Prepare and test your skills

Prepare and test your skills

Artificial intelligence (AI), machine learning (ML), and deep learning are nested fields with different scope. AI is the broad field of systems that perform tasks such as text generation, translation, classification, and prediction. ML is the part of AI where a model learns patterns from data and makes predictions on new data, while deep learning is a specialized form of ML that trains deep neural networks with many layers.
Concentric Euler diagram showing Artificial Intelligence as the outer set, Machine Learning nested inside it, and Deep Learning nested inside Machine Learning with defining traits.
AI, ML, and deep learning differ by scope and by how they learn. Traditional AI includes classification and predictive analytics, while generative AI creates new text, images, and other content. ML creates a model from examples, so data flows from datasets into a training process and then into a stored model. Deep learning uses multi-layer neural networks that need large datasets and accelerators such as GPUs and TPUs.
ML model types differ in what kind of data they learn from. Supervised models learn from labeled examples to predict a known outcome, unsupervised models find structure in unlabeled data, and reinforcement models learn from feedback on actions. The choice depends on the data available and the business question, such as prediction, grouping, or optimization.
The ML lifecycle is the ordered process that turns data into business predictions. Data provides the examples, training builds the model from those examples, and inference uses the stored model to score new data. Training and inference use resources differently, since training pays for compute to build the model and storage for the stored model, while queries against the model pay for query compute.
A four-step process flow showing data preparation leading to training, then evaluation and tuning, then deployment and prediction, with a feedback loop for monitoring and retraining.
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. BigQuery ML brings ML to the data in the warehouse, which reduces complexity because fewer tools are required and increases speed to production because large datasets do not need to be moved for Python frameworks. Models are stored in BigQuery datasets and can be accessed through the console, query editor, bq command-line tool, BigQuery REST API, Colab Enterprise notebooks, and business intelligence tools.
Imported and remote models extend what BigQuery ML can run. Previously trained TensorFlow models can be loaded into BigQuery for prediction, TensorFlow Lite supports smaller, faster inference on mobile and edge devices, and XGBoost supports efficient tabular learning. Remote models send the inference request from BigQuery to a Gemini Enterprise Agent Platform endpoint, so queries against remote models also incur charges from the platform hosting the remote model.
The same lifecycle can run on platforms that trade control for management. Google Kubernetes Engine (GKE) (opens in a new tab) provides managed Kubernetes for large-scale training and low-latency serving with GPU and TPU flexibility. Vertex AI (opens in a new tab) is a fully managed, end-to-end AI and ML platform that offloads infrastructure management, while Cloud Run is a serverless platform for containers that can scale to zero for smaller or event-driven inference.
| Service | Level of management | Choose when |
|---|---|---|
| Google Kubernetes Engine (GKE) | Managed Kubernetes with deep control | Control, custom pipelines, very large training, and low-latency serving are needed |
| Vertex AI | Fully managed end-to-end platform | Offloading infrastructure and accelerating MLOps matter most |
| Cloud Run | Serverless containers that scale to zero | Cost-effective serving of smaller models is the goal |
ML creates business value when it connects to measurable goals defined in a business use case (opens in a new tab). A value-driven approach starts from outcomes such as increased revenue, reduced costs (opens in a new tab), faster time-to-market, and higher satisfaction, then works backward to the AI solution. Reusing datasets, features, and trained models accelerates development and improves return on investment while keeping quality, security, and compliance intact.
Artificial intelligence (AI) and machine learning (ML) on Google Cloud are delivered as managed services that help teams govern data, reuse assets, automate work, and start from pre-trained models. The platform ranges from fully managed tools to 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.
Governing and reusing data 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, using logical lakes and zones, access controls, auto data quality checks, and a universal catalog for discovery.
Reusable data products let teams share work instead of rebuilding inputs. BigQuery sharing exchanges curated assets through data exchanges that hold listings, data clean rooms that protect sensitive data, and permissions that control publishing and viewing. Agent Platform Feature Store provides a central repository for organizing, storing, and serving features for online and offline use, while unstructured data is stored in Cloud Storage and governed through object tables and vector indexes.
Starting from pre-trained models means reusing models trained on vast datasets instead of training from scratch. Training from scratch requires significant compute and time, so reuse reduces cost and speeds development. Model Garden offers Google models such as Gemini and Imagen, open models such as Gemma and Llama, and third-party models from partners such as Anthropic.
Google Cloud also provides AI APIs that add intelligence without building models:
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.
Managed notebooks are shared, interactive environments where teams explore data and iterate without managing hardware. Agent Platform Workbench is a Jupyter environment that connects to BigQuery and Cloud Storage, schedules recurring executions, and runs on managed Spark clusters. Colab Enterprise is a collaborative notebook environment with security and compliance, where sharing is controlled by Identity and Access Management (IAM) (opens in a new tab) and changes are tracked with Git.
Automating the ML lifecycle with MLOps reduces manual work through pipelines and reusable assets:
Reliable pipelines depend on versioning in Artifact Registry, reuse of features and models through Feature Store and Model Registry, and monitoring with Cloud Logging and Cloud Monitoring.
BigQuery ML lets users create, evaluate, and run models with GoogleSQL queries or the console, with models stored like tables and views. Because large datasets stay in the warehouse, fewer tools are needed and models reach production faster. BigQuery ML can also access Agent Platform models and Cloud AI APIs for tasks such as text generation and translation.
| Service | What it is | Choose when |
|---|---|---|
| Google Kubernetes Engine (GKE) | Managed Kubernetes for containerized AI/ML | Deep control, very large training, minimal latency, and portability are needed |
| Vertex AI | Fully managed end-to-end platform | Rapid development and offloaded infrastructure matter most |
| Cloud Run | Serverless container inference | Scale-to-zero and event-driven serving of smaller models is the goal |
Controlling costs depends on visibility and pilots. FinOps practices assign owners, organize resources by organization, folder, and project, apply consistent labels, and export billing data to BigQuery for analysis. Teams start with pilots, then monitor key performance indicators (KPIs) and costs, make data-driven adjustments, refine objectives, and review results with stakeholders in a feedback loop.
Exam tip: BigQuery ML brings ML to the data so analysts can build with SQL, while Agent Platform Pipelines orchestrate full training and deployment workflows.
Continue with this study guide: Define artificial intelligence (AI) and machine learning (ML).
Human-led insights and autonomous learning differ in who produces the conclusion from data. Human-led insight depends on a person who explores dashboards and reports and makes the decision. Autonomous learning depends on a trained model that learns patterns from data and improves its classifications, predictions, or generated content over time.
Human-led insights are conclusions a person draws after exploring samples, measurements, and visualizations. Exploratory data analysis (EDA) summarizes the main characteristics of a dataset, often visually, before formal modeling begins. BigQuery supports this work because data flows from BigQuery tables to governance tools and then to the person for interpretation.
Governance tools help the person ask better questions without changing who decides:
Autonomous learning is improvement that happens when a model learns relationships directly from data rather than from hand-written rules. Traditional AI learns from existing data to classify or predict, such as spam detection, defect detection, or house-price prediction. Generative AI learns from training data and then generates new text, images, audio, or code, and models that handle more than one modality such as text and images are multimodal.
A model moves through an ordered training workflow in Vertex AI, where each step is recorded as an execution:
Hyperparameter tuning runs multiple trials with different values to maximize accuracy, while early stopping ends training when validation loss starts to increase. At the far end of autonomy, an agent plans and executes actions on its own to pursue a goal.
| Approach | What produces insight | How it improves | Who decides |
|---|---|---|---|
| Human-led analysis | Queries, profiles, and visual summaries a person reads | Analyst asks new questions manually | Person interprets result |
| Autonomous learning | Classifications or generated content from a trained model | Retraining, tuning, and feedback | Model produces output while people govern it |
Exam tip: Traditional AI classifies or predicts from historical data, while generative AI creates new content such as text, images, or code.
Retrospective analysis is backward-looking explanation, while predictive modeling is forward-looking forecasting. Retrospective analysis uses descriptive analytics in BigQuery to aggregate stored records into trends and explanations, such as why sales dropped last quarter. Predictive modeling with BigQuery ML (opens in a new tab) trains where the data lives and then runs inference on new data to forecast demand or flag fraud.
A BigQuery ML workflow runs in order from train to evaluate to predict. Explainable AI helps people understand why the model made its decisions. Predictive tasks include time-series forecasting and anomaly detection, classification and regression, clustering and recommendations, and contribution analysis.
| Approach | Core question | Example outcome |
|---|---|---|
| Retrospective analysis | What happened and why | Last month's return rate rose because of one defective batch |
| Predictive modeling | What will happen next | Next month's demand forecast and flagged risky transactions |
Structured data in rows, columns, and tables supports business intelligence reporting with fixed rules. Examples include transactional databases and BigQuery tables, where filters and thresholds produce dashboards and quality checks. Unstructured data such as text, images, audio, video, and code has no fixed table structure and requires AI that learns patterns from large training sets, as described in guidance on when to use generative or traditional AI (opens in a new tab).
Rule-based logic follows conditions people define in advance, while pattern recognition trains on historical examples and improves with feedback. Traditional AI classifies categories or predicts numbers from historical patterns, while generative AI creates new summaries, captions, documents, and code and explains unstructured content.
Pattern discovery and unstructured data insights are problems where machine learning succeeds when fixed rules do not. Pattern discovery finds hidden structure without predefined labels, while unstructured insights turn speech, text, images, video, and documents into usable information. Together they let a business group customers by behavior and extract meaning from sources that do not fit in a database.
Clustering is a descriptive ML approach that finds natural groupings in unlabeled data without being told the correct answers. Customer data flows into the clustering model, and the model returns groups of similar customers based on shared behavior. This differs from classification, which assigns data to known types, and from prediction, which forecasts a future outcome. For example, contact-center analytics can detect themes in calls and chats that were not defined ahead of time.
Pre-trained models turn language and media into structured information without custom training. Audio flows to Speech-to-Text, which returns recognized text, and text flows to Cloud Natural Language API, which returns categories, entities, and sentiment. Text can then flow to Cloud Translation or Text-to-Speech, while conversational tools such as Dialogflow CX, Dialogflow ES, and Agent Assist support virtual and human agents.
| API | Input source | What it returns |
|---|---|---|
| Speech-to-Text | Spoken audio | Recognized text for analysis |
| Cloud Natural Language API | Plain text or HTML | Categories, entities, and understanding |
| Cloud Translation | Text needing translation | Translated text |
| Text-to-Speech | Written text | Natural-sounding speech audio |
Pre-trained vision models process pixels at scale. An image flows to Cloud Vision, which returns labels and face detections, while video flows to Video Intelligence API, which returns annotation across frames. Choose Cloud Vision when the question is what is in a single image, and choose Video Intelligence API when the question is what happens across video.
Exam tip: Use Speech-to-Text to turn voice into text for analysis, and use Text-to-Speech to turn text into voice for playback.
Document AI (opens in a new tab) is a document processing platform that transforms unstructured documents into structured data. A document file flows to a Document AI processor, the processor applies an ML model, and the service returns Document objects with extracted fields and entities. Document AI can digitize with optical character recognition (OCR), extract entities and key-value pairs such as Name: Jill Smith, classify types, and split files.
Processors fall into Digitize, Extract, and Classify categories, and each Google Cloud project must create its own processor instances. Results can flow to Cloud Storage, BigQuery, and search systems. Choosing the right processor (opens in a new tab) and processing flow (opens in a new tab) follow this order:
Classification assigns data to discrete categories by learning from historical examples. Training data provides labeled examples of each category, such as emails marked as spam or not spam, so the model learns patterns that separate them. Once trained, the model classifies new emails, transactions, or images without a hand-written rule for every case.
Binary classification assigns two categories, while multi-class classification assigns three or more. Common scenarios include spam filtering, labeling good versus defective products for inspection, detecting fraudulent transactions, and sentiment analysis of reviews as positive, neutral, or negative. In each flow, data moves from the source record to the trained model, and the model returns a label that the workflow acts on.
| Task type | Output type | Example |
|---|---|---|
| Classification | Discrete category label | Email is spam or not spam |
| Regression | Continuous numerical value | Predicted house price from features and location |
Regression estimates a continuous numerical outcome such as inventory demand, future revenue, or optimal price. It is predictive rather than descriptive because it estimates a future or unknown number instead of summarizing the past. In Google Cloud, BigQuery ML supports regression with SQL or BigQuery DataFrames.
The BigQuery ML regression workflow runs in fixed order:
Evaluation uses ML.EVALUATE to compare predictions with actuals, including mean absolute error, mean squared error, and R-squared score. The R-squared score ranges from 0, meaning no variability explained, to 1, meaning all variability explained. Inference runs as batch inference for large asynchronous sets or online inference for fast real-time responses, though forecasting models support only batch inference.
Continue with this study guide: Discuss the types of problems that ML can solve.
Scaling decision-making through automation means using software to handle repetitive, high-volume work the same way every time. Instead of a person triggering each step manually, an automated workflow moves data from a source system to each destination in order. This increases efficiency because the same checks run consistently even as volume grows.
Orchestration is the automated coordination of several systems, while scheduling is the automated triggering of that work. A data pipeline is an orchestration of transformations described by a directed acyclic graph (DAG). When one pipeline depends on others, each pipeline becomes a sub-DAG inside a larger orchestration DAG.
Dependencies define how pipelines connect. A fan-in dependency occurs when pipelines A, B, and C merge into downstream pipeline D. A fan-out dependency occurs when pipeline D triggers downstream pipelines E, F, and G. If an upstream pipeline that produces shared data fails, dependent pipelines wait or cannot run at all.
Optimization follows an ordered progression from moving as is, to parallelizing by copying a shared sub-DAG into each dependent pipeline, to extracting common tasks into an independent pipeline that fans out. For example, a profit-and-loss pipeline that calculates monthly sales needed by marketing can become a shared pipeline that triggers both. Parallelizing creates duplication overhead, while extracting creates a cleaner DAG.
Exam tip: Orchestration coordinates what happens across systems, while scheduling triggers when that coordinated work starts.
Automating business decisions connects systems so an event in one system produces consistent updates in others. Application Integration links a customer relationship management (CRM) application such as Salesforce to BigQuery and ServiceNow. When a new opportunity is created in Salesforce, the integration retrieves details, validates them in BigQuery, and writes back a tracking order number.
Because the same validation and update steps run for every opportunity, teams see consistent results without manual handoffs.
Unstructured data such as text, images, audio, and video does not fit in rows and columns, so ML turns it into structured predictions, classifications, and embeddings. A BigQuery ML remote model (opens in a new tab) is a reference inside BigQuery that points to a model hosted in Vertex AI Inference. You create the reference with CREATE MODEL and run inference with ML.PREDICT, which sends data from BigQuery tables to Agent Platform and returns structured predictions to BigQuery.
Use remote models when models are already hosted by Agent Platform, when GPU support is needed, or when connecting BigQuery to generative models and Cloud AI APIs (opens in a new tab) for text and embeddings. A model is only as good as its inputs, so inaccurate prompts or data cause poor outputs.
Batch prediction and online prediction differ by size and speed. Built-in BigQuery ML inference is optimized for large-scale batch prediction inside BigQuery, while an Agent Platform endpoint provides faster online prediction for small, immediate requests. Models can be managed in Agent Platform for MLOps and Feature Store benefits, or exported to Cloud Storage object tables (opens in a new tab) for hosting elsewhere.
| Prediction option | Where it runs | Choose when |
|---|---|---|
| BigQuery ML built-in inference | Inside BigQuery | Large-scale batch prediction on tables |
| Agent Platform endpoint | On Agent Platform endpoint | Faster online prediction for immediate requests |
Leveraging large datasets means training on massive examples to find trends beyond human capacity. Large datasets are stored in BigQuery, processed with Dataflow and Apache Beam, and used to train models on Agent Platform. More examples and wider variety improve recognition, while labels with too few examples cannot be predicted confidently.
Evaluation metrics guide debugging and cost optimization (opens in a new tab). Low precision and recall can mean more data or cleaner labels are needed, while perfect scores can mean test data is too easy and the model may not generalize. Improvements include adding varied examples, removing rare labels, teaching nuances between similar labels, adding examples near decision boundaries, using custom data splits to avoid overfitting, reviewing confusion matrices, and allowing longer training with more node hours. A business use case (opens in a new tab) ties these improvements to revenue, cost, and experience goals.
High-quality, accurate data is essential because a model learns patterns from whatever examples it receives. If training data is inaccurate, incomplete, noisy, biased, or unrepresentative, predictions will be unreliable or unfair for missing cases. Diverse data that captures many conditions helps the model generalize to new or less common examples after deployment.
Unrepresentative data creates bias when the training set does not match real conditions where the model will be used. Three sourcing decisions shape bias: feature selection, data volume, and variation. A feature is an input attribute used for training, and features are how the model identifies patterns, so features must be relevant. Relevant features carry true signal, while irrelevant, noisy, or sensitive features teach misleading patterns.
Data volume scales with complexity, and more examples generally improve outcomes. Recommended minimums are 50 rows times the number of features for classification, 5,000 rows times the number of features plus 10 unique values in the time-series identifier times features for forecasting, and 200 times features for regression. Binary classification needs less data than multi-class because choosing between two classes is simpler. Missing or inconsistent values also reduce usefulness, so quality must be reviewed before training.
The Garbage In, Garbage Out (GIGO) principle states that bad inputs produce unreliable predictions because the model does not automatically correct bad data. Poor data enters through features, such as a fraud model that needs buyer, seller, amount, date, and history. A random identifier adds noise, while a future payment used to predict subscription leaks unavailable information.
| Data problem | Effect on model | Prevention |
|---|---|---|
| Missing or incomplete values | Less useful training; optimal results need complete rows | Correct gaps; leave blank only if nullable; keep consistent forecasting intervals |
| Noise, errors, or inconsistency | Model learns wrong patterns | Correct errors; unify spelling and formatting |
| Data leakage | Strong test scores but poor production | Remove features unavailable at serving time |
| Training-serving skew | Poor production quality | Use only features available in same form at serving |
| Biased data | Poor generalization and unfair decisions | Capture diversity; review fairness |
Data leakage happens when training uses a feature unavailable at serving, such as a future subscription payment to predict sign-up. Training-serving skew happens when training inputs differ from serving inputs, such as training on weekly temperatures to predict hourly temperatures. Both are prevented by understanding each field, flagging high correlations, and providing only serving-time features, as emphasized in guidelines for high-quality ML (opens in a new tab).
Exam tip: Biased or unrepresentative training data is a data problem, not just a model problem; if a model behaves unexpectedly, check the data first.
Relevant, sufficient, and diverse data gives the model the right signals. For retail marketing, relevant features include items, brands, prices, purchase counts over day, week, month, and year, stock levels, promotions, and shopper demographics. Diverse data covers variation, since a model trained only on winter purchases will not predict summer preferences well.
Clean, consistent, and correctly labeled data ensures accurate signals. Agent Platform can handle some missing values and impute for forecasting, but complete rows work better. Cleanup focuses on missing values, consistent forecasting intervals, and consistent spelling and formatting. After import, each column's variable type and nullability should be reviewed, since nullability controls whether missing values are allowed. Unlabeled images can be annotated in the console with multiple annotation sets for different tasks.
A dataset split into training, validation, and test sets gives each part a different role. If splits are not specified, Agent Platform uses 80% for training, 10% for validating, and 10% for testing, though defaults vary by model type and manual splits allow more control.
Using a separate validation set prevents overfitting to training data. Manual splits help when specific examples must be in a certain phase. After training, performance is summarized through evaluate functions, and unexpected behavior before or after deployment usually means returning to the data to improve completeness, diversity, or correctness.
Bias mitigation and fairness strategies are practices that keep automated systems from producing unfair outcomes. Bias mitigation addresses two entry points: the data used to train and evaluate models, and the algorithms and governance processes around those models. Google Cloud pairs a governed ML environment that reports fairness metrics and explainability with ethics policies, security controls, and data governance that improves quality and limits access to sensitive data.
Sources of bias include problems in training data and weaknesses in design and governance. On the data side, models produce unfair results when they are not trained on relevant data, when quality is poor, and when data is duplicated and managed inconsistently. On the algorithm and process side, bias persists when teams cannot see how a model reaches its output and when no ethics or security policies govern development. Reducing bias therefore requires better data feeding the model and stronger measurement, transparency, and governance around the model.
Fairness metrics and explainability are built-in checks that show how a model decides and whether outcomes are fair. Gemini Enterprise Agent Platform is the environment for developing and deploying models securely and under governance. It provides model explainability, which shows how a model reaches decisions, and fairness metrics, which measure whether outcomes are fair across groups. Because these features live inside the lifecycle, fairness checks happen during development rather than as an afterthought.
Data governance controls that reduce bias manage quality, access, and protection:
Together these controls improve relevance and quality while preventing unsafe use of sensitive data.
Explainability is the ability to show in human-understandable terms why a model made a specific prediction. A model takes input data and produces an inference, and explainability adds an explanation that connects that inference back to the inputs and behavior that shaped it. This connection is the foundation of responsible AI because it makes decisions trustworthy and open to review.
Vertex Explainable AI is a feature of Vertex AI that works alongside a deployed model to interpret inferences. When the model predicts for a new input, Vertex Explainable AI also produces information that helps a person interpret the result. Feature importance shows which inputs most influenced a specific inference, while model behavior describes how the model generally responds across many inferences. Together they let teams confirm the model reasons from relevant data rather than noise or bias.
Transparent decisions build trust and support compliance. When a lender, doctor, or customer can see why a case was approved or flagged, the decision moves from blind trust to informed trust. The same explanation lets auditors trace how inferences were reached and verify consistency and accountability.
Exam tip: Fairness metrics and model explainability come from the governed ML environment, while data quality, access control, and sensitive-data protection come from the data governance layer.
Google's AI principles require building AI that is socially beneficial, safe, and trustworthy. Google Cloud turns those requirements into prescriptive guidance organized by workload so teams can plan, design, build, and deploy with governance and grounding.
Trusted agentic AI solves open-ended problems through autonomous planning while staying grounded in real data from sources such as Maps and Calendar. Generative AI creates summaries and new content with controlled operation, often using retrieval-augmented generation (RAG) to ground answers in trusted sources and vector stores such as AlloyDB for PostgreSQL. MLOps provides the repeatable foundation for building and running models, while AI infrastructure such as Cloud Storage FUSE and Managed Lustre matches performance, cost, and scalability to each lifecycle stage.
Measuring fairness after deployment uses different metrics by AI type: predictive AI by accuracy and precision, generative AI by adoption, satisfaction, and content quality, and computer vision AI by accuracy. Monitoring follows an ordered cycle of monitoring KPIs, making data-driven adjustments, refining objectives, and establishing a stakeholder feedback loop.
Continue with this study guide: Discuss the importance of explainable and responsible AI
Artificial intelligence (AI) is the broad field of systems that perform tasks such as text generation, translation, classification, and prediction, machine learning (ML) is the part of AI where a model learns patterns from data and makes predictions on new data, and deep learning is a specialized form of ML that trains deep neural networks with many layers. Deep learning uses multi-layer neural networks that need large datasets and accelerators such as GPUs and TPUs.
Training builds the model from training data and stores it, while inference uses the stored model to score new data and generate predictions. The ML lifecycle runs in order from data preparation to training to evaluation and tuning to deployment and prediction, where training pays for compute to build the model and queries against the model pay for query compute.
BigQuery ML lets SQL practitioners create and run ML models with GoogleSQL queries where the data lives, which reduces complexity and increases speed to production because large datasets do not need to be moved. Google Kubernetes Engine (GKE) provides managed Kubernetes for large-scale training and low-latency serving with GPU and TPU flexibility, Vertex AI is a fully managed end-to-end AI and ML platform that offloads infrastructure management, and Cloud Run is a serverless platform for containers that can scale to zero for smaller or event-driven inference.
High-quality, accurate data is essential because a model learns patterns from whatever examples it receives, so if training data is inaccurate, incomplete, noisy, biased, or unrepresentative, predictions will be unreliable or unfair. This is the Garbage In, Garbage Out (GIGO) principle that bad inputs produce unreliable predictions, including data leakage where training uses a feature unavailable at serving and training-serving skew where training inputs differ from serving inputs.
Last updated