| In the era of intelligent computing,various artificial intelligence technologies have penetrated into all aspects of people’s life and affect people’s daily activities.Deep learn-ing technology based on neural network has achieved fruitful results and produced many applications,including molecular structure prediction,target detection,automatic driving,speech recognition and so on.In addition,the biological interpretability of this neuromor-phic algorithm is becoming more and more resonable,and the performance is much closer to human itself.Thus,it is considered to be one of the important ways to realize strong ar-tificial intelligence in the future.However,as an algorithm model derived from biological nervous system which operating information in analog domain,there are many problems in using binary digital computer for indirect implementation.Operations such as data format transformation,data structuring and data movement on the computing center pro-duce large computational resource overhead and time delay.These are not conducive to its further development under the trend of multi-sensor fusion in the future.In view of the above problems,this paper systematically studied and compared the similarities and differences between biological neural system and existing computing system in the imple-mentation of neural network algorithm,Further,we analyzed the information processing flow of these two system architectures and designed a new neural morphological comput-ing architecture based on the relationship between biological peripheral nervous system and central nervous system.Starting from basic computing architecture,this paper constructs an optoelectronic hybrid computing system integrating neuromorphic in-sensor computing from top to down.Firstly,in the top level design,the computational tasks in convolutional neural network algorithm are divided into modules and assigned to system software and hardware seper-ately?Then,the data interface between software and hardware is established as well as the mapping strategy between measured physical quantity and network parameters?Finally,in the bottom implementation,a neuron device which can simulate the function of optic neurons is prepared.The main research contents of this paper are summarized as follows:1.In terms of computing architecture,various mainstream neuromorphic algorithms in the current computing vision framework are systematically studied as well as their prob-lems and main bottlenecks.Aiming at the existing problems,following the processing mode of biological neural system for visual information,the algorithm tasks and cooper-ation strategies inside the computing system are reset and divided,constructiong a new neuromorphic vision framework.This model possesses better biological interpretability,greatly reduces the time and energy cost caused by information digitization,data structure and data movement.2.In the aspect of neuromorphic computing unit,a three-dimensional nano structure of optical synapse based on transition metal disulfides（TMDCs）is designed and real-ized.Based on the physiological activities of biological retina neurons and optic ganglia,the main functions of optical synapse are analyzed and refined.In the selection of pho-tosensitive layer materials,the photoelectric properties and carrier dynamics of TMDCs materials are characterized by steady-state and ultrafast spectroscopy?In terms of device structure fabricating,with the help of modern nanotechnology and coating technology,independent optical synaptic devices are prepared layer by layer.Test result shows that when a constant gate voltage is applied,the optical synapse has a ultra low dark current(～10-13A)and possesses the ability of producing a linear light response（up to 3 orders of magnitude）.Under the same light intensity,the embedded back gate structure can effec-tively manipulate the photocurrent of the device so as to achieve a reconfigurable synaptic effect.3.In the realization of neuromorphic algorithm function,algorithm modules are real-ized by system software and hardware seperately.Aiming at the matrix multiplication and matrix convolution calculation functions required by neuromorphic vision computing,a corresponding synaptic spatial topology is designed on the independent optical synapse to form optical-sensoring neuron unit（ONU）.A variety of equivalent matrix calculations are realized by time-division multiplexing ONU chip?In the aspect of software and hardware cooperation,the interface coupling method between chip output and logic neural network is studied.Network parameters are mapped to the actual physical quantities,realizing the docking and processing of system software input and hardware output data.4.The system framework of in-sensor computing is built and verified by experi-ments.The spatial modulation of beam and efficient loading of information are realized by using the self-built beam shaping system and digital micro mirror（DMD）system?With the self built micro area imaging system,a high-precision optical alignment system is realized?The observation results of micro imaging after carrier modulation are highly consistent with the simulation results?The signal detection system with very low noise is also realized?A 3×3 array size optical synapse array is prepared to form an ONU chip,and the synapse pre-configuration based on simulation parameters and the extraction of ONU computing results are carried out.Test results show that our system can effectively complete the calculation process of convolution layer and full connection layer,proving the effectiveness of the integrated neuromorphic in-sensor computing system.The main innovations of this paper include:Imitating the processing process of bi-ological neural network for external information,combined with the algorithm model,a more efficient sensory computing integrated neuromorphic computing architecture is designed?The basic calculation elements of neural morphology computing are analyzed,and the computing process is accelerated by using the physical properties of new mate-rials and devices?With the help of the transmission characteristics of optical carrier,the in-sensor computing method is further accelerated and the corresponding algorithm model is improved?A neuromorphic in-sensor computing system is constructed,and the entire algorithm function is realized through the coupling of software and hardware. |
PDF Full Text Request