Implement artificial neural network hardware systems by stacking them as “neuron-synapse-neuron” structural blocks

With the emergence of new industries such as artificial intelligence, Internet of Things and machine learning, the world’s largest companies are focusing on developing next-generation artificial intelligence semiconductors capable of processing large amounts of data while efficiently consuming power.

Neuromorphic computing, inspired by the human brain, is one of them. As a result, devices mimicking neurons and biological synapses are being developed one after another based on emerging materials and structures, but research into integrating individual devices into a system to verify and optimize them is still under way. default.

For large-scale artificial neural network hardware to become practical in the future, it is essential to integrate artificial neurons and synaptic devices, and it is necessary to reduce mass production costs and energy consumption by manufacturing devices with the same materials and structures.

A team led by Dr. Joon Young Kwak from the Neuromorphic Engineering Center at the Korea Institute of Science and Technology (KIST) has implemented embedded element technology for artificial neuromorphic devices capable of connecting neurons and synapses like “Lego blocks” to build large-scale artificial neural network hardware. The study is published in the journal Advanced functional materials.

The team fabricated vertically stacked memristor devices using hBN, a two-dimensional material advantageous for high integration and ultra-low power implementation, to demonstrate biological neurons and synapse characteristics.

Since the team designed artificial neurons and synaptic devices with the same material and structure, unlike conventional CMOS silicon-based artificial neuron imitation devices with complex structures using multiple devices, the devices developed by the team ensured ease of process and scalability of the network, paving the way for the development of large-scale artificial neural network hardware.

By integrating and connecting the developed devices, the team also successfully implemented the “neuron-synapse-neuron” structure, the basic unit of an artificial neural network, in hardware to demonstrate information transmission based on a spike signal, like the human brain does. works.

By experimentally verifying that the modulation of spike signal information between two neurons can be adjusted based on the synaptic weights of the artificial synaptic device, the researchers show the potential of using emerging hBN-based devices for hardware systems of Large-scale, low-power AI. .

“Artificial neural network hardware systems can be used to efficiently process large amounts of data generated in real-world applications such as smart cities, healthcare, next-generation communications, weather forecasting, and autonomous vehicles,” said Dr. Joon Young Kwak of KIST. .

“This will help improve environmental issues such as carbon emissions by significantly reducing power consumption while surpassing the scale limits of existing silicon CMOS-based devices.”