Learning with noisy labels is part of Artificial Intelligence, Big Data and Smart Data, as well as Digital Transformation. It describes a situation where artificial intelligence (AI) is trained with data whose „labels“ or assignments are not always entirely correct.
In everyday life, this can happen, for example, in an online shop: an AI system is supposed to recognise whether photos show shoes or T-shirts. However, some images are accidentally mislabelled – for example, a T-shirt has been labelled as a shoe. These „noisy labels“ are like typos in the data that can confuse the AI.
Nevertheless, a good AI can learn to recognise the right patterns and ignore such errors. Especially with very large amounts of data, it is rarely possible to prevent a few labels being incorrect. With sophisticated methods, the AI develops a certain „fault tolerance“ and still delivers usable results in the end, for example in product recommendations or image recognition.
Learning with noisy labels is therefore particularly important because real-world data is rarely perfect. It ensures that AI systems remain robust and practical – even when not all data is entirely clean.
Deutsch 
English (UK)