As a rule, for training neural networks, a set of marked-up photographs is used – the necessary objects on it are enclosed in rectangles. Facebook specialists were able to make sure that the set of photos did not have to be marked up in advance – the computer itself “sees” what is needed on it.
We are used to the fact that in order to train neural networks, we need to collect a set of photographs with the necessary objects, for example, with cats, which must be marked in special rectangles. The model will then be able to define objects (cats) on its own using new input data, that is, new images.
Facebook specialists used the method of the so-called “semi-supervised” learning. It allows you to get rid of the stage of marking objects. The DINO system itself is able to find objects of interest in videos with people and animals of any markings.
It is argued that such a system even performs better than traditional teaching. DINO can “understand” that dogs and cats, for example, are visually similar to each other. The system has context and metadata in memory.
Read also:
Named habits that help you lose weight without dieting and sports
Differences found in the brains of coffee lovers
Mint tea turns out to be unhealthy
Coronavirus turned out to be deadly for recovered
The safest water for health has been named
