Figure 1 Schematic diagram of sound localization integrated with brain storage and calculation
The research group of Associate Professor Gao Bin of Tsinghua University Institute of Integrated Circuits used the continuous conductance modulation characteristics of memristors to construct a new type of brain like computing system based on memristors array, which was applied to the task of sound localization: the network consisted of 60 input neurons and 7 output neurons, and the input and output information were the acoustic signals received by both ears and the azimuth angle in the range of -90 degrees to 90 degrees, respectively, and all output layer neurons * * * decided the prediction angle.
Figure 2. The hardware implementation of sound localization network based on memristor array is inspired by brain mechanism.
Although the traditional memristor programming strategy can greatly reduce the hardware overhead of online training, it is difficult to meet the accuracy requirements of positioning tasks. In order to overcome this problem, the research group proposed a multi-threshold updating strategy that tolerated the discreteness of memristors. In the process of weight updating, multiple judgment thresholds are introduced, and a corresponding number of operation pulses are applied according to the interval where the weight updating value is located. The analysis results show that the multi-threshold updating strategy achieves the balance between the training accuracy and hardware overhead of voice network.
Fig. 3 is a multi-threshold updating strategy considering memristor characteristics and hardware overhead.
In 1K TiN/TaOy/HfOx/TiN memristor array, the research team successfully demonstrated the task of sound localization for CIPIC HRTF data set samples. The experimental results show that compared with the traditional training scheme, the accuracy of network detection is improved by about 45.7% after adopting the multi-threshold updating strategy. In addition, compared with the previous CMOS ASIC scheme, this technology shows higher positioning accuracy and energy consumption advantage of 184 times. This work provides a new solution for realizing a brain-like positioning system with low energy consumption and high performance.
Fig. 4 Demonstration of sound localization function and hardware evaluation results based on memristor array.
Recently, the related research result of this work is entitled "Brain-inspired sound localization for in-situ training based on memristors" in Nature? Newsletter published online (natural newsletter). Associate Professor Gao Bin and Professor Wu Huaqiang of Tsinghua University Institute of Integrated Circuits are co-authors of this paper, and Ying Zhou, Gao Bin and Jasmine Zhang Tian are co-authors of this paper. This research is supported by the major scientific and technological innovation 2030 project "Brain Science and Brain-like Research", Jieqing Project of National Natural Science Foundation, key projects and scientific exploration award.
Associate Professor Gao Bin has been devoted to the research of brain like computing based on memristors for many years. He was invited to write a summary of brain chips in journals such as Nature Electronics, Nature Communications and Proceedings of IEEE, give a special report at conferences such as ASP-DAC and A-SSCC, and serve as the chairman of model sub-committee at flagship conferences such as IEDM and EDTM.