A research team at Korea Advanced Institute of Science and Technology (KAIST) has developed a highly-density neuromorphic semiconductor that imitates the human brain.

Professors Choi Yang-kyu and Choi Sung-Yool of the School of Electrical Engineering at KAIST led the research.

Neuromorphic hardware is an artificial intelligence function made in the form of hardware that imitates the human brain. This originates from the idea that the human brain can accomplish complicated tasks but consumes only 20 watts of energy. Unlike the traditional von Neumann method, neuromorphic hardware draws attention because it can carry out AI functions with very low electricity.

The joint-research team developed neuromorphic hardware of neurons and synapses using a single transistor. The hardware was produced with a commercialized standard silicon process, increasing the possibility of commercializing neuromorphic hardware systems.

Han Joon-Kyu, a Ph. D candidate at the School of Electrical Engineering at KAIST, was the lead author, and Oh Jung-yeop was the second author of the research. The research was published in the August online edition of Science Advances under the title of “Cointegration of single-transistor neurons and synapses by nanoscale CMOS fabrication for highly scalable neuromorphic hardware.”

To materialize neuromorphic hardware, synapses that remember the connection between neurons that produces spikes when signals synchronize are needed. However, because neurons and synapses that have footholds in digital or analog circuits take up a lot of space, there are limits in their degree of integration. Since the human brains use about 100 billion neurons and 100 trillion synapses, the degree of integration must be improved for use in mobiles and the Internet of Things (IoT).

To do so, various materials and structural bases of neurons and synapses were suggested. Still, most of them could not be made with the standard Si CMOS process, making it difficult to commercialize and mass-produce them.

To solve the problem, the research team utilized a single transistor widely used for the standard Si CMOS process to imitate the movements of biological neurons and synapses. Neuromorphic hardware was produced in the fabricated 8-inch wafer through co-integration.

The newly produced neuromorphic transistor has the same structure as currently made transistors for memory and system hardware. The transistor can precede memory function and logical operation. This research has its meaning in showing the possibility of new neuromorphic simulation. Applying a new principle of operation on the currently produced transistors, a new neuromorphic transistor was made, with a similar structure but different functions. Like a coin that has a head and tail, the neuromorphic transistor proved for the first time in the world that it could function as both neuron and synapse.

The degree of integration was highly increased by substituting the neuron, based on complicated digital and analog circuits, with a single transistor. In addition, this new technology can reduce the cost by simplifying the production process with the same structure of synapses. While the current neuron circuit needs 21,000 units of two-dimensional space, the newly developed neuromorphic transistor only requires six units. The degree of integration is 3,500 times higher.

Using the neuromorphic hardware, the research team simulated brain functions like gain modulation and coincidence detection. It also proved face image recognition and letter pattern recognition possible.

KAIST expects the neuromorphic hardware to increase integration and decrease the cost, moving a step closer to their commercialization.

“Using complementary metal-oxide-semiconductor (CMOS)-based single transistor, we proved the possibility of neuron and synapsis simulation,” Han said. “Using the commercialized CMOS process, neuron, synapses, and co-integrating additional circuits at the same time in one single wafer, we increased the degree of integration in neuromorphic hardware, moving one step closer to commercializing the neuromorphic hardware.”