Prepare and test your skills

Prepare and test your skills

Artificial intelligence (AI), machine learning (ML), and deep learning are related fields with different scope. AI is the broad field of systems that perform tasks such as text generation, machine translation, classification, and prediction. Machine learning (ML) is the part of AI where a model learns patterns from data and then uses those patterns to make predictions on new data. Deep learning is a specialized form of ML that trains deep neural networks with many layers.
AI, machine learning (ML), and deep learning differ by scope and method. AI covers any system that performs intelligent tasks, including generative tasks like text generation and traditional tasks like classification and predictive analytics. ML achieves those tasks by creating a model from examples, so data flows from datasets into a training process and then into a stored model. Deep learning achieves them with multi-layer neural networks that need large datasets and accelerators, which is why TensorFlow (opens in a new tab) optimizes models for training and inference and TensorFlow Lite optimizes existing TensorFlow models for reduced size and faster inference on mobile and edge devices.
Three nested ovals showing AI as the outer set, Machine Learning inside AI, and Deep Learning inside Machine Learning, with examples for each.
ML model types include supervised, unsupervised, and reinforcement learning, which differ in how the model learns from data. Supervised models learn from labeled examples, unsupervised models find structure in unlabeled data, and reinforcement models learn from feedback on actions. The choice of model type depends on the data available and the business question, such as prediction, grouping, or optimization. A model selection guide helps map ML models to the actions a user wants to accomplish.
The ML lifecycle is the ordered process that turns data into business predictions. Data provides the examples the model learns from, training creates the model from those examples, and inference uses the trained model to score new data. Training builds and stores the model, while inference runs queries against the stored model to generate predictions.
A four-step process flow from data preparation to training to evaluation and tuning to deployment and prediction, showing training builds the stored model and inference queries it for predictions.
Because training creates the model and inference applies the model, the two stages use resources differently. Training is charged for the compute used to build the model and storage is charged for the stored model. Queries against models run in BigQuery, and remote models also incur charges from the platform hosting the remote model.
Exam tip: Training builds and stores the model, while inference runs queries against the stored model to generate predictions.
BigQuery ML (opens in a new tab) is a Google Cloud capability that lets SQL practitioners create and run ML models with GoogleSQL queries or the console, without programming in Python or Java. BigQuery ML brings ML to the data in the warehouse, which reduces complexity because fewer tools are required and increases speed to production because moving and formatting large amounts of data for Python frameworks is not required. Models can also connect outward, since BigQuery ML can access Gemini Enterprise Agent Platform models and Cloud AI APIs for AI tasks and can register BigQuery ML models to Model Registry for deployment to online prediction endpoints.
You can work with BigQuery ML capabilities by using the following interfaces, which all operate on the same models stored in BigQuery datasets:
Imported and specialized models extend what BigQuery ML can run. Previously trained TensorFlow models can be loaded into BigQuery for prediction, TensorFlow Lite supports deployment on edge devices, and XGBoost implements algorithms under the gradient boosting framework for highly efficient tabular learning. Remote models call Agent Platform models from BigQuery, so the query starts in BigQuery and the inference request flows from BigQuery to the Agent Platform endpoint.
The same lifecycle can run on different Google Cloud platforms that trade control for management. Google Kubernetes Engine (GKE) (opens in a new tab) provides a scalable, managed Kubernetes platform for training very large models and serving inference with low latency, high throughput, and accelerator flexibility across GPUs and TPUs. GKE depends on the customer to manage container orchestration, while fully managed options offload that work for faster time to value or scale to zero.
| Service | Level of management | Choose when |
|---|---|---|
| Google Kubernetes Engine (GKE) | Managed Kubernetes with deep control and portability | Control, custom pipelines, large-scale training, and low-latency serving are needed |
| Vertex AI | Fully managed, end-to-end AI and ML platform | Offloading infrastructure management and accelerating MLOps matter most |
| Cloud Run | Serverless platform for containers that can scale to zero | Serving smaller models or event-driven inference cost-effectively is the goal |
ML creates business value when the lifecycle connects to measurable goals. A business-value approach starts from outcomes such as increased revenue or reduced costs, faster time-to-market or quicker issue resolution, and increased satisfaction or improved retention, then works backward to the AI solution. Reusing datasets, features, pipeline templates, and trained models across projects accelerates development, reduces infrastructure cost, and frees resources, so governance and reuse directly improve return on investment while keeping data quality, security, and compliance intact.
Machine learning solutions on Google Cloud can be built, trained, deployed, and scaled through different services depending on the data and the skills of the people using it. BigQuery ML (opens in a new tab) lets data analysts create and run ML models with SQL inside BigQuery, so the data stays in the data warehouse. Vertex AI is a fully managed, unified AI development platform for building and using generative AI. Agent Platform, the Gemini Enterprise Agent Platform, is the end-to-end platform for AI and ML in Google Cloud, and Cloud AI APIs provide pre-trained AI tasks such as text generation and machine translation.
BigQuery ML lets data analysts create and run machine learning models by using GoogleSQL queries or the Google Cloud console. Models are stored in BigQuery datasets, similar to tables and views. BigQuery ML brings ML to the data instead of moving data to Python or Java frameworks, which reduces complexity and speeds production. Data analysts can use existing SQL skills and business intelligence tools to build and evaluate models.
Data analysts can work with BigQuery ML through several interfaces:
BigQuery ML offers advantages over other approaches to using ML with a cloud data warehouse. It democratizes ML by empowering data analysts to build and run models. It does not require programming in Python or Java, since training and AI access use SQL. It increases speed of model development and innovation by removing the need to move data from the data warehouse. A model can also use data from multiple BigQuery datasets for training and prediction.
BigQuery ML supports internally trained models, externally trained models, remote models, and imported models. Imported models can be TensorFlow, TensorFlow Lite, or XGBoost models. The CREATE MODEL statements for imported models do not process any bytes and do not incur BigQuery charges. Remote models make calls to Agent Platform models, so queries against remote models also incur charges from Agent Platform.
BigQuery ML also lets data analysts access Gemini Enterprise Agent Platform models and Cloud AI APIs to perform AI tasks like text generation or machine translation. This means pre-trained AI capabilities can be used without building a model from scratch. A model selection guide helps map ML models to the actions a user wants to accomplish.
BigQuery ML integrates with Agent Platform (opens in a new tab), the end-to-end platform for AI and ML in Google Cloud. Data analysts can register BigQuery ML models to Model Registry in order to deploy these models to endpoints for online prediction. This connects model development in BigQuery to deployment and scaling on Agent Platform, so the flow moves from the dataset in BigQuery to Model Registry and then to the online prediction endpoint.
Colab Enterprise notebooks in BigQuery let data analysts perform ML workflows using SQL, Python, and other popular libraries and languages. Notebooks support experimentation and operationalization by combining the data warehouse with flexible coding environments. Vertex AI is also listed as a fully managed, unified AI development platform for building and using generative AI.
Data preparation and governance help ensure that ML models are trained on relevant, high-quality data. A data governance framework with clear policies, procedures, and roles manages data throughout its lifecycle and improves quality, security, utilization, and reduces redundancy. Google Cloud provides services to implement this framework.
Reusing datasets and features across multiple ML projects provides significant efficiency and cost benefits. It avoids redundant data engineering and feature development, accelerates model development, reduces infrastructure costs, and frees up resources. Data clean rooms manage access to sensitive data for AI and ML projects, so sharing can happen without losing control.
MLOps practices reduce costs for technology and personnel by automating ML activities and reducing duplication. Automation helps avoid duplication of ML activities and reduces the workload for data scientists and ML engineers. Pipelines define dependencies between steps, manage data flow, and automate the execution of the entire workflow.
Google Cloud offers several orchestration options for ML workflows:
For model evaluation and tuning, begin an AI project with experiments before scaling up. Track metadata such as dataset version, model parameters, and model type. Google Cloud services can streamline and automate evaluation and tuning to reduce time to production and resources. Versioning tracks and controls different versions of pipelines and components to ensure reproducibility and auditability. Monitoring pipeline execution uses Cloud Logging and Cloud Monitoring to identify and address issues.
BigQuery ML charges for the compute resources used to train models and run queries against models. Queries against models always run in BigQuery and use BigQuery compute pricing. Remote models also incur charges from Agent Platform because they make calls to Agent Platform models. Storage used by trained models is charged using BigQuery storage pricing.
BigQuery ML is supported in the same regions as BigQuery. Queries that use BigQuery ML functions and CREATE MODEL statements are subject to BigQuery query job quotas and limits, in addition to BigQuery ML-specific limits. BigQuery ML is not available in the Standard edition.
Exam tip: BigQuery ML brings ML to the data: training and prediction run in BigQuery, while remote models call Agent Platform models and incur Agent Platform charges.
A common ML solution on Google Cloud can flow from data preparation to deployment. The flow connects governance, development, deployment, and automation into one chain where each stage depends on the previous stage.
A concrete example is a vision analytics solution. A Dataflow pipeline processes large-scale image files with Cloud Vision API and stores the results in BigQuery. BigQuery ML can then use those results to train pre-built models. This shows how data processing, pre-trained APIs, and BigQuery ML connect in one flow.
Machine learning (ML) creates business value when it supports a measurable business goal rather than existing on its own. The value shows up as increased efficiency and productivity, lower costs, better customer experiences, and competitive advantage. To capture that value, an organization starts from the desired business outcome and works backward to the ML solution, then measures results with return on investment (ROI) and model-level metrics.
An AI or ML use case should be defined with a business value-driven decision approach, meaning the business goal is chosen first and the technology is chosen to serve it. The process moves through five ordered steps where each step depends on the decision made in the previous step.
Return on Investment (ROI) is a key measure of AI project success. ROI can be tracked through direct financial gains such as increased revenue or reduced costs, operational efficiency such as faster time-to-market or quicker issue resolution, and customer experience such as higher satisfaction scores or improved retention. Business constraints such as security, privacy, industry compliance, and country regulatory requirements must also be identified during goal setting.
Different kinds of AI produce value that is measured in different ways, so the metric must match the goal. Business-level metrics capture the outcome the organization cares about, while model-level metrics capture whether the AI itself is performing well. Generative AI use cases (opens in a new tab) require special attention to adoption, satisfaction, and content quality, not just technical correctness.
| Metric category | What it measures | Examples |
|---|---|---|
| Direct financial | Money gained or saved | Increased revenue, reduced costs |
| Operational efficiency | Speed of business processes | Faster time-to-market, quicker issue resolution |
| Customer experience | How customers respond | Satisfaction scores, retention |
| Industry-specific | Outcomes tied to a sector | Retail: revenue lift and churn; healthcare: patient time and patient outcomes; finance: fraud reduction |
| Predictive AI | Model correctness | Accuracy and precision |
| Generative AI | How the output is received | Adoption, satisfaction, content quality |
| Computer vision AI | Model correctness | Accuracy |
Exam tip: Predictive AI and computer vision AI are measured by accuracy and precision, while generative AI is measured by adoption, satisfaction, and content quality.
ML value is not fixed at launch; it is improved through a repeating optimization cycle that ties monitoring back to business objectives. To validate ROI hypotheses, an organization starts with pilot projects and then repeats the following cycle to keep the AI application aligned with business goals.
This cycle keeps the AI application aligned with business goals and lets the organization adjust before costs or performance drift too far. Continuous monitoring feeds adjustments, adjustments lead to refined objectives, and stakeholder review starts the next round of monitoring.
ML operations (MLOps) is a set of practices that automate and streamline the ML lifecycle from model development to deployment and monitoring, and its primary benefit is a reduction in technology and personnel costs. Common MLOps practices that reduce cost include version control for code, data, and models, continuous integration and continuous delivery (CI/CD) pipelines, observability, periodic retraining, and automated testing.
Effective data management and governance also drive cost savings. A data governance framework with defined policies, procedures, and roles improves data quality, enhances security, improves utilization, and reduces redundancy. Reusing datasets and features across ML projects avoids redundant data engineering, accelerates model development, and reduces infrastructure costs. BigQuery ML (opens in a new tab) supports this by letting SQL practitioners build and run ML models with GoogleSQL queries or the Google Cloud console, bringing ML to the data instead of moving data to ML frameworks, which reduces complexity and increases speed to production.
Artificial Intelligence (AI) is the broad field of systems that perform tasks such as text generation, machine translation, classification, and prediction. Machine Learning (ML) is the part of Artificial Intelligence (AI) where a model learns patterns from data and then uses those patterns to make predictions on new data. Deep learning is a specialized form of Machine Learning (ML) that trains deep neural networks with many layers.
Training builds and stores the model from data examples, while inference uses the stored model to score new data and generate predictions. The Machine Learning (ML) lifecycle runs in order from data preparation to training to evaluation and tuning to deployment, where queries against the model run where the data lives.
BigQuery ML lets analysts create and run Machine Learning (ML) models with GoogleSQL queries or the console where the data lives, without programming in Python or Java. It brings Machine Learning (ML) to the data in the warehouse, which reduces complexity because fewer tools are required and increases speed to production because moving and formatting large amounts of data is not required.
Machine Learning (ML) creates business value when the lifecycle connects to measurable goals such as increased revenue or reduced costs, faster time-to-market or quicker issue resolution, and increased satisfaction or improved retention. A business-value approach starts from those outcomes and works backward to the Artificial Intelligence (AI) solution, then tracks return on investment through direct financial, operational efficiency, and customer experience metrics.
A financial services organization is developing a machine learning system to identify fraudulent transactions.
The project team established the following workflow requirements:
Which machine learning paradigm is used during model training, and which lifecycle stage represents generating predictions on new data?
Last updated