ROBIOTIC

Lexikon

Transfer Learning

At a glance

Transfer Learning is one of several learning methods used for machine learning. The method is mainly used when an artificial intelligence is to be trained for a special purpose, e.g. for specific tasks such as face recognition or automated text analysis.

Learning from older models

What is special about the method is that it uses other AI models that have already been trained with suitable data in advance. In transfer learning, the artificial intelligence tries to learn from the findings of these models, so to speak. New systems can thus also acquire the data that complex deep learning systems have accumulated over a long period of time. This is comparable to the transfer of knowledge in humans: instead of laboriously and lengthily working out all the important information on a topic ourselves, we can attend a course, consult an expert or read a book. The experts have already done the work and can now provide us with the most important findings in a short time.

Advantages of Transfer Learning

Learning by knowledge transfer offers several advantages that make it attractive for IoT applications, among others. Above all, the method can be used to improve the efficiency and precision of machine learning.

Faster results
Because an AI can use the accumulated knowledge of previous AI systems, it usually achieves a usable data model much faster with transfer learning.

Training of new AI's
Transfer learning is not only efficient, but also fundamental: AI's need a training data set so that they can learn which problem to solve or which goal to pursue. AI systems for face recognition, for example, can be filled with the results of a deep learning model to get first clues that show them which distinctive features distinguish a face from other patterns. The new AI can thus begin to focus directly on the differentiation of individual faces, for example, and does not have to learn for itself which patterns it can classify as faces in general.

Continuous improvement
Transfer learning can also be used to allow the individual learning processes to build on each other again and again. This way, the system can be further optimized with each iteration and go into more and more detail. In connection with IoT, this could mean, for example, that the system continuously develops optimized solutions by learning from the previous interaction of the networked devices.

Kontakt