---
title: "Machine learning: definition and example – Carl-Uwe Würfel"
description: "Machine learning explained: how software learns from data, the four types of learning, why training data matters, and how it differs from fixed rules and RPA."
url: https://carl-wuerfel.de/en/glossary/machine-learning
author: Carl-Uwe Würfel
---

1. [Home](https://carl-wuerfel.de/en)
  2. [Knowledge](https://carl-wuerfel.de/en/knowledge)
  3. [Glossary](https://carl-wuerfel.de/en/glossary)
  4. Machine learning

# Machine learning

Machine learning is a branch of artificial intelligence in which software learns its rules from example data instead of having them programmed step by step.

## On this page

  1. How a model learns from data
  2. The four types of learning
  3. Training data determines what a model can do
  4. Machine learning, fixed rules and RPA
  5. Models age in production
  6. What good implementation looks like
  7. Where the term stops working
  8. Frequently asked questions
  9. Sources

A model can only be as good as the data it learns from, and it ages when the real world changes.

_Figure: A cycle using the example of a retailer with a few thousand products that has so far planned orders based on experience and spreadsheets. A model learns from several years of sales history, seasonality and promotions. Before launch, it is tested against the previous year 's manual plan. It then suggests weekly order quantities, and the planners still make the call. It is compared every week: if accuracy holds, it moves on to the next week. If accuracy drops, for example because a new online channel changes buying behavior, the model is retrained on the new data._

**Example**

A retailer with a few thousand products has so far planned orders based on experience and spreadsheets. A model learns from several years of sales history, seasonality and promotions and suggests weekly order quantities; the planners still make the call. Before launch, the model is tested against the previous year's manual plan, and in production it is compared every week. When a new online channel changes buying behavior, accuracy drops, and the model is retrained on the new data.

Conventional software follows rules a person wrote down: if the subject line contains "prize," move it to spam. In machine learning, the software instead gets many examples, say thousands of emails that people have marked as spam or not spam, and works out for itself what spam looks like. The result is called a model. It applies what it learned to new cases it has never seen. In his 1997 textbook, Tom Mitchell describes the field as the study of computer algorithms that improve automatically through experience.

Machine learning is a branch of [artificial intelligence](https://carl-wuerfel.de/en/glossary/kuenstliche-intelligenz). Inside it sits [deep learning](https://carl-wuerfel.de/en/glossary/deep-learning), which uses neural networks with many layers, and inside that sits today's generative AI with its [large language models](https://carl-wuerfel.de/en/glossary/llm). Arthur Samuel of IBM used the term in 1959 for a checkers program that got better by playing. Today machine learning runs spam filters, product recommendations, fraud detection, sales forecasts and visual inspection in quality control.

For executives, one consequence matters most: a model can only be as good as the data it learns from, and it ages when the real world changes. Most of the work therefore lies not in the algorithm but in getting the data, maintaining it, testing and ongoing monitoring.

## How a model learns from data.

It starts with a question that data can answer, for example: how many units of a product will sell next week? That requires historical examples with features (season, price, promotions, weather) and, where available, the actual outcome. The data is cleaned and split. The larger part is used for training; the smaller part is held back.

During training, an algorithm adjusts the model's internal parameters until its predictions fit the examples as closely as possible. Then comes the most important step: testing on the held-back data the model has never seen. If it does much worse there than in training, it has memorized the examples instead of finding general patterns. This is called overfitting.

Only when the model holds up on unseen data does it go into production, where it makes predictions for new cases; this step is called inference. A model usually delivers a probability or an estimate, not certainty. How that uncertainty is handled, whether a person decides above a certain threshold or the system acts automatically, is a decision for the company, not the model.

## The four types of learning.

In supervised learning, the model learns from examples with a known correct answer, called labels. A spam filter trained on emails that people have marked is the standard example; the European Commission uses it in its [AI Act](https://carl-wuerfel.de/en/ai-act) guidelines too. Most forecasting and classification tasks in business fall into this category.

In unsupervised learning there are no labels. The model looks for structure on its own, such as groups of similar customers or transactions that deviate from the usual pattern. This suits segmentation and anomaly detection but delivers no ready-made interpretation; people have to work out what a group it finds actually means.

Self-supervised learning creates labels from the data itself by hiding part of it and having the model predict it. This is how large language models are pretrained: they learn to predict the next piece of text. Finally, in reinforcement learning, a system tries out actions and learns from rewards and penalties which strategy reaches the goal. Richard Sutton and Andrew Barto describe the field in their standard textbook. It is used in game-playing programs, robotics and the [fine-tuning](https://carl-wuerfel.de/en/glossary/fine-tuning) of language models with human feedback.

## Training data determines what a model can do.

A model learns exactly what is in its training data and nothing else. If certain cases are missing, such as a new sales channel or a rare type of defect, it cannot recognize them reliably. If the labels are inconsistent because different people marked the same cases differently, the model learns that inconsistency too.

Historical data also carries historical decisions. A model that learns from past hiring or lending decisions picks up their patterns, including disadvantages for particular groups. The model is not neutral; it simply applies existing biases faster and more uniformly. One way to check is to evaluate the error rate separately for each group rather than only on average.

For high-risk systems, Article 10 of the EU AI Act requires training, validation and test data to be relevant, sufficiently representative and, to the best extent possible, free of errors and complete in view of the intended purpose. After the July 2026 amendment, these obligations apply from December 2, 2027, and from August 2, 2028 for AI in regulated products (as of November 2026). They are a useful yardstick for any project. If the data includes personal information, the GDPR also applies, including to training.

## Machine learning, fixed rules and RPA.

Not every automation needs machine learning. If the rules are clear, stable and can be written down, such as "invoices over 10,000 euros go to management," a fixed rule is simpler, cheaper and easier to trace. Machine learning pays off where there are too many factors to capture as rules but enough examples exist: fraud patterns, swings in demand, visual defects on a production line.

Robotic process automation ([RPA](https://carl-wuerfel.de/en/glossary/rpa)) is something else. A software robot clicks through screens in a fixed sequence and moves data between systems. It learns nothing and recognizes no patterns. In practice the two are often combined: a model reads and classifies an incoming document, and an RPA workflow enters the result into a legacy system.

The law draws the line too. Under the European Commission's guidelines on the AI system definition, systems that only execute rules defined by people fall outside the AI Act. The same goes for simple predictions that just project an average, even if they could technically count as machine learning. For planning, this means checking first whether a rule will do, and using machine learning only where rules stop working.

## Models age in production.

A trained model is a snapshot of the world its data came from. When that world changes, through new products, different buying behavior, new fraud schemes or changed processes, its predictions gradually get worse. Practitioners call it data drift when the input data shifts and concept drift when the relationship between input and outcome changes. The model does not report either on its own.

Running a model therefore includes ongoing monitoring. Predictions are compared with the actual outcome once it is known, and if quality drops below an agreed threshold, the model is retrained or someone steps in. This discipline is called MLOps, modeled on DevOps in software development.

In 2015, Google researchers described how the actual learning code makes up only a small part of a production machine learning system. The rest is data collection, data validation, configuration, monitoring and interfaces, and that is where most maintenance costs arise. Anyone planning a project should budget for years of operation, not just for building the first model.

## What good implementation looks like.

Good projects start with a baseline. Before anyone builds an elaborate model, it is clear how well the current process performs and how well a simple rule or an average would do. Only a model that clearly beats that baseline justifies the effort.

The second sign is a metric that fits the business goal. A high overall hit rate is of little use if the rare but expensive cases are missed. In fraud detection, what counts is how many real fraud cases are caught and how many false alarms the team has to work through. The test data should come from a later period than the training data, just as it will in production.

Then come traceability and ownership. Data sources are documented, model versions are stored, and there is a fallback if the model fails or behaves oddly. For every automated decision, it is clear when a person reviews it. When hiring a vendor, ask about the baseline, the metric, the test data and monitoring in production. A proposal that only describes training covers the smaller part of the work.

## Where the term stops working

Machine learning finds correlations in data, not causes, and it fails in situations that do not appear in its training data. It should be distinguished from artificial intelligence as the umbrella term, from deep learning and generative AI as subfields, and from rule-based automation such as RPA, which learns nothing. Descriptive statistics such as averages and dashboards are not machine learning, even when they use the same data.

## Frequently asked questions

What is the difference between machine learning and artificial intelligence?

Artificial intelligence is the umbrella term, and machine learning is one branch of it. In machine learning, software learns its rules from data. AI also includes knowledge-based systems whose rules and facts were encoded by experts.

What is the difference between machine learning and deep learning?

Deep learning is a branch of machine learning that uses neural networks with many layers. It is strong with images, speech and text but needs a lot of data and computing power. For tabular data such as sales figures, simpler methods like decision trees are often enough.

Is ChatGPT machine learning?

Yes. The language models behind ChatGPT are deep learning models and therefore part of machine learning. They were pretrained with self-supervised learning on large amounts of text and then fine-tuned with human feedback.

What are supervised and unsupervised learning?

In supervised learning, a model learns from examples with a known correct answer, such as emails marked as spam. In unsupervised learning there are no such answers; the model finds groups and anomalies in the data on its own.

How much data does machine learning need?

There is no fixed minimum. It depends on how many factors are involved and how often the important cases appear in the data. A useful first test is whether a simple model beats the current process on held-back data.

What is overfitting?

Overfitting means a model has memorized its training data instead of finding general patterns. It scores very well in training but much worse on new data. The only way to detect it is with test data the model did not see during training.

When is machine learning the wrong choice?

When the rules are clear and stable, when there is little example data, or when every decision has to be fully explainable. In those cases a fixed rule is simpler, cheaper and easier to check. Machine learning pays off where rules break down under the variety of cases.

## Further reading

  * [**AI glossary**](https://carl-wuerfel.de/en/glossary)
  * [**AI automation**](https://carl-wuerfel.de/en/ai-automation)
  * [**Data quality**](https://carl-wuerfel.de/en/glossary/datenqualitaet)
  * [**Fine-Tuning**](https://carl-wuerfel.de/en/glossary/fine-tuning)
  * [**RPA: Robotic Process Automation**](https://carl-wuerfel.de/en/glossary/rpa)
  * [**LLM: Large Language Model** Term](https://carl-wuerfel.de/en/glossary/llm)
  * [**AI bias** Term](https://carl-wuerfel.de/en/glossary/ki-bias)
  * [**Artificial intelligence (AI)** Term](https://carl-wuerfel.de/en/glossary/kuenstliche-intelligenz)
  * [**Generative AI** Term](https://carl-wuerfel.de/en/glossary/generative-ki)

## Sources

  1. [Tom M. Mitchell: Machine Learning, McGraw Hill, 1997](https://www.cs.cmu.edu/~tom/mlbook.html)
  2. [Arthur L. Samuel: Some Studies in Machine Learning Using the Game of Checkers, IBM Journal of Research and Development, 1959](https://doi.org/10.1147/rd.33.0210)
  3. [Richard S. Sutton, Andrew G. Barto: Reinforcement Learning: An Introduction, 2nd edition, 2018](https://incompleteideas.net/book/the-book-2nd.html)
  4. [D. Sculley et al.: Hidden Technical Debt in Machine Learning Systems, 2015](https://papers.nips.cc/paper/5656-hidden-technical-debt-in-machine-learning-systems)
  5. [European Commission: Guidelines on the definition of an AI system, 2025](https://digital-strategy.ec.europa.eu/en/library/commission-publishes-guidelines-ai-system-definition-facilitate-first-ai-acts-rules-application)
  6. [European Parliament and Council: Regulation (EU) 2024/1689 (AI Act), Art. 10, 2024](https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32024R1689)
  7. [European Parliament and Council: Regulation (EU) 2026/1744 (Digital Omnibus on AI), 2026](https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32026R1744)