In classic software development, a human defines how a program responds to which input. Machine learning (ML) is different: an algorithm receives many examples and derives a model from them on its own, which can also be applied to new, unknown cases. This is particularly useful where rules can hardly be written down completely, for example when deciding whether an email is spam.
Machine learning is the foundation of most of today's applications of artificial intelligence, from forecasts to language models.
How does machine learning work?
An ML model goes through two phases. In training, an algorithm adjusts the model's parameters so that it represents the sample data as well as possible. In use, also called inference, the finished model processes new data and delivers a prediction or classification.
Three forms of learning are distinguished:
- Supervised learning: The training data contains the correct answer, for example "invoice" or "payment reminder" as the category of a document.
- Unsupervised learning: The method looks for structures without predefined answers, for example for groups of similar orders.
- Reinforcement learning: A system learns through feedback in the form of rewards or penalties which actions lead to the goal.
Deep learning is a subtype of machine learning that works with multilayer neural networks.
Typical examples
- Spam and phishing detection in email traffic
- Sales and demand forecasts for purchasing and warehousing
- Detection of unusual sign-ins or access in IT security
- Automatic assignment of incoming requests to departments
- Predictive maintenance based on machine data
What companies should keep in mind
- Data quality: A model is only as good as its data. Gaps, errors, and outdated records lead to wrong results.
- Biases: If the training data reflects one-sided decisions from the past, the model adopts them. This problem is referred to as bias.
- Ongoing monitoring: If conditions change, a model's accuracy declines. Regular review and, if necessary, retraining are required.
- Traceability: Especially for decisions that affect people, it should be possible to explain how a result comes about.
- Assess effort realistically: Ready-made models exist for many standard tasks. Your own training is only worthwhile if sufficient suitable data is available.
How TOMORIS relates
An example of automatic classification is MailAutomate, the structured mail automation from TOMORIS with AI classification.
View AI consulting