A promising optoelectronic synapse for reservoir computing based on alpha-indium selenide

Reservoir computing is an emerging computational framework based on the use of recurrent neural networks (i.e. networks in which data persists or reproduces according to patterns). This framework has the potential to reduce data processing time, while improving the energy efficiency of neuromorphic devices.

Researchers from Peking University and the Beijing Academy of Artificial Intelligence recently introduced a novel artificial synapse based on alpha-indium selenide (α-In₂Se₃ ), which could help more efficiently recreate biological neural processes in neuromorphic devices. This synapse, presented in an article published in Natural electronicscould have very valuable implications for reservoir calculation applications.

“Our idea came from the need for a simple strategy that can be used to exploit the dynamic responses of a physical system for computation, and computation of physical reservoirs is a promising framework for achieving this goal,” said Yuchao Yang, one of the researchers who conducted the study, TechXplore told TechXplore.

“In₂Se₃ is a very interesting material and a good platform for reservoir computing, and its rich physical properties support the creation of multi-mode and multi-scale reservoir computing systems, which we hope will expand application scenarios of physical reservoir computing.”

Reservoir computing is based on the use of artificial synapses capable of directly executing deep learning algorithms, without having to transfer data between a memory and a processing unit. The van der Waals semiconductor material α-In₂Se₃ has many advantageous optoelectronic, ferroelectric and semiconductor properties, which make it an ideal candidate for the fabrication of these artificial synapses.

“In₂Se₃ simultaneously has two interesting inherent physical properties, namely ferroelectric switching and optoelectronic response,” Yang explained. “We built a planar device to use in-plane ferroelectric polarizations for the electrical synapse, while also introducing light as a third terminal to enable optoelectronics. answer. This unique structure effectively combines the two physical properties and can exploit the coupling of ferroelectrics and optoelectronics for heterosynaptic plasticity and higher-level computational functionality.”

The temporal dynamics of the optoelectronic synapse created by this team of researchers can be controlled using electrical and optical stimuli. This means that it can ultimately artificially replicate the brain’s innate plasticity (i.e. its ability to adapt over time), while directly processing information.

“A large majority of previous research in neuromorphic computing only used the device as a non-volatile element, whereas we take advantage of more complex nonlinear dynamics to power computing,” Yang said.

“Compared to previous reservoir systems with a fixed mechanism and functions, our optoelectronic synapse has both ferroelectric and optoelectronic properties, thus providing two distinct and physically coupled mechanisms for reservoir computation. This allowed us to achieve a reservoir computing system based on mixed signal input, demonstrating high adaptability and improved network performance, while here multi-scale signal processing is achieved by modulating the relaxation time of the device with a voltage light or rear grill.

To evaluate the performance of their artificial synapse, Yang and his colleagues used to build a multi-mode reservoir computing system. They then tested the performance of this system in a handwritten digit recognition task and a QR code recognition task. They found that it achieved promising results, successfully tackling both of these data processing tasks with accuracies above 80%.

The artificial synapse made in this study may soon open up exciting new possibilities for reservoir computation. Additionally, the reservoir computing system created using this synapse could be further developed to tackle other complex information processing and data analysis tasks.

“Our demonstration of a multi-mode, multi-scale reservoir computing system fundamentally expands the processing capability of reservoir computing systems,” Yang added. “In our recent study, we focused on computational applications, but in the future we would also like to realize a fully integrated neuromorphic system, including information sensing.”

© 2022 Science X Network