Overfitting
Excessive adaptation to details of training data that reduces a model’s ability to generalise to new examples.
What happens in overfitting?
Overfitting means that a model adopts details of its training data so closely that it does not generalise adequately to new relevant examples. These details may include chance patterns or features with little relevance to the real task. Good results on training examples alone are therefore insufficient for assessment. Investigation considers results on separate data, the data split and the course of learning, among other factors.
In a fictional spare-part project, certain item classes often appear against the same background colour in training photos. The model could use that background as a shortcut instead of learning the important characteristics of the parts. The team therefore plans checks with previously unused images and changed backgrounds. It also examines whether the data covers the intended use and how different model variants generalise. A single incorrect assignment under different lighting does not establish overfitting; other causes need investigation too.
Sources
Related terms
- Machine learningModels learn patterns from data and apply them to new inputs.
- LLM evaluationLLM evaluation tests model outputs against defined tasks and criteria. One good example does not establish overall quality.
- Training dataThe examples from which a model learns relationships or parameters during training.
- Validation dataData used to assess and select model variants during development.
- Test data in machine learningData reserved for final assessment whose results were not used to select the model variant being assessed.
- UnderfittingA model fails to capture the relationships needed for its task adequately and therefore produces weak predictions.
Your project with ASES
Describe your project to us. We will get back to you with an initial assessment.