Machine learning & AI

For use in next-generation AI computing, artificial neurons simulate complicated brain functions.

For the purpose of computing, scientists have developed artificial neurons that are atomically thin and capable of processing both electric and light signals. The material improves a neural network’s capacity for complex problem solving by allowing separate feedforward and feedback paths to exist simultaneously.

In order to create machine learning systems that are quicker and use less energy, researchers have been looking into how to duplicate the adaptable computational capabilities of biological neurons for a number of decades. Utilizing memristors is one promising strategy: electronic components with the ability to modify their conductance to store a value and then use that value for in-memory processing.

However, the difficulty of integrating both feedforward and feedback neuronal signals has been a major obstacle to replicating the intricate processes of biological neurons and brains using memristors. These instruments support our mental capacity to learn complex errands, utilizing prizes and blunders.

A group of specialists at the College of Oxford, IBM Exploration Europe, and the College of Texas have reported a significant accomplishment: the advancement of molecularly flimsy counterfeit neurons made by stacking two-layered (2D) materials. Nature Nanotechnology has published the findings.

“The demand for new materials, devices, and unique ideas into how they may be utilized creatively makes the entire sector extremely interesting. This research explores the potential of 2D materials for unique computing paradigms rather than transistors, and so constitutes a new toolkit.”

Professor Harish Bhaskaran at the Advanced Nanoscale Engineering Laboratory, University of Oxford,

By making the electronic memristors more responsive to both electrical and optical signals, the researchers improved the memristors’ functionality in the study. This empowered the synchronous presence of discrete feedforward and criticism within the organization. Thanks to the advancement, the team was able to develop neural networks in which the winner takes all: programs for computational learning that have the potential to solve difficult machine learning issues, such as clustering and combinatorial optimization issues that require unsupervised learning.

Because 2D materials are only made up of a few layers of atoms, their exotic properties can be fine-tuned depending on how the layers are laid out. In this study, graphene, molybdenum disulfide, and tungsten disulfide were used in a stack to create a device with a different conductance depending on the power and duration of light or electricity shining on it.
These analog storage devices, in contrast to digital ones, function similarly to synapses and neurons in our biological brains. The simple component considers calculations, where a succession of electrical or optical signals shipped off the gadget produces continuous changes in how much electronic charge is put awayon of electrical or optical signals shipped off the gadget produces continuous changes in how much electronic charge is put away. Similar to how our brain processes a combination of excitatory and inhibitory signals, this process is the foundation for threshold modes in neuronal computations.

“This is a highly exciting development,” said lead author Dr. Ghazi Sarwat Syed, a research staff member at IBM Research Europe Switzerland. A novel idea that goes beyond the fixed feedforward operation that is commonly used in current artificial neural networks has been presented in our study. These current proof-of-principle results demonstrate an important scientific advancement in the broader fields of neuromorphic engineering and algorithms, enabling us to better emulate and comprehend the brain, in addition to the potential applications in AI hardware.”

The experimental work was carried out by Drs. Syed and Yingqiu Zhou, Oxford labmates and DPhil students. Through the optoelectronic physics of low-dimensional systems, their implementation, according to Dr. Zhou, who is now a postdoctoral researcher at Denmark Technical University, captures the essential components of a biological neuron.

They make the observation that our heterostructure stack design has resulted in the formation of atomically abrupt semiconductor junctions. The stack specifically delivers a heterojunction that serves as the neuronal membrane, while the neuronal soma is formed by the graphene electrodes that are in contact with the heterojunction. This way, just like in real neurons, the neuronal state is represented in the soma but changed by membrane changes.

The amount of computational power required for artificial intelligence applications has grown at an exponential rate, outpacing the development of new hardware based on conventional processors. New methods need to be studied right away, like the work done by co-lead author Professor Harish Bhaskaran at IBM Research Zurich and the Advanced Nanoscale Engineering Laboratory at the University of Oxford.

Teacher Bhaskaran said, “This whole field is super-energizing, as material developments, gadget advancements, and novel bits of knowledge into how they can be inventively applied all need to meet up. The power of 2D materials, which are not transistors but rather novel computing paradigms, is the subject of this work, which provides a brand-new toolkit.”

“The use of such 2D structures in computing has been spoken about for years, but only now are we finally seeing the payoff after spending over seven years in development,” stated co-author Professor Jamie Warner of the University of Texas at Austin. By gathering wafer-scale 2D monolayers into complex ultrathin optoelectronic gadgets, this will empower the beginning of new data handling approaches utilizing 2D materials in view of economically adaptable manufacturing strategies.”

“Our discoveries have a greater exploratory nature than real exhibits at the framework level,” Dr. Syed says. “We are certain that our current proof-of-principle results demonstrate an important scientific interest in the broader fields of neuromorphic engineering, enabling us to better emulate and comprehend the brain. Although we intend to expand on this concept in the future,

Professor Bhaskaran says that exciting new research is important for new ideas in the future, but this is not technology that should be in mobile phones in two years.

More information: Ghazi Sarwat Syed et al, Atomically thin optomemristive feedback neurons, Nature Nanotechnology (2023). DOI: 10.1038/s41565-023-01391-6

Topic : Article