports reveal patterns and surprises.
Teams validate return on investment by starting with pilots and following an iterative cycle:
Core ML processes are the repeatable stages that turn raw data into useful predictions in the cloud. The foundational workflow moves from data collection to training to evaluation to prediction, with Google Cloud services supporting each stage. BigQuery ML, Agent Platform, and Google Kubernetes Engine (GKE) provide different places to run that workflow depending on skill set and control needs.
The foundational ML workflow is the ordered sequence every machine learning project follows from data to results. Data flows from collection and preparation into training, then into evaluation and tuning, and finally into prediction for real use:
A four-step process flow showing the foundational ML workflow in order: data collection and preparation, training, evaluation and tuning, and prediction (inference), with the Google Cloud services that support each stage.
Each stage depends on the one before it. Training depends on prepared datasets and features, evaluation depends on completed training runs, and prediction depends on a tuned model stored in a registry or dataset.
Collecting, governing, and reusing data means unifying distributed data, controlling access, and sharing assets across projects. Knowledge Catalog is an intelligent data fabric that unifies distributed data and automates governance without consolidating datasets, which improves quality, security, and utilization while reducing redundancy and cost. Good governance is what makes later training and prediction reliable.
Google Cloud provides peer tools that keep ML data discoverable throughout the lifecycle:
Reusable data products, datasets, and features avoid redundant engineering. Teams discover shared assets through Knowledge Catalog and BigQuery sharing instead of rebuilding the same inputs for every model.
Training models in the cloud is the process where a model learns patterns from data so it can make predictions later. BigQuery ML (opens in a new tab) lets analysts create and run models with GoogleSQL queries or the console, with models stored in BigQuery datasets like tables and views. Training queries run where the data already lives, so large datasets do not need to be moved and reformatted for Python frameworks.
Teams that need larger or more custom training use managed pipelines and platforms. Agent Platform Pipelines and Kubeflow Pipelines orchestrate steps such as data transformation, training, deployment, and versioning with defined dependencies and data flow. Google Kubernetes Engine (GKE) (opens in a new tab) provides scale for very large jobs with faster startup nodes, job queueing with Kueue, and access to GPUs and TPUs. To save time and compute, teams can start from Model Garden, which offers pre-trained Google, open, and third-party models for image, language, and other tasks.
Evaluating and tuning models is the stage where a trained model is tested, compared, and improved before production. Agent Platform provides experiments and evaluation tools that track dataset version, model parameters, and model type across runs. Predictive models are measured on accuracy and precision, while generative and vision models are measured on adoption, satisfaction, content quality, and accuracy.
Interactive notebooks support cost-effective experimentation before scaling up. Agent Platform Workbench is a Jupyter environment that connects to BigQuery (opens in a new tab), Cloud Storage, and Agent Platform for exploration and scheduled updates. Colab Enterprise is a collaborative managed notebook environment with Google Cloud security that integrates with Agent Platform, BigQuery, and Gemini assistance. Notebook work can be versioned with Git so changes can be tracked and reverted.
Making predictions with trained models, also called inference, is when a finished model generates results on new data. Model Registry on Agent Platform stores, manages, and deploys trained models so they can be reused in pipelines or deployed to endpoints for online prediction. BigQuery ML runs prediction queries directly in BigQuery, while remote models can call Agent Platform models and incur charges in both services.
Low-latency serving uses specialized inference paths. GKE supports GPUs and TPUs for inference and provides GKE Inference Gateway for intelligent routing and load balancing. Pre-built Cloud AI APIs, such as Cloud Vision API, Cloud Natural Language API, Speech-to-Text API, Video Intelligence API, Document AI API, and Dialogflow API, let applications get predictions without building models from scratch.
Choosing where to run ML in Google Cloud means matching the platform to control, expertise, and scale needs. BigQuery ML brings ML to the warehouse for SQL users, Agent Platform offloads infrastructure with a fully managed path, and GKE offers deep control for customized platforms.
| Platform | What it provides | Choose when |
|---|---|---|
| BigQuery ML | In-warehouse training and prediction with SQL | Analysts want to build with SQL and BI tools without moving data or coding in Python |
| Gemini Enterprise Agent Platform | Fully managed end-to-end platform with Model Registry and pre-trained APIs | Teams want rapid MLOps and offloaded infrastructure management |
| Google Kubernetes Engine (GKE) | Managed Kubernetes with GPU and TPU support and Inference Gateway | Teams need custom pipelines, kernel-level control, large-scale training, and portability |
Exam tip: Choose BigQuery ML when the data already lives in BigQuery and the team works in SQL, and choose GKE when the workload needs custom infrastructure control and portability across Kubernetes environments.
Artificial intelligence and machine learning are related but different ideas that work together in intelligent cloud solutions. Artificial intelligence is the broader field focused on enabling machines to perform human-like tasks, such as understanding language, recognizing images, and making decisions. Machine learning is a subset of AI that uses data-driven models to learn patterns from examples instead of following only fixed rules. Together, AI defines the goal of intelligent behavior and ML provides a way to achieve it by learning from data.
A nested set diagram showing Machine Learning as a smaller ellipse fully contained inside the larger Artificial Intelligence region, with human-like AI tasks outside and data-driven learning inside.
Artificial intelligence is the field of computer science that builds systems capable of performing tasks that normally require human intelligence. These tasks include understanding spoken or written language, seeing and interpreting images, reasoning through problems, and automating routine work. An AI system can combine many capabilities, such as language understanding, vision, and decision-making, to complete a useful job. In a cloud environment, AI services allow applications to add these intelligent capabilities without building them from scratch.
Machine learning is a subset of AI in which models learn patterns from data and use those patterns to make predictions on new data. The process starts with training data flowing from collected examples to the model, then the model adjusts itself to recognize relationships in that data. After training, the trained model can analyze new inputs and predict an outcome, such as forecasting demand or classifying an image. This data-driven approach means the system improves as it receives more representative data, rather than relying only on manually written instructions.
How AI and ML contribute to intelligent cloud solutions is best seen as an overall experience powered by a learning engine. An AI application depends on one or more ML models to provide predictive and analytical capabilities, such as recommendations, anomaly detection, and automated assistance. Data flows from cloud storage and applications to the ML model for training and prediction, then the prediction flows back to the AI application where it guides an action or decision. This relationship lets organizations automate work and gain insights, with AI delivering human-like interaction and ML continuously learning patterns from changing data.
Artificial intelligence is the broader field focused on enabling machines to perform human-like tasks such as understanding language, recognizing images, and making decisions, while machine learning is a subset of AI that uses data-driven models to learn patterns from examples to make predictions on new data. Together AI defines the goal of intelligent behavior and ML provides a way to achieve it by learning from data.
The foundational ML workflow runs in order from data collection and preparation to training to evaluation and tuning to prediction, also called inference. Each stage depends on the one before it, with training depending on prepared datasets and features, evaluation depending on completed training runs, and prediction depending on a tuned model stored in a registry or dataset.
Use BigQuery ML when analysts want to build and run models with SQL where the data already lives without moving data for Python frameworks. Choose Google Kubernetes Engine (GKE) when the workload needs deep infrastructure control, very large model training, minimal-latency serving, and portability across Kubernetes environments.
A digital retail company is designing an intelligent customer engagement solution on Google Cloud. During a planning session, stakeholders want to understand how Artificial Intelligence (AI) and Machine Learning (ML) relate to one another to meet two core business objectives:
Which statement accurately distinguishes between AI and ML and explains how each contributes to this cloud solution?
Last updated