With the development of artificial intelligence, the internet of things, cloud computing, and other new technologies, human society is developing from information to intelligence. The construction of an intelligent society poses higher requirements for the development of information technology. The system needs to high-efficiently collect, process the external environment information in real-time, and make timely decisions. Correspondingly, the amount of collected data boosting exponentially. Thus, the computing power and energy efficiency of the conventional system with separated sensing, computation, and memory will not meet the real-time processing requirements of intelligent terminals. Studying a low-power intelligent information processing system that integrates sensing, computation and memory is a considerable trend in information technology. At present, the development of sensors provides the possibility for real-time collection of perceptual data information. For further processing sensed data, spiking neural networks (SNNs), due to its event-driven and sparse coding, have become a promising candidate to build a high-efficient processing unit, which is also considered as the next generation of neuromorphic computing technology. To realize the integrated intelligent processing system with the spiking mechanism, building an efficient interface (called afferent nerve in biology) to bridge the gap between the spiking data processing unit (SPU) and the sensor is urgent. The interface can convert the analog signal sensed by the sensor into the spiking signals that could be processed by the SPU in real-time. However, due to the limited inherent dynamics of traditional CMOS devices, the conventional interface circuits are complex and not easy to be integrated into large-scale. Moreover, CMOS devices nearly reach its physical bottleneck. To build an efficient spiking sensing interface, developing new principle devices with inherent dynamic characteristics and scalability is an effective solution.
Recently, Prof. Ming Liu’s group, from the Key Laboratory of the Institute of microelectronics of Chinese Academy of Sciences (IMECAS), and Prof. J. Joshua Yang’s group from the University of Massachusetts, Amherst, using a NbOx Mott memristor developed by the IMECAS, teamed up to build an high-efficient artificial spiking afferent nerve (ASAN) for the interface between the sensors and the SPU. NbOx Mott memristor, as an emerging device, features simple structure, high scalability and reliability, high integration density, and rich dynamics (insulation-metal-transition). The constructed interface circuits successfully convert the analog electrical signals sensed by sensors to the frequency-dependent spiking signals, which simplifies the information interface between the sensor and the SPU. The frequency has a quasi-linear relationship to the input voltage under the nonharmful input intensity and tends to stop firing under the harmful input intensity, closely resembling the biological afferent nerve. To verify the application of the interface, the researchers further connect the ASAN with a piezoelectric sensor to build an artificial spiking mechanoreceptor system without any external power supply. The system can respond to the pressure signal and convert the pressure intensity into the corresponding spiking frequency for being further processed by the SPU.
In addition, the interface circuit can be easily extended to process signals from various sensors, such as smell, taste, sight, hearing, temperature, magnetic field, and humidity. Except for constructing a variety of sensor systems, the ASAN (interface circuit) also features the leaky integration-and-fire spiking behavior of neurons, which is suitable for the realization of SNNs. Therefore, the ASAN can be further used to construct complex SNNs to process information and has a great potential to realize a high-efficient intelligent terminal system that integrates sensing, computation, and memory.
The above research results were published in Nature Communications on January 2, 2020 (DOI:10.1038/s41467-019-13827-6). Xumeng Zhang, a Ph.D. candidate in the IMECAS, is the first author of this article. Ming Liu (an Academician of Chinese Academy of Sciences in IMECAS), Qi Liu (a researcher in IMECAS), and J. Joshua Yang (a professor at the University of Massachusetts) are co-authors. This work has been partly funded by relevant projects such as NSFC, Ministry of science and technology, Chinese Academy of Sciences, etc.
a. Schematic of biological mechanoreceptor system. The receptors in the skin receive external stimulation, and the afferent nerve generates action potential and transmits it to the cortex system for further processing.
b. Schematic of artificial spiking mechanoreceptor system, including piezoelectric sensor and afferent spiking afferent nerve (ASAN). The ASAN generates spiking signals that can be processed by SNNs.
Source：https://www.nature.com/articles/s41467-019-13827-6
