Skip to content

New artificial neuron device that reduces energy requirements for neuronal calculations

29/12/2023

Scientists from the prestigious University of San Diego have conceived a revolutionary device for artificial neuron innovative capable of performing neural calculations with energy and spatial efficiency between 100 and 1.000 times greater than current devices based on CMOS technology.

The recent discovery was disclosed on March 18 in a study published in the prestigious magazine Nanotechnology from Nature.

Understanding and HEALING your traumas / Complete Audiobook

To feed one of the layers of linked artificial neurons In the neural network system, it is imperative to apply a mathematical function called nonlinear activation function. However, this demanding application consumes significant computing power and requires numerous circuits due to the need to transfer data between memory and the external processor.

Thus, the brilliant researchers at UC San Diego designed a sophisticated nanometer-sized device capable of executing this activation function with astonishing efficiency.

Duygu Kuzum is a professor of electrical and computer engineering at UC San Diego's distinguished Jacobs School of Engineering.

"Neural network computing in hardware is becoming increasingly inefficient as neural network models grow in complexity and dimension," Kuzum said. "Therefore, we have developed a unique nanoscale artificial neuron device that executes these calculations in hardware in a tremendously efficient manner, in terms of space and energy consumption."

The research was led by Kuzum and doctoral student Shangheon Oh, in collaboration with UC San Diego physics professor Ivan Schuller, who directs a DOE Energy Frontier Research Center. This center actively participates in the development of hardware implementations for artificial neural networks with a focus on energy efficiency.

Description of the innovative device

This technological innovation is based on the rectified linear unit or ReLU, which is one of the most recurrent activation functions in the training of neural networks. This process requires hardware that can undergo progressive changes in resistance, which was a key goal for the engineers. The device can gradually transition from an insulating state to a conductive state with a minimal amount of heat.

This transition, known as the Mott transition, takes place in an incredibly thin layer of vanadium dioxide, located above a nanowire heater made of titanium and gold. The aforementioned layer of vanadium dioxide gradually heats up when current flows through the nanowires, causing a slow and controlled change from an insulating to a conductive state.

Oh had the distinction of being the lead author of the study.

“The architecture of this ingenious device is amazingly interesting and innovative,” commented Oh. "In this context, we make current flow through a nanowire located above the material with the aim of heating it and inducing a tremendously gradual change in resistance."

Implementation of the revolutionary device

To carry out the implementation, the team created a series of activation artifacts and a series of synaptic devices, and then integrated both on a printed circuit board designed specifically for this purpose. They were later interconnected, resulting in a hardware version of a neural network.

This neural network was used to process an image through edge detection, thus identifying the contours and edges of the objects present in the image. The integrated hardware system proved capable of performing operaconvolution tions, vital for various types of deep neural networks.

"At this point, this is simply a demonstration of concept," Kuzum explained. "It's a tiny system in which we've just stacked one layer of synapses with one layer of activation. “By stacking more of these, we could create a more complex system suitable for various applications.”

READ MORE ARTICLES ABOUT: Data Science with AI.

READ THE PREVIOUS POST: Scientists seek to perfect AI systems with new models of "neurons".