The research on new sensor computing devices of the Semiconductor Institute has made progress

With the development of new information interaction fields such as artificial intelligence, Internet of Things and smart medical treatment, it is increasingly difficult for computer systems based on the traditional von Neumann architecture and the increase in computing power brought about by process iteration to meet the needs of data processing and complex neural network model operations. As an efficient and low-power information processing model that simulates the human brain, neuromorphic devices have natural advantages in information processing. At present, artificial synaptic devices represented by memristors are widely used in neuromorphic computing and build multiple types of neural networks. However, traditional artificial synaptic devices store fixed weights, redeployment is time-consuming and laborious, and cannot be adaptively adjusted to input changes.

Similar to synapses, charge-based semiconductor energy storage devices enable the regulation and retention of storage weights under low-energy conditions. The unique properties of ion migration make it possible to build artificial synaptic devices that simulate synaptic gap information transmission. Studies have shown that battery-like energy storage devices can be used as artificial synaptic devices for low-energy calculations. Therefore, the use of semiconductor energy storage devices to design a new sensor computing system to solve the problems of high write noise, nonlinear difference and diffusion under zero bias will be an important research direction in the field of brain-like computing.

Recently, Wang Lili, researcher of the State Key Laboratory of Semiconductor Superlattices of the Institute of Semiconductors of the Chinese Academy of Sciences, Shen Guozhen, professor of Beijing Institute of Technology, and Fan Zhiyong, professor of the Hong Kong University of Science and Technology, cooperated to design a new sensor computing integrated system based on the adjustable flexible energy storage device (FMES) system using micro-nano processing (Figure 1). The system realizes the control of ion accumulation and dissipation through the regulation of the resistance value in the system without changing the external stimulus, and is expected to realize the coupling of the sensing signal and the storage weight W. The FMES system can be used to build neural networks that implement a variety of neuromorphic computational tasks, making the recognition accuracy of the handwritten set of numbers to about 95% (Figure 2). In addition, the FMES system can simulate the adaptability of the human brain to achieve adaptive recognition of similar target datasets, and the adaptive recognition accuracy after training can reach about 80% (Figure 3), avoiding the time and energy loss caused by recalculation. Future research can be based on this achievement, combined with the integration of different types of sensors on chip, to further realize the multi-modal sense of the integrated architecture.

Figure 1: A new sensor computing integrated system based on an adjustable flexible energy storage device (FMES) system. a, the structure of the biological synapse, b, FMES device schematic, c, FMES device optical image, d, read and write operation decoupling schematic diagram and corresponding circuit diagram, e, different voltage pulses under the post-synaptic current.

Figure 2: Accuracy of neuromorphic calculations. a. Neural network structure; b. Hardware neural network composed of FMES devices; c. Numerical mapping of synapse weights in various resistance states of initial and training; d. Classification accuracy of 100 training cycles, PSV distribution of FMES equipment: pre-training (e) and post-training (f).

Figure 3: Adaptive simulation of digital recognition. a, Dataset1 and dataset2, b, the weight mapping image of the artificial neural network, the recognition accuracy of c, dataset1 and dataset2, d, the adaptive adjustment of the weight of the artificial neural network, e, the adaptive weight mapping image, f, the relationship between the recognition accuracy and the number of map images, g, the classification recognition before and after adaptation, h-i, the probability of each sample.

The research results were published in the National Science Review under the title of Neuromorphic-Computing-Based Adaptive Learning Using Ion Dynamics in Flexible Energy Storage Devices. (Source: Institute of Semiconductors, Chinese Academy of Sciences)

Related paper information:

Special statement: This article is reproduced only for the purpose of disseminating information, and does not mean to represent the views of this website or confirm the authenticity of its contents; If other media, websites or individuals reprint from this website, they must retain the “source” indicated on this website and bear their own legal responsibilities such as copyright; If the author does not wish to be reprinted or contact us for reprint fees, please contact us.

Source link

Related Articles

Leave a Reply

Your email address will not be published. Required fields are marked *

Back to top button