柔性人工突触: 面向智能人机交互界面和高效率神经网络计算的基础器件

Benefiting from the fast progress of artificial intelligence, a number of evolutions in the areas of human-machine interface, bio-inspired sensing systems, robots and prosthetics have been achieved. However, because of data explosion and the requirement for intelligent human-machine interaction, novel technology should be developed to overcome current bottlenecks. Different from existing extensive-energy-consumption neural networks implemented at the software level based on the conventional von Neumann architecture, the human brain only has a power consumption of about 20 W. Therefore, it is of a great importance to design a new neuromorphic computing system that is executed by parallel operation with high speed and low power consumption resemble to the human brain. Artificial synapses, either based on transistor or memristor structure, can be used as basic building blocks to achieve large-scale neural network parallelism. In addition, the spike based information processing in biological systems can also be mimicked with the employment of thue artificial synapses, which is beneficial to the construction of intelligent human-machine interface. Therefore, much efforts have been made to optimize the performance of the synaptic devices in terms of materials, fabrication process and structures. Consequently, a series of biorealistic synaptic behaviors, such as visual information reprocessing, movement control and learning-forgetting process, have been emulated using flexible artificial synapses. Despite the great achievement in the study of artificial synapses, several underlying mechanisms have not been uncovered. First, the modulation of the post-synaptic signals varies from each individual, which requires specific analysis in order to make it to be compatible with the neural signal. In addition, dendrites in biological systems can collect, integrate, and modulate thousands of pre-synaptic input signals, and transmit these signals to post-synaptic neurons. That is to say, spatiotemporal information can be modulated. Therefore, exploration of the underlying mechanism and optimization of the device structure mimicking the biological dendrite can contribute to the simulation of dynamic logic induced by spatiotemporal synaptic stimulation. Besides, most of the reported researches were performed on the rigid substrates, which are not compatible with the biological systems. Therefore, fabrication of the devices on flexible substrates and investigation of the relationship between the electrical properties and interface quality are critical. Herein, an overview of the recent progress of the artificial synapses is presented in terms of device structure, material selection, and working mechanism. Future challenges, research directions, and possible applications are also discussed. This review is hoped to provide a guidance for the design and fabrication of the flexible artificial synapses towards neuromorphic computing and intelligent human-machine interface.