Artificial neurons are faster than human brain

Michael Schneider, a physicist at the National Institute of Standards and Technology (NIST), and his colleagues successfully built superconducting computer chips that simulate neuronal cells that can process information faster and more effective than the human brain.

This is an important milestone in the development of modern computer equipment capable of mimicking biological systems, and opening the door to machine-learning software. more natural. The results are published in Science Advances magazine on January 26.

Artificial intelligence software is increasingly imitating the brain more. Algorithms such as Google's automatic image classification and language learning program use artificial neuron networks to perform complex tasks. However, because computer hardware is not normally designed to run a simulation algorithm for brain operation, machine-learning tasks require greater computing power than the human brain. .

Neurons have the function of storing and transmitting information in the brain.(Photo: CNRI / SPL).

NIST is one of the few research groups trying to develop hardware that mimics the human brain, or neuromorphic chip , in hopes that it will run brain simulation software more effectively. In conventional electronic systems, transistors process information at regular intervals with precise values ​​(bits 1 or 0), but neuromorphic devices can store a small amount of information from multiple sources, The sea changes it to create another type of signal and emits only electrical impulses when needed, like biological neurons. As a result, neuromorphic devices require less energy to operate.

However, these devices are still ineffective, especially when they transmit information through space, or synapses (synapses), between transistors. Therefore, Schneider's team created artificial synapses including electrodes made of superconducting niobium (Nb) - conductive without resistance. They fill the gap between electrodes with thousands of magnetic nanoparticles (nanocluster) of manganese (Mn).

By changing the magnetic field in the synapse, Mn nanoparticles can be oriented in different directions. This allows the system to encode information at both the electrical and magnetic levels, creating a much greater computing power than other neuromorphic systems that do not occupy additional physical space.

They can transmit electrical signals faster than human neurons, but use only 1 / 10,000 energy levels compared to a biological synapse.

The problem is that synapses can only operate at temperatures close to absolute zero and need to be cooled with liquid helium. Steven Furber, a computer engineer at Manchester University, said this is not feasible if chips are used in small devices. However, Schneider said that cooling these devices requires much less energy than operating a conventional electronic system with comparable computing power.

"It takes a long time before the chips can be applied to real computers. In addition, they also face stiff competition from many other brain simulation devices that are being developed. " Carver Mead, electrical engineer at the California Institute of Technology (USA), said.