Research published in Nature Nanotechnology shows the creation of atomically thin artificial neurons by stacking two-dimensional (2D) materials to expand the functionality of electronic memristors by making them responsive to both optical and electrical signals for the operation of winner-take-all neural networks with separate feedforward and feedback paths within the network. 

 For years, scientists have explored how to create machine learning systems that mimic the computational abilities of biological neurons, with the goal of making them faster and more energy-efficient. Memristors, electronic components that can store a value by modifying their conductance, have shown promise in achieving this goal. However, a significant obstacle has been integrating both feedforward and feedback neuronal signals, which are necessary for complex cognitive abilities such as learning through rewards and errors.

Now, researchers from the University of Oxford, IBM, and the University of Texas at Austin have developed atomically thin artificial neurons using stacked two-dimensional (2D) materials. By making the electronic memristors responsive to both optical and electrical signals, the team was able to create separate feedforward and feedback paths within the network, enabling the development of winner-take-all neural networks capable of solving complex machine learning problems.

The 2D materials used in the study, such as graphene, molybdenum disulfide, and tungsten disulfide, have unique properties that can be fine-tuned by layering them. In this case, the researchers used a stack of three 2D materials to create a device that responds to light by changing its resistance, much like an artificial retina. Unlike digital storage devices, these devices are analog and operate similarly to biological synapses and neurons, allowing for computations based on gradual changes in stored electronic charge.

Lead author Dr Ghazi Sarwat Syed, now a Research Staff Member at IBM Research Switzerland, said the team's work represented a significant scientific advancement in neuromorphic engineering and had the potential for various AI hardware applications. Dr Yingqiu Zhou, a former DPhil student and lab colleague at Oxford, added that their implementation captured the essential components of biological neurons and was based on the optoelectronic physics of low-dimensional systems.

As the demand for computational power in AI applications has grown, traditional processors have struggled to keep up, making new hardware research critical. Co-lead author Professor Harish Bhaskaran at the Advanced Nanoscale Engineering Laboratory at Oxford said the entire field was "super-exciting," with the need for materials and device innovations, as well as creative applications. Co-author Professor Jamie Warner at the University of Texas at Austin said the use of 2D structures in computing had been discussed for years, but the development of new information processing approaches using 2D materials based on industrially scalable fabrication methods was now possible thanks to advances in materials production that yield wafer scale monolayers.

Full article can be found here:

Atomically thin optomemristive feedback neurons

Ghazi Sarwat Syed, Yingqiu Zhou, Jamie Warner, Harish Bhaskaran, Nature Nanotechnology (2023)

https://www.nature.com/articles/s41565-023-01391-6