Israeli scientists have created a new neural network architecture for recognizing objects and images

Researchers from Israel have developed a new neural network architecture that surpasses classical convolutional networks in solving problems related to image recognition and computer vision. The work was published in the journal Scientific Reports. The results were announced on Monday by the press service of Bar-Ilan University.

"We managed to show that a simple machine learning system architecture based on dendritic trees is able to classify objects and images better than multilayer deep convolutional networks can do. This opens the way for the creation of more efficient algorithms and nature-like artificial intelligence systems," said a professor at Bar-Ilan University in Ramat Gan (Israel) Ido Kanter, whose words are quoted by the press service of the university.

Professor Kanter and his colleagues have developed a new neural network architecture that will significantly simplify and speed up their work. According to scientists, according to its principles of operation, it is more similar to real human brain tissues than to the currently existing multi-layered convolutional neural networks in which information is sequentially processed by dozens or even hundreds of separate sets of neurons.

A new form of neural networks Israeli mathematicians have found out that the device of machine learning systems can be radically simplified if, when developing neural networks, they use peculiar trees from analogs of nerve cells similar in structure to the dendrites of natural neurons.

In this case, the machine learning system is organized in the form of several "trees" consisting of branching sets of analogs of neurons connected to several subsequent sets of cells at once.

The key feature of these "trees" is that each of their branches is connected to only one exit (one of the variants of the final answer). According to the researchers, this greatly simplifies and speeds up calculations compared to classical convolutional neural networks, which, when searching for an answer, involve a huge number of potentially related and unrelated neurons.

The scientists tested the work of this approach on a standardized set of images from the MNIST database, which is traditionally used to evaluate the effectiveness of machine learning and computer vision systems. The tests showed that the created approach correctly classified 99.07% of the images, which is comparable to the quality of the classic LeNet-5 neural network (99.05%).

The creation of this neural network architecture, scientists hope, will lead to the development of specialized devices that will calculate the work of such "dendritic" machine learning systems as quickly as possible. This will reduce the energy costs of artificial intelligence, the researchers concluded.