Researchers develop atomically thin artificial neurons for enhanced neural network computing

Scientists have achieved a significant breakthrough in the field of artificial neurons by developing atomically thin versions capable of processing both light and electric signals. These ultra-thin neurons, created through the stacking of two-dimensional materials, offer enhanced computational capabilities and can handle complex problem-solving tasks.

The integration of feedforward and feedback pathways, crucial for cognitive abilities such as learning from rewards and errors, has been a major challenge in replicating the complexity of biological neurons. However, researchers from the University of Oxford, IBM Research Europe, and the University of Texas have successfully addressed this obstacle. By making electronic memristors responsive to optical and electrical signals, they have enabled the coexistence of separate feedforward and feedback paths within neural networks. This breakthrough has paved the way for the development of winner-take-all neural networks, which have the potential to tackle intricate problems in machine learning, including unsupervised learning and combinatorial optimization. The findings of this study have been published in the journal Nature Nanotechnology.

In this groundbreaking study, scientists utilized the unique properties of two-dimensional (2D) materials, which consist of just a few layers of atoms. By stacking three specific 2D materials (graphene, molybdenum disulfide, and tungsten disulfide), they engineered a device that exhibited changes in conductance based on the intensity and duration of light or electricity exposure.

Unlike traditional digital storage devices, these analog devices function in a manner similar to the synapses and neurons found in our biological brains. Their analog nature enables computations where a series of electrical or optical signals induce gradual modifications in the stored electronic charge. This mechanism forms the foundation for threshold modes in neuronal computations, mimicking the brain’s processing of excitatory and inhibitory signals.

Dr. Ghazi Sarwat Syed, the lead author and a Research Staff Member at IBM Research Europe Switzerland, expressed immense excitement about this breakthrough. The study introduces a new concept that surpasses the fixed feedforward operation commonly employed in existing artificial neural networks. Beyond its potential applications in AI hardware, these initial results signify a significant scientific advancement in the fields of neuromorphic engineering and algorithms, enhancing our ability to emulate and comprehend the workings of the human brain.

The experimental work for this groundbreaking research was carried out by Dr. Ghazi Sarwat Syed and Dr. Yingqiu Zhou, who were DPhil students and lab colleagues at Oxford. Dr. Zhou, currently a Postdoctoral researcher at Denmark Technical University, emphasizes that their implementation captures the fundamental aspects of biological neurons through the optoelectronic physics of low-dimensional systems.

The researchers achieved atomically precise semiconductor junctions by designing a heterostructure stack. This stack incorporates a heterojunction, acting as the neuronal membrane, while the graphene electrodes that interface with the heterojunction function as the neuronal soma. This setup allows the neuronal state to be represented in the soma but influenced by changes in the membrane, akin to the behavior of actual neurons.

As the demand for computational power in artificial intelligence applications has surged, traditional processors have struggled to keep pace. This underscores the urgent need for exploring new techniques, such as the ones investigated by co-lead author Professor Harish Bhaskaran from the Advanced Nanoscale Engineering Laboratory at the University of Oxford and IBM Research Zurich laboratory.

Professor Bhaskaran expresses enthusiasm, stating that the entire field is highly exciting as it combines materials innovations, device innovations, and novel insights into their creative applications. This work introduces a new toolkit that harnesses the power of 2D materials not only in transistors but also for novel computing paradigms.

Co-author Professor Jamie Warner from the University of Texas at Austin highlights the significant progress made after over seven years of development, where the use of 2D structures in computing has been discussed for a long time. By assembling wafer-scale 2D monolayers into intricate ultrathin optoelectronic devices, this research opens up new possibilities for information processing using 2D materials, thanks to industrially scalable fabrication methods.

Dr. Syed acknowledges that their findings are exploratory in nature and represent proof-of-principle results rather than fully realized system-level demonstrations. However, he believes that these results hold great scientific interest in the broader fields of neuromorphic engineering, enabling a better understanding and emulation of the brain.

Professor Bhaskaran points out that while these research developments are exciting for future innovation, it is important to note that this technology is unlikely to be integrated into mobile phones within the next two years.

Source: University of Oxford

Leave a Comment