Chickens in a hurry: drones will mimic the thinking of birds

During an experiment with ordinary chickens, Russian scientists for the first time described the mechanism of birds' thinking, which allows them to distinguish some objects from others. This natural mechanism works much more efficiently than existing methods used in modern electronic neural networks, which require large amounts of data and significant computing power to solve similar problems. Experts plan to use the open architecture to create self-learning robots and drones capable of performing complex tasks previously accessible only to humans. For more information, see the Izvestia article.

How birds think

Researchers from the Institute of Ai at MIPT and Moscow State University have described for the first time the mechanism of thinking of birds, which allows them to distinguish objects with different properties. This natural brain process is much more efficient than its electronic counterparts. Conventional neural networks require supercomputers and terabytes of data for this, while any chicken can quickly learn through trial and error to distinguish, for example, edible from inedible. Scientists will be able to use the architecture discovered by researchers to create robots or drones that will be able to self-learn in the process and perform tasks that were previously accessible only to humans.

A special experiment with ordinary chickens helped the developers understand how birds think. A plexiglass plate with beads of different colors was placed in the experimental chamber. Food was scattered between them, the number of granules of which approximately corresponded to the number of beads. The bird, pecking, had to learn to distinguish edible grains from inedible beads, forming the "inedible" category. With the help of cameras, experts monitored the behavior of birds and analyzed the logic of their decision-making.

— We have developed a new experiment design in which chickens were presented with three surfaces with beads in sequence. The goal was to determine the moment of category formation in the learning process. The results showed that when presented sequentially, a category is formed, however, the categorization algorithm depends on the order in which the colors are presented. This phenomenon can be related to both the innate functional systems of chickens and their individual experiences," said Ekaterina Diffine, a graduate student at the Lomonosov Moscow State University Faculty of Biology.

Flash reception: magnetic storm affected drones, orientation of birds and health of earthlings

Have the vagaries of space weather become record-breaking and what consequences can we expect from them?

The behavior of the birds turned out to be natural: on average, the frequency of errors — pecking on beads instead of food — decreased, and the birds gradually formed a new category of objects — "inedible". Scientists have described this pattern using a mathematical model that allows them to predict the sequence of actions of a bird.

— It turned out that the observed sequences of single behavioral acts can be caused by some hidden states, that is, they have an additional structure. After eliminating the pecking on the feed, the additional structure disappears. Thus, the previously discovered "structure" is explained, apparently, by both the inclusion of reinforced actions and the color of the beads," explained Evgeny Dzhivelikyan, a research engineer at the Center for Cognitive Modeling at the Institute of AI at MIPT.

Experts also studied the role of brain asymmetry by teaching birds with one eye closed. Due to the almost complete overlap of the optic nerves in birds, information from each eye is processed mainly by the opposite hemisphere. When the bird was tested with the other eye open three hours after training, the number of errors increased dramatically. This means that the memory of the category initially remains only in the hemisphere that received the visual experience.

Continuously learning AI

The research paves the way for the creation of a new generation of AI capable of learning, preserving and transferring knowledge in a similar way to biological systems. The transfer of information from the right hemisphere to the left indicates a possible evolutionarily formed mechanism for optimizing learning, said Yaroslav Seliverstov, a leading expert in the field of AI at University 2035.

— Transferring these principles to AI can lead to the creation of systems capable of true continuous learning. An architecture that mimics this asymmetric process will allow you to quickly adapt to new conditions without erasing previously acquired proven knowledge," he said.

The technology could find applications in AI to create robots or drones for search and rescue missions and referral systems. They will not only analyze historical user data, but also flexibly and safely explore new, potentially useful options for humans, while minimizing "erroneous actions," the expert emphasized. For example, an agricultural drone will be able to distinguish between weeds and cultivated plants. It is also the key to more efficient computer vision systems for medical diagnostics, where the model must confidently distinguish artifacts from pathologies, constantly learning from a new data stream.

Brain research does not always become the basis for creating new algorithms and models of AI. Despite certain similarities between text processing in large language models and brain processes, the learning process in these systems proceeds in different ways, explained Alexander Kugaevskikh, associate professor at the ITMO Faculty of Software Engineering and Computer Technology.

— This study can theoretically help in those campaigns when the AI does not conduct several iterations of training, but remembers an object and its signs at a glance. However, so far there have been practically no successful practices of such neurophysiological research, despite numerous attempts. But that doesn't mean it shouldn't be done," the scientist said.

The Bird of God Bites: The Emotional Ups and Downs of Birdwatcher

The famous book by the British ornithologist John Alec Baker has been published in Russian.

The ability of animals to distinguish categories of objects in the external environment is quite difficult to verify in neurobiological experiments. It would seem that a simple division into "edible" and "inedible" is an elementary operation, but scientists still have only a faint idea of how it is implemented on neural networks, said Igor Bondar, an expert at the Vladimir Zelman Center for Neurobiology and Neurorehabilitation at the Skoltech Center for Bio and Medical Technologies.

— We must pay tribute to the colleagues who conducted an accurate experiment demonstrating at the behavioral level such a brain function as categorization. Of course, the first results will require further confirmation, but such a confident start in a complex topic is impressive," he said.

The statements of the authors of the study about the possibility of creating a new AI architecture still look exaggerated, says Anton Averyanov, CEO of the ST IT group of companies, TechNet NTI market expert.