Associate Prof. S.N. Piramanayagam (left) and Assistant Prof. Renshaw Wang with the setup where the electrical measurements were carried out. Photo credit: M.Fadly.
Neuromorphic computing is an emerging technological paradigm that aims to create computers inspired by the functioning of the human brain. In a new study, researchers from Nanyang Technological University (NTU Singapore) and the Indian Institute of Science (IISc) have developed a tiny electronic device mimicking a synapse, an elementary connection within the brain. For the first time, the team incorporated a metallic channel into an electrolyte-gated transistor, resulting in a device that can efficiently learn and forget as brains do.
As modern computers grow increasingly powerful, their high power consumption and the associated generation of waste heat have become ever more serious problems. Artificial intelligence (AI) software is particularly power-hungry: for example, the AI system that first beat a professional human player at Go in 2017 consumed 170 kW of power, the equivalent of 2800 incandescent light bulbs. By contrast, the human brain consumes only around 20 W – one-third of an incandescent light bulb – while taking up a much smaller volume.
One of the main reasons computers use so much power is that they are based on a 70-year-old scheme called the von Neumann architecture, which delegates computation and memory storage to separate devices. Many computing tasks, especially in AI, require data to shuttle frequently between the CPU and memory units. By analogy, imagine you are cooking and are required to run to a grocery store each time you need another ingredient! Despite the great success of the von Neumann architecture over the past seven decades, its high power expenditure and waste heat have caused researchers to look for alternatives.
One possible solution is to take inspiration from biological brains, in which neurons and synapses serve respectively as the computing units and memory units. Unlike in the von Neumann architecture, neurons and synapses are not packaged in different locations, but are closely intertwined – specifically, synapses form the connections between neighboring neurons. This layout handily avoids the data transfer bottlenecks of the von Neumann architecture. Neuromorphic computing is an emerging research field that aims to develop computers based on artificial neurons and synapses.
The best way to implement artificial neurons and synapses, however, remains an unresolved question. In a notable advance, the NTU/IISc team have developed an artificial synapse based on a common electronic component called an electrolyte-gated transistor. Their work was published in the journal ACS Applied Materials and Interfaces in March 2022.
To create the artificial synapse, the researchers introduced a small but significant tweak to the usual design of an electrolyte-gated transistor: the replacement of a semiconductor “channel” with a metal channel made from a cobalt thin film. This metal channel is compatible with current silicon technologies, making the device cheap and simple to manufacture.
A key advantage of using cobalt in the device is that it can have magnetic properties different from conventional semiconductors. This allows information reading and processing to take place in a more energy-efficient manner.
Left: Schematic of the biological synapse as the functional connection between neurons. Right: Learning, forgetting, and relearning are realized in the artificial synapse similar to the human brain.
In a series of experiments, the authors showed that their transistor device is capable of performing operations that can be characterized as “learning”, “forgetting”, and “relearning”, which are some of the fundamental cognitive functions occurring in biological brains.
“By applying voltage pulses to the gate electrolyte, we can modulate the conductance of the device by sending ions into or out of the cobalt channel,” says S.N. Piramanayagam, associate professor in Physics and Applied Physics, NTU and corresponding author of the study. “These changes in conductance mimic the chemical activity of biological synapses.”
By applying a sequence of voltage pulses to the device, the researchers were able to emulate the distinction between short-term and long-term memory. Short-term memory plays a crucial role in how the human brain processes immediate tasks, whereas long-term memory is responsible for learning. Working in a similar manner, the transistor can transfer information between short-term to long-term memory.
As P. Monalisha, the PhD student who carried out the research, explains: “For these devices, it can be easier to retrieve a long-stored memory than a fresh memory, which is quite similar to how our own memory works.”
The project was led by Associate Professor S. N. Piramanayagam and Nanyang Assistant Professor Xiao Renshaw Wang from NTU’s School of Physical and Mathematical Sciences (SPMS). In the future, the researchers aim to explore the use of the cobalt-based synaptic transistor for practical neuromorphic computing tasks.
About the author:
Associate Professor S.Piramanayagam is faculty member at the School of Physical and Mathematical Science, NTU Singapore. He has 30 years of experience in the field of magnetism. He has carried out research in magnetic nanostructures, magnetic recording and spintronics materials.