Canadian scientists have modeled a neural network modeled on the human brain. She handled cognitive tasks more flexibly and efficiently than traditional systems.
The authors of the project are scientists from the Montreal Neurological Institute-Hospital and the Quebec Institute for Artificial Intelligence. They conducted a detailed study of magnetic resonance imaging data in a large repository of Open Science, reconstructed patterns of brain connections and designed a neural network based on them. The result is a system of I / O modules that truly resembles a biological brain. To test the health of the system, a cognitive task related to memory was set. We are not talking about bit-by-bit data recording, but about interpreting the information received and assimilating it as knowledge – following the pattern of the human brain. Next, the researchers monitored how the system performed it.
The authors of the project are sure that it is unique for two reasons. Previous work on brain connections (connectomics) involved describing the organization of the structure of the brain without regard to how the process of computation is carried out and basic functions are realized. In addition, traditional modern neural networks have an arbitrary structure, not always tied to the models of the organization of brain connections. Therefore, by integrating brain connectomics into the architecture of a neural network, the authors of the study hoped to obtain additional information about the physiological implementation of cognitive processes in the brain in order to derive new principles for designing neural networks based on these data.
As a result, it was found that a neural network built on the model of human brain connections – a neuromorphic neural network – performs cognitive tasks related to memory more flexibly and efficiently than analogs based on traditional arbitrary architecture.
“The project combines two powerful scientific disciplines. Neuroscience and artificial intelligence have common roots, but recently they have diverged. Using artificial neural networks will help us understand how the structure of the brain ensures the implementation of its functions. In turn, using empirical data to create neural networks will open up design principles for creating better artificial intelligence. Thus, these areas will help each other and enrich our understanding of the brain, ”commented one of its authors, Bratislav Misic.