Computer taught to perceive optical illusions, to understand their nature

SL, Понедельник 24 Сентябрь 2018 - 16:26:02

Scientists have created a neural network, subject to optical illusions of human nature. It will help not only to understand the nature of this curious phenomenon, but also to improve computer vision algorithms. The achievement is described in a scientific paper published in the journal Psychological Review group, headed by Thomas Serra (Thomas Serre) from brown University in the United States.

Psychologists know a huge number of visual illusions. Their extensive collection can be seen, for example, the site of the psychological faculty of Moscow state University named after M. V. Lomonosov. However, the neurophysiological basis of the illusion still largely remain a mystery.

This applies in particular to the class of illusions that could be called context-sensitive. In them the visual impression of the object is distorted due to its specific environment. For example, the blue circle in the center of the green looks turquoise and the gray circle in the middle of the pink – greenish. Interestingly, in the first case, the object becomes more similar to the context (the turquoise color is more green than blue), and the second – Vice versa (shades of green further away from pink than grey).

To understand the mechanism of this phenomenon, the authors have created a neural network. They mimicked the structure and physiology of the visual cortex of man. Researchers were interested in how neurons "talk" to each other and adjust the "conclusions" of each other in the perception of complex stimuli.

One of the innovations was the creation of a feedback between the artificial neurons. They've enhanced or suppressed response of neurons, "perceiving" the Central object, depending on the response of cells occupied by the context. The authors consider the likely existence of such connections in the visual parts of the cortex, but this hypothesis remains to be tested.

Such channels transmit messages was an innovation in artificial intelligence technology. Algorithms for deep learning provide a similar relationship between the different layers of the neural network, but not between neurons of one layer.


The researchers presented their work in various context-dependent illusion. Scientists regulate the force feedback, while the response of artificial neurons did not like that which demonstrate their natural analogues in primates.

After that, they tested how the neural network "sees" these illusions. It turned out that artificial intelligence confuses the color and shape exactly the same as people.

To check whether the constructed system is redundant, the authors selectively removed some of the feedbacks. However, the results of artificial intelligence has deteriorated.

"Our model is simple, which is necessary and sufficient to explain the behavior of the visual cortex in relation to contextual illusions," says Serre.

Now a word for the neurophysiologists, who will confirm or deny the existence of such backward linkages in the human brain.

Capacity development is not confined to the study of human vision. It can be useful in order to improve the view of the computer.

In fact, a context-dependent illusion, as is clear from the name, arise from the manners of our brain to perceive an object one way or the other depending on its environment. Illusion is an example of how this approach can be wrong. However, for such error you need a very specific combination of object and context. No wonder that optical illusions don't surround us at every step, and specially designed and carefully collected.

In most cases, the ability to consider the context very useful. It makes a significant contribution to the perfection of our visual perception, and in particular, not enough computers to get us to "catch up". But in conditions when artificial intelligence drives a car, the quality of vision depends on our security.

The authors hope that their work will help to teach computers in the analysis of the images to give greater attention to context. In this case, the modern world will be easier and safer.

We will remind that earlier "News.Science" (nauka.vesti.ru) wrote about artificial intelligence that mimic mental illness. We are not spared the attention and computer vision, describing the systems that produce faces in the crowd and differentiate men from women the smile.




эти данные от SlavaLand.Ru
( http://slavaland.ru/part/content/content.php?content.2160 )