The team of Academician Huang Ru and Professor Yang Yuchao has made important progress in the research of biomimetic photosynapses and multi-modal, multi-scale reserve pool calculation

The dynamic evolution of physical systems can be applied to information processing
as an efficient computing resource.
Driven by input, the physical system evolves according to internal laws, and the result of evolution realizes the transformation of input information and serves as an important basis for
achieving specific computing functions.
Hardware neural networks built on this are more energy efficient and faster than
digital systems.
Reservoir computing is one of the computing paradigms that has attracted much attention, which has the advantages of easy training, low hardware overhead, and can be implemented
using a variety of electrical and optical hardware.
The key characteristics required by reserve pool computing hardware are nonlinearity, which maps input information nonlinearly into high-dimensional space, and short-term memory, which means that the state of the system is determined
by both the current input and the most recent input.
Dynamic memristor devices that meet the above requirements are used in hardware to implement reserve pool calculation, and show broad application prospects, and have made significant progress
in speech digital recognition, chaotic system prediction, solving second-order nonlinear tasks and real-time neural activity analysis.

However, most of the current memristor devices are single physical mechanisms, lack of cross-modal synaptic plasticity, lack of relaxation time tunability, which leads to the reserve pool based on memristor still faces problems such as nonlinear mapping mode fixation, inability to process multi-modal signals, and lack of multi-scale feature extraction ability, thus limiting the information processing ability of the reserve pool, and it is difficult to realize a multi-modal and multi-scale reserve pool computing system
.

Figure 1.
α-In₂SE₃ photomemristor and multimodal, multi-scale reserve pool calculation

In view of the above key problems, the research group of Professor Huang Ru-Yang Yuchao, academician of the School of Integrated Circuits/Advanced Innovation Center of Integrated Circuits, Peking University, constructed a multifunctional photoelectric memristor device based on α-In₂SE₃, and used its dynamic characteristics of photoelectric coupling to realize multi-modal fusion and multi-time scale reserve pool calculation
.
Among them, van der Waals semiconductor α-In₂Se₃ has both ferroelectric and optoelectronic characteristics, and the memristor device based on this structure embodies rich synaptic characteristics under the excitation of electrical and optical pulses, including PPF, PPD, EPSC, IPSC, LTP/LTD and other long- and short-term plasticity
.

Figure 2.
α-In₂SE₃ photomemristor with rich synaptic properties

More importantly, the physical mechanism of photohybrid enables the device to have cross-modal synaptic plasticity, providing a variety of nonlinear mapping modes
for reserve pool calculation.
Based on this, the research team proposed a technical scheme of multi-modal fusion, constructed a reserve pool computing system for multi-modal fusion, and successfully realized the multi-modal handwriting digit recognition and two-dimensional code recognition tasks, which reached 86.
1% and 98.
6% accuracy, respectively, and the accuracy remained basically unchanged
under low fluctuations (C2Cσ/μ=0.
1).
In addition, the back-gate voltage/illumination as the third end can further achieve heterologous synaptic plasticity, thereby effectively regulating the DC and pulse characteristics of the device, especially the relaxation time of the device can be shortened by more than an order of magnitude
.
The device solves the problem that the relaxation time of the existing memristor is relatively fixed, and further realizes the multi-scale reserve pool calculation, which can capture the characteristics of MSO 5 time series at different scales by placing the device at different relaxation times_{, so as to successfully realize the accurate prediction of the sequence (NRMSE=0.
105).}

Figure 3.
Multimodal handwritten digit recognition and MSO₅ time series forecasting

This study uses multiple physical mechanisms in the α-In₂SE₃ memristor to realize cross-modal synaptic plasticity and relaxation time tunability, providing a variety of nonlinear mapping modes, innovative schemes of multimodal fusion, and multi-scale feature extraction capabilities, which lays a solid foundation
for the application of reserve pool computing to more complex tasks and broader scenarios.

The results, titled "An optoelectronic synapse based on α-In₂Se₃withcontrollable temporal dynamics for multimode and multiscale reservoir computing," were published in Nature Electronics Electronics).

Liu Keqin, a 2018 doctoral student at the School of Integrated Circuits of Peking University, is the first author, and Huang Ru and Yang Yuchao are the corresponding authors
.

Huang Ru-Yang Yuchao's team has been deeply engaged in the research of memristors, brain-like computing, storage and computing integrated intelligent chips for a long time, and has published more than 130 papers in journals and conferences such as Nature Electronics, Nature Reviews Materials, Nature Nanotechnology, Nature Communications, Science Advances, and IEDM.
2 selected TOP0.
1% ESI hot papers, 11 selected TOP1% ESI highly cited papers, the research work has formed an important influence
in the world.

The research work was supported
by the National Key Research and Development Program of China, the National Outstanding Youth Fund, the Post-Moore Major Research Program of the National Natural Science Foundation of China, the 111 Program, and other projects, as well as the Peking University-Baidu Foundation, Fok Ying Tung Education Foundation and Tencent Foundation.