What is machine learning?

Machine learning: definition

With machine learning, nobody writes every rule by hand anymore. Instead, a learning method receives many example records and works out for itself which features lead to which outcome. The result is a model. It can be applied to new data, for instance to assign an inquiry to a category or to predict a machine failure.

Machine learning is a subfield of artificial intelligence and the technical basis of most AI applications today. Deep learning, learning with deep neural networks, is in turn a subfield of machine learning and powers language models and image recognition. In daily life you encounter it in spam filters, recommendations in online stores or automatic text recognition.

How machine learning works

There are three types of learning. In supervised learning, the training data includes the right answer, such as “invoice” or “complaint.” In unsupervised learning, the method looks for groups or outliers on its own. In reinforcement learning, a system tries out actions and is rewarded for good results.

A project usually runs in these steps: collect and clean data, train a model, test it on unseen data, integrate it into the application and monitor it continuously. The effort often lies less in the algorithm than in data preparation and stable operation.

Real-world example: predictive maintenance at a supplier

A metalworking company runs several CNC machines. Sensors deliver data on vibration, temperature and power consumption. A model learns from past failures which patterns precede a bearing failure and reports anomalies in good time. Maintenance is then scheduled for a quiet period instead of the machine stopping in the middle of an order.

At first a single machine was enough as a pilot. Only when the alerts were reliable and the maintenance team trusted them was the method rolled out to other equipment. The example shows that the value only arises in combination with a clear process after the warning.

Advantages, limits and typical mistakes

Machine learning detects patterns that are hard to describe by hand and improves with better data. At the same time it is only as good as its training data: biased, too small or outdated datasets lead to wrong results. Whether machine learning is needed at all is something Nano Concepts clarifies with you—often a simple rule is enough.

Typical pitfalls are unclear target metrics, no domain experts on the project and a model that is no longer checked after launch even though reality changes. When models prepare decisions about people, for example in job applications or credit, data protection law and the EU AI Act also apply.

• Define a specific question and a measurable success criterion
• Check the data situation before building a model
• Start with a pilot and compare results against a simple rule
• Monitor the model after launch and retrain it when needed
• Involve the business department early