Prepare and test your skills

Prepare and test your skills

Assessing human-led insights versus autonomous learning means deciding whether a person or a model produces the conclusion from data. Human-led insight depends on a person who explores data, reads dashboards and reports, and makes the decision. Autonomous learning depends on a machine learning model that learns patterns from data exposure and improves its classifications, predictions, or generated content over time.
Human-led insights from data analysis and business intelligence are conclusions that a person draws after exploring samples, measurements, and visualizations. Exploratory data analysis (EDA) summarizes the main characteristics of a dataset, often with visual methods, to show what the data can tell before formal modeling begins. EDA is especially useful when a dataset is first received and the team does not yet know its quality or patterns. In Google Cloud, BigQuery (opens in a new tab) supports this human-led work because data flows from BigQuery tables to governance tools and then to the person for interpretation. The governance tools help the person ask better questions without changing who decides:
Autonomous learning from machine learning models is improvement that happens when a model learns relationships directly from data rather than from hand-written rules. Machine learning is a subfield of artificial intelligence that uses a statistical and data-driven approach to create AI. Traditional AI (opens in a new tab) learns from existing data to classify information or predict future outcomes, such as classifying email as spam or not spam, detecting defective products in images, or predicting house prices. Generative AI learns from training data and then generates new data with similar characteristics, such as text, images, audio, or code, and models that work with 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 in Vertex AI runs multiple trials with different hyperparameter values within set limits to maximize predictive accuracy. Early stopping ends training when loss on a validation dataset starts to increase, which means generalization is getting worse. At the far end of autonomy, an agent is software that plans and executes a series of actions on its own to pursue a goal, even in novel situations.
Combined human review and autonomous learning is a feedback loop where people improve the data that models rely on and models improve the information people review. Knowledge Catalog is a Gemini-powered context layer that builds a dynamic context graph from structured and unstructured data to help teams discover assets, verify data quality, and ground generative AI applications to reduce hallucinations. Data stewards provide human-in-the-loop reviews of AI-generated descriptions, standardize business vocabularies with glossaries, and attach domain-specific context to catalog entries. Explainable AI in Vertex AI then gives people tools to understand model inferences, including feature importance and model behavior. Anti Money Laundering AI (AML AI) (opens in a new tab) shows how the loop works in practice because training data flows from core banking data and suspicious activity information in the customer's Google Cloud environment to AML AI, and risk scores flow back to investigators. AML AI produces monthly risk scores with explainability output to support alerting and investigation. Investigation results are then fed back into AML AI to update models and risk scores, so human decisions change future model output.
The two approaches differ in who produces insight, how improvement happens, and who decides:
| Approach | What produces the insight | How it improves | Who decides |
|---|---|---|---|
| Human-led analysis | Queries, profiles, quality checks, and visual summaries that a person reads | An analyst asks new questions and explores new data manually | The person interprets the result and makes the decision |
| Autonomous learning | Classifications, predictions, or generated content from a trained model | Retraining on new data, hyperparameter tuning, and feedback such as investigation results | The model produces the output while people govern, monitor, and explain it |
Exam tip: Traditional AI classifies or predicts from historical data, while generative AI creates new content such as text, images, or code.
Contrasting retrospective analysis with predictive modeling means comparing backward-looking explanation with forward-looking forecasting. Retrospective analysis uses descriptive analytics and business intelligence to summarize what happened and why. Predictive modeling uses machine learning to learn patterns from that history automatically and forecast what is likely to happen next.
Retrospective analysis of historical data is descriptive analytics that aggregates stored records into trends, comparisons, and explanations. The analysis starts with data already stored in systems such as BigQuery, and queries, dashboards, and reports summarize that data into context for decisions. Because the output describes known events, it helps people spot problems, understand changes in behavior, and decide what to investigate further. Choose retrospective analysis when the question is about the past, such as why sales dropped last quarter or which product had the highest return rate.
Predictive modeling for future outcomes is machine learning that trains a model on historical data so the model can run inference on new data. BigQuery ML (opens in a new tab) keeps training data in BigQuery tables, so the model learns where the data already lives and then produces predictions on new serving data without requiring data movement. SQL practitioners can build and evaluate models with existing SQL tools through the Google Cloud console, the bq command-line tool, the REST API, or Colab Enterprise notebooks. This approach automates pattern discovery that is difficult to find by manual review. A BigQuery ML workflow moves through an ordered lifecycle where each stage depends on the one before it:
BigQuery ML supports predictive tasks that go beyond historical reporting:
TimesFM, ARIMA_PLUS, and ARIMA_PLUS_XREG modelsAnomaly detection (opens in a new tab) shows how prediction differs from simple reporting because it identifies deviations from a baseline, such as a sharp rise in product return rate that might indicate a defect or fraud. If labeled examples exist, the ML.PREDICT function uses supervised models to flag future cases. If labels do not exist, AI.DETECT_ANOMALIES or ML.DETECT_ANOMALIES uses unsupervised models to find unusual patterns automatically.
Dependencies between retrospective and predictive approaches describe how backward-looking data and forward-looking models need each other in order. Predictive modeling depends on retrospective data because historical records provide the training examples the model learns from. Retrospective analysis then gains new value when predictions are compared with what actually occurred, which creates better historical context for the next decision. The choice depends on the core question being asked:
| Approach | Core question | How it works | Example outcome |
|---|---|---|---|
| Retrospective analysis | What happened and why | Aggregates and summarizes past data into reports and insights | Last month's return rate rose because of one defective product batch |
| Predictive modeling | What will happen next | Trains a model on past data and runs inference on new data | Next month's demand forecast and a flag for transactions likely to be fraudulent |
Exam tip: Choose descriptive analytics when you need historical context for what happened, and choose machine learning when you need an automated forecast of future trends from that history.
Evaluating data types and processing methods means matching the kind of data to the right kind of processing. Structured data in rows, columns, and tables supports business intelligence reporting with fixed rules. Unstructured data such as text, images, audio, video, and code requires artificial intelligence that can find patterns and learn from examples.
Structured data for business intelligence reporting is information stored in organized rows, columns, and tables that BI tools can query directly. Examples include tabular data, transactional databases, and tables in BigQuery, Google Cloud's data warehouse for analytics at scale. Because the format is consistent, teams can apply rule-based logic such as filters, thresholds, and calculations to produce descriptive reports, dashboards, and quality checks. Knowledge Catalog helps manage these environments by holding business, technical, and operational metadata in a central inventory for discovery, profiling, and data quality monitoring.
Unstructured data requiring artificial intelligence is content with no fixed table structure that gains meaning only when a model learns patterns from large training sets. Models that work with more than one kind of data are called multimodal (opens in a new tab) models, and they can perform tasks such as image captioning, video summarization, audio transcription, and text classification. Knowledge Catalog bridges both worlds by building a dynamic context graph from structured and unstructured data, so an AI tool can link a transactional record to a related product photo in Cloud Storage without manual pipelines. Generative AI (opens in a new tab) is especially strong at working with this unstructured content: The main uses form a set of peer capabilities that all start from natural language or mixed media input:
Rule-based logic compared with pattern recognition is a comparison between following instructions people define and learning relationships from data. Rule-based logic follows conditions set in advance, such as if a value passes a limit then flag it for review, and it works well for traditional data analysis that gains understanding through samples, measurement, and visualization before a model is built. Pattern recognition in machine learning works differently because the model trains on historical examples, makes predictions on new data, and then uses results to update and improve over time. First the model trains on labeled examples, then it predicts on new data, and only after that feedback does the model improve. Traditional AI uses pattern recognition to classify categories or predict numbers from historical patterns. Classification assigns a category, such as labeling email as spam or not spam or identifying good versus defective products for manufacturing inspection. Regression predicts a continuous number, such as predicting house prices from house features and location. Generative AI expands beyond classification and prediction to create new summaries, uncover hidden correlations, and generate new text, images, or video that reflect the style of the training data.
| Processing method | What it does with data | Choose when |
|---|---|---|
| Traditional AI | Learns from historical tabular data and images to classify categories or predict numbers | The task is spam filtering, defect detection, or forecasting from past records |
| Generative AI | Learns patterns from training data to generate new content and explain unstructured data | The task is summarization, chat, captioning, document understanding, or code creation |
Exam tip: Traditional AI classifies or predicts from existing historical data, while generative AI creates new text, images, video, or code with similar characteristics to its training data.
Human-led insights are conclusions that a person draws after exploring samples, measurements, visualizations, dashboards, and reports and then makes the decision. Autonomous learning is improvement that happens when a machine learning model learns relationships directly from data exposure and improves its classifications, predictions, or generated content over time while people govern, monitor, and explain it.
Use retrospective analysis when the question is about the past, such as why sales dropped last quarter, because it aggregates stored records in BigQuery into trends, comparisons, and explanations. Use predictive modeling when you need an automated forecast of what will happen next, because BigQuery ML trains a model on that historical data and then runs inference on new serving data without requiring data movement.
Traditional artificial intelligence (AI) learns from existing historical data to classify information or predict future outcomes, such as classifying email as spam or not spam or predicting house prices. Generative AI learns from training data and then generates new data with similar characteristics, such as text, images, audio, video, or code, and is especially strong at working with unstructured content.
A retail organization relies on business intelligence (BI) dashboards to monitor structured sales tables and daily inventory counts. Now, the company wants to automatically analyze new customer feedback that includes:
The analytics team initially tried building manual, rule-based filters, but the rules failed to categorize most of the submissions.
Why is an artificial intelligence and machine learning (AI/ML) approach better suited for this requirement than traditional rule-based data analysis?
Last updated