A team of engineers at the Pennsylvania State University introduced memory resistors (memristors) fabricated from graphene - mimicking neural networks through a similarly analog nature as the brain itself.
The research team demonstrated the possibility of a multi-level non-volatile memristive synapses, according to their paper, published in the journal Nature Communications. Digital states, which can either be on or off or a value of one or zero, is the basis of modern computing systems. On the other hand, an analog computer - like the human brain - allows for any number of possible states. Instead of a discrete one or zero state, analog computers can present any continuous value between one and zero.
Meeting The Growing Computing Demands
Neuromorphics - computing technologies that mimic how the brain works - has been around for more than four decades, as explained by Saptarshi Das, team leader in the study and an assistant professor of engineering science and mechanics at Penn State. In fact, neuromorphics was a concept first developed by Caltech scientist and engineer Carver Mead to describe the use of VLSI systems holding electronic analog circuits that emulate the neurobiological architecture of the brain.
The accompanying press release from Penn explains that although the limitations of digital computing systems are apparent, the need for faster data processing and larger memory requirements underscores the need for more advanced systems - such as those that mimic the complex workings of the brain.
"We have powerful computers, no doubt about that, the problem is you have to store the memory in one place and do the computing somewhere else," noted Das.
To eliminate this conflict, the Penn State team turned to artificial neural networks (ANN), which, according to first author Thomas Shranghamer, seeks "to emulate the energy and area efficiencies of the brain." He compared the actual size of a human brain to modern supercomputers that occupy incredibly large spaces.
Graphene Memristive Synapses
Researchers achieved a programmable conductance over a group of graphene field effect transistors (GFET). As a memory device, it can be programmed to retain more than 16 conductance states, surpassing digital binary states in conventional oxide-based memristors. To create their GFETs, researchers fabricated a chemical vapor deposition (CVD) graphene, which was then moved to a 50 nanometer alumina substrate. The alumina works as the back-gate oxide layer, in contrast with a stack that contains platinum, titanium nitride, and silicon (Pt/TiN/p++-Si) that is the back-gate electrode.
By applying a controlled electric field on the graphene sheet, the fabricated ANN can be reconfigured in a similar way as the synapses - the gap between neurons where the electric nerve impulse is transmitted.
Researchers also demonstrated the graphene memristive synapses' retention capability and its switching endurance. Also, as an artificial neural network, researchers demonstrated the memristors' capability for weight assignment through the use of k-means clustering. K-means refer to a method of vector quantization and one of the most commonly used algorithms especially in unsupervised machine learning systems.
"What we have shown is that we can control a large number of memory states with precision using simple graphene field effect transistors," Das noted. Researchers are optimistic that with the growing interest in neuromorphic computing, their work will gain support and be ramped up to a commercial scale.
