Scientists from the Institute of Mathematical Problems of Biology of the Russian Academy of Sciences have created a neural network that controls its "gaze" and searches for objects in a perceived picture in much the same way as the organs of vision and the human brain do, according to an article published in the journal Neural Networks.
“The developed model offers a simple and unexpected explanation for a very complex cognitive process of searching for and recognizing objects in a picture perceived by our eyes,” says Yakov Kazanovich from the Institute of Mathematical Problems of Biology of the Russian Academy of Sciences in Pushchino. According to him, the neural network created by his team should help neurophysiologists understand how real human vision works.
Over the past ten years, hundreds of programmers and dozens of large IT companies have created countless machine vision systems capable of recognizing various objects in a perceived picture and classifying them. Modern robots, search engines and drones can use this data for a variety of purposes - for example, to bypass obstacles or search for a client when delivering a parcel.
Despite tremendous advances in this area, scientists still know virtually nothing about how human and animal vision works and how we manage to automatically classify and recognize even objects that we have never seen before.
Therefore, as Casanovic says, many features of human consciousness, perception of reality and vision still remain a mystery to neurophysiologists and psychologists. For example, scientists have argued for a long time about why a person can very easily find “contrasting” objects in a huge variety of other structures that are unlike him, but at the same time have difficulty finding several figures hidden in a small number of similar objects.
Kazanovich and his colleague Roman Borisyuk took a big step towards solving this problem by creating an artificial intelligence system, which, when solving these problems, behaves in exactly the same way as a person.
Its main feature, as scientists say, is that it consists of a multitude of relatively independent structures, the so-called "ensembles", in which neurons produce special vibrations. One of these structures becomes a kind of "conductor" who controls the work of the other "ensembles" and gives them tasks, while the other ensembles are essentially objects that the neural network "sees" in the picture.
"Ensembles" constantly compete with each other for influence on the "conductor" and on the operation of the entire neural network as a whole. The way this competition proceeds, as shown by the experiments and calculations of Casanovich, almost perfectly reflects the principle of human vision and is similar to the "sliding" of our gaze over the picture when looking for objects of different degrees of "contrast".
Promotional video:
This model, scientists hope, will help neurophysiologists not only find similar structures in the brains of humans and apes, but also understand how they work, which will bring us closer to creating "natural" machine vision systems.
