Scientists of Tyumen state University presented to the world scientific community the development has no analogues biomorphic neuroprocessor based on a new component of nanoelectronics — combined memristor diode crossbar, the press service of the TSU. The study is published in Microelectronic journal.
Existing neural processors, according to scientists of Tyumen state University, is designed for hardware acceleration of calculations in artificial neural networks with simple neurons and maintain computer vision, machine learning and other systems with a weak artificial intelligence (AI).
Processing information and making decisions in such processors is done by selecting the most plausible solution based on pre-determined associations.
Neuroprocessor, developed at TSU, according to its creators, is able to generate new associations (new knowledge) in biologically similar mechanism that allows you to talk about a possible transition from weak to strong artificial intelligence. Under strong AI implies the ability to comprehend new knowledge.
According to TSU, biomorphic neuroprocessor implements a pulse neural network hardware based on advanced electric and original software (Neural Computing and Applications) biomorphic models of the neuron.
The authors claim that their CPU is capable of, apart from solving traditional tasks of information processing, to reproduce the work of a cortical column of the brain.
“In contrast to existing neural processors with simple neurons presented us biomorphic neuroprocessor gives you the opportunity to make decisions not only based on predetermined associations, but also based on new associations formed in the process of signal processing in a dynamically changing environment,” explained Professor of applied and technical physics, Tyumen state University Sergey Udovichenko.
According to the Professor, neuroprocessor, in essence- the unique design of the new generation, and is the carrier of artificial intelligence.
Developed specialized hardware can be solved technical problem — increasing the speed and efficiency of calculations in comparison with the current computing resources (personal computers, servers, and supercomputers).
The effect is achieved through the use of mixed analog-digital computing, including using a bipolar memristor integrated into a combined memristor diode crossbar.
Today, scientists of Tyumen state University have successfully demonstrated that information processing is made memristive diode crossbar — weighing, addition, and routing of pulses, as well as associative learning and generation of new associations. Still, associative learning was demonstrated only in hardware neural networks, are constructed on a discrete memristor.
Currently, the research team continues testing the new system. According to scientists, the launch of a unique neuroprocessor in small-scale production will be possible by 2025. Research supported by grants RFBR № 19-07-00272 and No. 19-37-90030.