How images learn to think
Self-driving cars, automatically reordering refrigerators, service robots that bring drinks to the table, doors that open with a fingerprint, self-sorting digital image archives – these future scenarios already exist or are very close to becoming reality. Often it is imaging technologies or image evaluation tools based on artificial intelligence that even make all of that possible.
Trial and error
One primarily understands artificial intelligence (AI) as self-learning systems. While one used to have to permanently program something in, like a program that, for example, recognises faces, software is so flexible today that it formulates the program itself. So-called "neuronal networks" are used as basic technology to this purpose, which, similar to human beings, learn on the basis of trial and error which operation they need to execute to come to the desired result. When several layers of neuronal networks are used in combination, the results improve considerably. One then refers to "deep learning".
hotoDirector © CyberLink
In order that neuronal networks achieve the desired result, feedback concerning the accuracy of their interim results and a large quantity of data are helpful. Ideally, neuronal networks are therefore already supplied with data already evaluated by humans from the start, e.g. manually tagged images. The more of this they have, the more easily and better they can recognise patterns and the more reliable the end results.
The German stock photography provider EyeEm has, for example, made use of its large inventory of data and developed an application on the basis of artificial intelligence, which automatically suggests keywords for images. Google, Apple and Microsoft already use this automatic recognition of scenes as a standard in their search engines. An important segment of AI technologies is that of facial recognition. Image administration programs can in this way provide automatically compiled collections with images of certain persons on the basis of already read-in photos. Microsoft even goes so far as to attempt to automatically recognise emotions. In this way it could become possible, for example, to easily find a baby photo with a defiant facial expression for a souvenir photo album.