Thin Artficial Neurons Made of 2D Materials Could Advance AI Computing

A groundbreaking achievement has been announced by a team of researchers from the University of Oxford, IBM Research Europe, and the University of Texas. They have successfully developed atomically thin artificial neurons by stacking 2D materials.

The researchers enhanced the functionality of electronic memristors to respond to both optical and electrical signals.

This innovation enabled the creation of separate feedforward and feedback paths within the network, resulting in the development of winner-take-all neural networks.

Winner-take-alll Neural Networks

Winner-take-all neural networks are a type of computational learning program that have the potential to solve complex problems in machine learning.

They work by selecting the largest output signal from a group of neurons and suppressing the output of all other neurons in the group, resulting in only one neuron being activated at a time.

This enables the network to make a single, decisive response to a given input, and is often used for tasks such as classification and clustering.

The neural networks that the researchers developed could tackle intricate problems in machine learning, including unsupervised learning in clustering and combinatorial optimization.

2D materials are composed of a few layers of atoms, providing them with unique properties that can be fine-tuned based on the material's layering. To create the device, the researchers used a stack of three 2D materials: graphene, molybdenum disulfide, and tungsten disulfide.

This particular device has the ability to modify its conductance in response to the intensity and duration of light or electrical signals it receives.

In contrast to digital storage devices, these devices are analog in nature and operate similarly to neurons and synapses in the human brain.

Analog devices can produce gradual changes in stored electronic charge in response to a sequence of electrical or optical signals delivered to the device.

This feature allows for the creation of threshold modes for neuronal computations, similar to how the human brain processes a combination of excitatory and inhibitory signals.

Highly Exciting Development

"This is a highly exciting development. Our study has introduced a novel concept that surpasses the fixed feedforward operation typically utilized in current artificial neural networks," lead author Dr. Ghazi Sarwat Syed, a Research Staff Member at IBM Research Europe Switzerland, said in a statement.

The team's new technique explores the power of 2D materials for novel computing paradigms, according to co-lead author Professor Harish Bhaskaran.

The use of 2D structures in computing has been discussed for years, but this new development enables the start of new information processing approaches using industrially scalable fabrication methods.

This development could be crucial as the computational power required for AI applications has outpaced the development of new hardware based on traditional processors, according to the team.

The study's findings were published in Nature Nanotechnology.