Analog domain time signal photon peak processing method and system architecture based on synaptic delay plasticity

Abstract

The invention discloses an analog domain time signal photon peak processing method and system architecture based on synaptic delay plasticity. The system architecture comprises an analog domain time signal input, a detector array, a photon neuron peak coding layer, a synaptic delay adjusting module, a photon neuron peak output layer, an optical decision module and a multifunctional output module. According to the method, the real-time weight-free photon spike processing of the simulation domain time signal can be realized by combining the advantages of biological nerve mimic synaptic parameter diversity and all-optical high-speed processing, combining the characteristics of photon neuron threshold judgment and time domain integration and a pulse neural network algorithm. The method is of great significance in improving real-time processing of simulation domain time signals such as flight time signals and promoting application of photon peak processing in low-delay scenes.

Description

Technical Field

[0001] The present invention relates to the field of photon information processing, and in particular to a method and architecture for processing photon spikes of analog domain time signals based on synaptic delay plasticity. Background Art

[0002] Analog domain time signal processing has high requirements for low latency and low power consumption in future scenarios such as unmanned driving and virtual reality. At present, traditional processing of time signals such as time-of-flight (ToF) signals requires the use of time-to-digital converters (TDC) to convert them into the digital domain for precise time calculation, and then perform neural network calculations or other complex algorithm operations in the digital domain. Among them, the conversion of TDC to the digital domain alone requires a large delay and power consumption. Algorithms such as neural networks are processed in the electrical domain based on the von Neumann architecture, and also face the millisecond-level delay bottleneck caused by the data transmission between the central processing unit (CPU) and the memory. The overall processing efficiency is low and cannot meet the growing real-time requirements of electronic information systems for signal processing.

[0003] Biological brains can perform high-speed parallel information processing with ultra-low power consumption and strong fault tolerance. Photonic technology has advantages over electrical processing such as high speed and large bandwidth. Photonic neuromorphic, which organically integrates biological neuron and synaptic characteristics with photonic technology, can provide a disruptive solution for real-time processing of time signals in the future. Therefore, conducting research on the principles and architecture design of photonic neuromorphic processing of time signals is of great significance to improving the performance of information processing systems.

[0004] At present, simulating the spike processing mechanism of the human brain to improve signal processing efficiency has become a research hotspot for the new generation of information technology. In 2013, the European Union proposed the "Human Brain Project (HBP)" to develop cutting-edge information technology in the fields of future neuroscience and future computing. European scientists have tried to use a spike timing coding-based method to process tactile information. By using the characteristics of sparse coding of spike signals, the frame coding delay of electronic skin is reduced from more than 100 milliseconds to a few milliseconds, which is comparable to the biological nervous system (Chen L, et al. Spike timing-based coding in neuromimetic tactile system enables dynamic object classification. Science, 2024, 384(6696): 660-665.). In addition, in terms of hardware, the University of Münster in Germany and others have developed a photon spike processing chip that integrates all-optical synapses and all-optical neurons. Benefiting from the inherent high bandwidth and fast data transmission rate of light, the processing delay of tasks such as pattern classification of digital signals is reduced to the microsecond level (Feldmann J, et al. All-optical spiking neurosynaptic networks with self-learning capabilities. Nature, 2019, 569 (7755): 208-214.). Xidian University in China has conducted in-depth research on the algorithm optimization of pulse neural networks, and cooperated with photonic neuromorphic unit devices to achieve efficient static image processing with soft and hard coordination. (Han Y, et al. y Experimental demonstration of delay-weight learning and pattern classification with a FP-SA-based photonic spiking neuron chip [J]. Journal of Lightwave Technology, 2023.)

[0005] Therefore, conducting research on the principles and architectural design of photonic neuromorphic processing of temporal signals has great scientific significance and application value for breaking through the performance bottleneck of current information processing systems and meeting the stringent requirements of future electronic information systems for real-time and low power consumption. Summary of the invention

[0006] The present invention is inspired by the biological delay plasticity mechanism. The stimulation signal in the brain will have a delay effect after being transmitted through the synaptic structure. The specific delay amount can be adjusted by the thickness of the myelin sheath and axon. In the thalamus system, auditory brainstem system and other nervous systems, the synaptic delay plasticity can be used to adjust the peak moment to complete signal processing, and realize functions such as sound positioning, direction selection, and visual perception. This technology uses synaptic delay plasticity to adjust the photon spike moment obtained by analog domain time signal encoding, and extracts timing features to complete tasks such as distance perception and three-dimensional object classification. It is a new technology based on photon technology, biology, and brain-like intelligence. This technology uses the advantages of high-speed transmission of light and the diversity of parameters of biological synapses, combined with pulse neural network algorithms, etc., to form a real-time analog domain time signal spike processing system, which effectively alleviates the pressure of high-speed signal processing in scenarios such as intelligent driving and virtual reality.

[0007] The purpose of the present invention is to propose a method and system architecture for processing photon spikes of analog domain time signals based on synaptic delay plasticity. The method combines the advantages of bio-neuromorphic synaptic parameter diversity and all-optical high-speed processing, and utilizes the characteristics of photon neuron threshold judgment and time domain integration, as well as the pulse neural network algorithm, to achieve real-time weightless photon spike processing of analog domain time signals.

[0008] The technical solution of the present invention is as follows:

[0009] The analog domain time signal photon spike processing method and system architecture based on synaptic delay plasticity is characterized by completing the real-time coding learning and judgment of the time signal in the pure analog domain without weight control and digital domain conversion, including analog domain time signal input, detector array, all-optical spike processing module, optical decision module and multi-functional output module. The all-optical spike processing module includes the photon neuron spike coding layer, the synaptic delay adjustment module and the photon neuron spike output layer.

[0010] The analog domain time signal input module is used to receive time signals in multiple scenarios. The input signal is N-channel time signals. The detector array is used to convert optical analog signals to electrical analog signals. It is composed of N detector units and can detect N-channel analog domain time signals respectively. The converted electrical analog signal is loaded into the photon neuron spike coding layer. The photon neuron spike coding layer includes N photon neuron units, which correspond to the detectors one by one. Each photon neuron unit uses the time domain integration characteristics to encode the detected time signal into photon spikes at different times. The photon neuron spike output layer includes M photon neuron units, which are fully connected to the N photon neurons in the photon neuron spike coding layer and have N*M branches. A synaptic unit is set on each branch to adjust the delay of spike transmission. All N*M synaptic delay adjustment units constitute a synaptic delay adjustment module. The optical decision module is composed of P photon neurons, which are connected to the M photon neuron units of the photon neuron spike output layer in a fully connected manner. The outputs of the M photon neurons in the photon neuron spike output layer are combined and input into each photon neuron of the optical decision module respectively. The photon neurons generate spike responses at different times through time domain integration and threshold judgment operations. Under different types of time signal inputs, the multi-functional output module completes multi-functional outputs such as target recognition, object classification, and coincidence detection according to the spike response time.

[0011] The implementation platform can be a system composed of discrete devices, an integrated chip platform, or a system that uses a mixture of discrete devices and integrated chips. When using discrete devices for implementation, the connection between modules and components can be by means of limiting the propagation path of the light beam, such as optical fiber; when using integrated chips for implementation, the connection between components can be by means of integrated waveguides, etc.

[0012] The analog domain time signal input adopts but is not limited to a time of flight signal (ToF).

[0013] The detector array can be realized by methods such as, but not limited to, single photon avalanche diode (SPAD), avalanche photodiode (APD), etc.

[0014] The multi-channel analog signals obtained by the detector array are loaded into the photon neuron spike coding layer, which can be directly modulated or externally modulated. Direct modulation can use a current source or a voltage source to change the bias current of the laser. External modulation can use methods including but not limited to amplitude modulation (AM), phase modulation (PM), etc., and devices such as silicon-based Mach-Zehnder intensity modulators and lithium niobate Mach-Zehnder intensity modulators can be used.

[0015] The photon neurons in the photon neuron spike coding layer, the photon neuron spike output layer and the optical decision module can adopt but are not limited to photon neurons based on the injection locking architecture. Specifically, the injection locking architecture injects the output of a laser (master laser) into another isolated laser (slave laser) through an optical isolator or a circulator, uses an adjustable optical attenuator or an optical amplifier to adjust the intensity of the injected light, and a polarization controller to adjust the polarization of the injected light, wherein the laser can adopt but is not limited to a Bragg semiconductor laser (DFB-LD) and a vertical cavity surface emitting laser (VCSEL). In addition, photon neurons based on the self-pulsation state of microrings, two-stage lasers containing saturated absorbers, photon neurons based on polarization competition of vertical cavity surface emitting lasers, and photon neurons based on phase change materials (PCM) embedded in microring resonators can also be used.

[0016] The synaptic delay adjustment module may adopt, but is not limited to, optical fiber delay line, optical waveguide delay line, slow light delay line or spatial free propagation delay and other components that achieve delay effects through the propagation time of light.

[0017] The encoding and feature extraction principles of the all-optical spike processing module are as follows.

[0018] Photon neurons have time domain integration characteristics. Let the time interval for the photon neuron to integrate the input signal be t i , the threshold is T, the response time is t j The initial moment when the detector receives the time signal is recorded as t0. i Within the interval, the time signal has an integral effect on the concentration of photogenerated carriers and light field intensity in the photon neuron, that is, the influence of the latter moment will be superimposed on the previous moment. When the change of photogenerated carriers or light field intensity exceeds the threshold T, a photon spike will be stimulated at the output end. The moment of the photon spike is t j It is related to the moment when the threshold is exceeded, and more fundamentally to factors such as the input time signal amplitude, duration, time interval, etc. The photon neurons in the photon neuron spike coding layer can encode signals of different durations, amplitudes, and time intervals into spike sequences at different times.

[0019] The learning of delay parameters in the synaptic delay adjustment module is based on the Hebbian rule: if a neuron A repeatedly and continuously excites neuron B, the synaptic connection between A and B will be strengthened (Feldman D E. The spike-timing dependence of plasticity [J]. Neuron, 2012, 75 (4): 556-571.). According to the analysis of the above literature, the time difference between the target and the actual output peak is used to update the synaptic parameters according to the negative exponential function. The specific rules for updating the synaptic delay parameters adopted in this patent are as follows:

[0020]

[0021] Where Δd is the change in the delay parameter for each input sample, Δt is the difference between the ideal output time of the peak and the actual output time, Δt = t o -t d , t o Refers to the actual output time after the input signal passes through the synapse, t d It is the ideal output time set according to the supervisory signal, and m refers to the maximum value of the delay parameter for a single update. When the actual peak output is later than the ideal output time, Δt>0, it is necessary to reduce the delay in the transmission process and adjust the delay parameter of the synapse to a smaller value. On the contrary, when the actual peak output is earlier than the ideal output time, Δt<0, it is necessary to increase the delay in the transmission process and adjust the delay parameter of the synapse to a larger value.

[0022] Compared with the prior art, the present invention has the following advantages:

[0023] 1. Real-time: The present invention realizes photon spike processing of analog domain time signals based on photonic technology and brain-like intelligence, giving full play to the advantages of efficient and real-time analog signal processing, effectively improving the efficiency of time signal processing and improving the performance of information processing systems.

[0024] 2. Intelligence: The present invention constructs an intelligent delay learning network based on the pulse neural network algorithm to realize signal feature extraction and delay parameter learning and training, thereby being able to effectively cope with the recognition and analysis of information in various scenarios.

[0025] 3. Energy saving: The present invention constructs a synaptic delay adjustment module based on the principle of delay plasticity, avoiding the power consumption caused by additional weight adjustment devices, thereby reducing the energy consumption of photon spike processing.

[0026] 4. Taking advantage of the ability of photon neuromorphism to directly process analog signals, the original time signal is encoded into spike signals at different times using the time-domain integration characteristics of photon neurons, without the need to use a time-to-digital converter to convert it to the digital domain for processing. Then, the pure delay parameter learning algorithm of the pulse neural network inspired by synaptic delay plasticity is used to train and learn based on the characteristics of the spike moment, and the delay of the photon synapse is adjusted according to the trained delay parameters, completing the flight time signal processing in the three-dimensional object classification with ultra-low delay. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 It is a schematic diagram of the flow of the method for processing photon peaks of time signals of the present invention;

[0028] Figure 2 Schematic diagram of an embodiment of a photonic neuron unit and a synaptic delay adjustment unit based on an injection locking architecture;

[0029] Figure 3 An example diagram of an implementation of a classification task for analog domain time-of-flight signal processing of a three-dimensional object. DETAILED DESCRIPTION

[0030] The technical solution of the present invention is described in detail below in conjunction with the accompanying drawings and embodiments, and detailed implementation methods and structures are given, but the protection scope of the present invention is not limited to the following embodiments.

[0031] See also Figure 1 , Figure 1 The figure is a flow chart of the method for processing photon spikes of analog domain time signals based on synaptic delayed plasticity of the present invention. Figure 1 It can be seen that the analog domain time signal photon spike processing method based on synaptic delay plasticity includes:

[0032] Analog domain time signal input 1, receiving time signals from different scenes;

[0033] The detector array 2 converts the received optical analog signal into an electrical analog signal;

[0034] All-optical spike processing module 3:

[0035] - Photon neuron spike coding layer 3-1: encodes electrical analog signals into photon spikes at different times.

[0036] -Synaptic delay adjustment module 3-2: adjusts the delay of photon spikes during transmission.

[0037] - Photon neuron spike output layer 3-3: outputs the adjusted photon spike signal and feeds it back to the photon neuron spike coding layer 3-1 for iterative optimization.

[0038] Optical decision module 4: performs combining and thresholding on the spike signals output by the photon neuron spike output layer 3-3 to generate a spike response;

[0039] Multifunctional output module 5: According to the peak response time, it completes multifunctional outputs such as target recognition, object classification, and coincidence detection.

[0040] The analog domain time signal input 1 is connected to the input end of the detector array 2, the output end of the detector array 2 is connected to the first input end of the photon neuron spike coding layer 3-1, the output end of the photon neuron spike coding layer 3-1 is connected to the synaptic delay adjustment module 3-2, and the synaptic delay adjustment module 3-2 adjusts the moment of the photon spike in the link and then inputs it into the photon neuron spike output layer 3-3. The photon neuron spike coding layer 3-1, the synaptic delay adjustment module 3-2 and the photon neuron spike output layer 3-3 together constitute the all-optical spike processing module 3, the photon neuron spike output layer 3-3 is connected to the second input end of the photon neuron spike coding layer 3-1, and the output result after the photon spike processing is fed back to the photon spike coding layer, so as to achieve parameter optimization and improvement of processing effect in repeated iterations. The output end of the all-optical spike processing module 3 , that is, the output end of the photon neuron spike output layer 3 - 3 , is connected to the optical decision module 4 , and the output end of the optical decision module 4 is connected to the input end of the multifunctional output module 5 .

[0041] Figure 2 The specific implementation of the photon neuron unit based on the injection locking architecture and the synaptic delay adjustment unit based on the optical delay line are shown. The photon neuron spike coding layer 3-1, the photon neuron spike output layer 3-3 and the optical decision module 4 all contain multiple photon neuron units. The photon neuron unit includes a master laser 100, a modulator 200, an adjustable attenuator 300, a circulator 400, a polarization controller 500, and a slave laser 600. The output end of the master laser 100 is connected to the first input end of the modulator 200, the output end of the modulator 200 is connected to the input end of the adjustable attenuator 300, and then connected to the first input end of the circulator 400. The transmission direction of the optical signal in the circulator 400 is counterclockwise. The second input end of the circulator 400 is connected to the input end of the polarization controller 500, and the output end of the polarization controller 500 is connected to the port of the slave laser 600. The original output light of the slave laser interacts with the injected light field and is merged into an optical signal. After passing through the polarization controller 500, it is input to the second input end of the circulator 400, and then transmitted to the output port of the circulator 400.

[0042] When the photon neuron unit is in the photon neuron spike coding layer 3-1, the input signal is the electrical signal output by the detector array 2, the modulator 200 is an electro-optic modulator, and the output of the detector unit is loaded to the RF input port of the electro-optic modulator by a cable. The output end of the photon neuron unit is a photon spike signal, which completes the spike coding of the analog domain time signal detected by the detector, and is subsequently input into the synaptic delay adjustment unit 3-2. The synaptic delay adjustment unit is composed of an optical delay line 700, which is used to adjust the delay of the photon spike during the transmission process. The output end of the optical delay line 700 is connected to the input end of the photon neuron spike output layer 3-3.

[0043] When the photon neuron unit is in the photon neuron spike output layer 3-3, the input signal is the output signal of the synaptic delay adjustment module 3-2, which can be an optical signal. The modulator 200 uses an all-optical modulator to directly load the photon spike sequence after synaptic delay adjustment on the second input end of the all-optical modulator, or the signal can be converted into an electrical signal using a photoelectric detector. The modulator 200 uses an electro-optical modulator to load the converted electrical signal to the RF input port of the electro-optical modulator using a cable. When the photon neuron is in the optical decision module 4, the input signal is the output signal of the photon neuron spike output layer 3-3, which can be an optical signal. The modulator 200 uses an all-optical modulator to directly load the photon spike sequence after all-optical spike processing on the second input end of the all-optical modulator, or the signal can be converted into an electrical signal using a photoelectric detector. The modulator 200 uses an electro-optical modulator to load the converted electrical signal to the RF input port of the electro-optical modulator using a cable.

[0044] The analog domain time signal photon spike processing method based on synaptic delay plasticity can be used to perform tasks including but not limited to three-dimensional object classification and target recognition. Figure 3The system architecture diagram for implementing the classification task of analog domain flight time signal processing for three-dimensional objects. Here, the ModelNet40 dataset is taken as an example. The depth information of real objects such as cones is detected by flight time ranging to obtain analog domain time signal input 1. Here, it is assumed that the N detection points of the object are simultaneously detected by flight time to obtain N time signal inputs. N detectors are set in the detector array 2, corresponding to the N time signal inputs. Then the output end of each detector unit is connected to a photon neuron unit to complete the spike coding of the N time signals, so the photon neuron spike coding layer 3-1 has N photon neuron units. The photon neuron spike output layer 3-3 has M photon neuron units, and a photon spike network with N*M branches is formed between the two layers in a fully connected manner. Each photon neuron in the coding layer 3-1 has M output ports, and each port is connected to a synaptic delay adjustment unit. Different synaptic delay adjustment units have different adjustment amounts for the spike moment. The output port of each synaptic delay adjustment unit is connected to the photon neuron unit in the photon neuron spike output layer 3-3. There are a total of N*M synaptic delay adjustment units, which together constitute the synaptic delay adjustment module 3-2. Each time a sample in the ModelNet40 data set is input, the output of the photon neuron spike output layer 3-3 is fed back to the photon neuron spike coding layer 3-1 for iterative optimization. At the same time, the M photon neurons in the photon neuron spike output layer 3-3 are connected to the P photon neurons in the optical decision module 4 in a fully connected manner, and three-dimensional object classification is performed according to the moment when the photon spike is output in the optical decision module 4. Figure 3 It is judged as a "cone" and the processing result is output by the multi-functional output module 5.

Claims

1. An analog domain time signal photon spike processing system architecture based on synaptic delay plasticity, characterized in that: include: The analog domain time signal input module is used to receive analog domain time signals in multiple scenarios, that is, N analog domain time signals; The detector array is composed of N detector units and is used to convert the optical analog signal into an electrical analog signal and respectively detect N analog domain time signals; An all-optical spike processing module, comprising a photon neuron spike coding layer, a synaptic delay adjustment module and a photon neuron spike output layer, for completing the coding, delay adjustment and output of time signals in a pure analog domain; wherein the photon neuron spike coding layer has N photon neuron units, corresponding one-to-one to the detector units, and each photon neuron unit uses the time domain integration characteristic to encode the detected time signal into photon spikes at different times; the synaptic delay adjustment module has N*M synaptic delay adjustment units, which are arranged on each branch between the photon neuron spike coding layer and the photon neuron spike output layer, for adjusting the delay of spike transmission; the photon neuron spike output layer has M photon neuron units, which are fully connected to the N photon neurons of the photon neuron spike coding layer, and output the adjusted spike signal; The optical decision module is composed of P photon neurons, which are connected to the M photon neuron units of the photon neuron spike output layer in a fully connected manner. The outputs of the M photon neurons are combined and input into each photon neuron respectively. Each photon neuron generates a spike response at different times through time domain integration and threshold judgment operations. The multi-function output module completes target recognition, object classification, and coincidence detection multi-function output according to the peak response time.

2. The analog domain time signal photon spike system architecture based on synaptic delay plasticity according to claim 1, characterized in that: The output end of the analog domain time signal input 1 is connected to the input end of the detector array 2, the output end of the detector array 2 is connected to the first input end of the photon neuron spike coding layer 3-1, the output end of the photon neuron spike coding layer 3-1 is connected to the synaptic delay adjustment module 3-2, and after the synaptic delay adjustment module 3-2 adjusts the moment of the photon spike in the link, it is input to the photon neuron spike output layer 3-3. The photon neuron spike output layer 3-3 is connected to the second input end of the photon neuron spike coding layer 3-1, and the output result after the photon spike processing is fed back to the photon spike coding layer, so as to achieve parameter optimization and improvement of processing effect in repeated iterations. The output end of the all-optical spike processing module 3, that is, the output end of the photon neuron spike output layer 3-3, is connected to the optical decision module 4, and the output end of the optical decision module 4 is connected to the input end of the multi-function output module 5.

3. The analog domain time signal photon spike system architecture based on synaptic delay plasticity according to claim 1, characterized in that: The photon neuron unit includes a master laser 100, a modulator 200, an adjustable attenuator 300, a circulator 400, a polarization controller 500 and a slave laser 600; the output end of the master laser 100 is connected to the first input end of the modulator 200, the output end of the modulator 200 is connected to the input end of the adjustable attenuator 300, and then connected to the first input end of the circulator 400, the transmission direction of the optical signal in the circulator 400 is counterclockwise, the second input end of the circulator 400 is connected to the input end of the polarization controller 500, the output end of the polarization controller 500 is connected to the port of the slave laser 600, the original output light of the slave laser interacts with the injected light field and is merged into one optical signal, which is input to the second input end of the circulator 400 after passing through the polarization controller 500, and then transmitted to the output port of the circulator 400.

4. The analog domain time signal photon spike processing system architecture based on synaptic delay plasticity according to any one of claims 1 to 3, characterized in that: The implementation platform can be a system composed of discrete devices, an integrated chip platform, or a system that uses a mixture of discrete devices and integrated chips. When discrete devices are used for implementation, the modules and components can be connected by optical fibers or other methods that limit the propagation path of the light beam. When integrated chips are used for implementation, the components can be connected by integrated waveguides or other methods.

5. The analog domain time signal photon spike processing system architecture based on synaptic delay plasticity according to any one of claims 1 to 3, characterized in that: The detector array 2 is implemented by, but not limited to, single photon avalanche diode (SPAD), avalanche photodiode (APD) and other methods.

6. The analog domain time signal photon spike processing system architecture based on synaptic delay plasticity according to claim 5, characterized in that: The multi-channel analog signals obtained by the detector array are loaded into the photon neuron spike coding layer 3-1, which can be directly modulated or externally modulated; direct modulation can use a current source or a voltage source to change the bias current of the laser; external modulation can use methods including but not limited to amplitude modulation (AM), phase modulation (PM), etc., and devices such as silicon-based Mach-Zehnder intensity modulators and lithium niobate Mach-Zehnder intensity modulators can be used.

7. The analog domain time signal photon spike processing system architecture based on synaptic delay plasticity according to any one of claims 1 to 3, characterized in that: The photon neurons in the photon neuron spike coding layer, the photon neuron spike output layer and the optical decision module adopt photon neurons based on injection locking architecture, photon neurons based on microring self-pulsation state, two-stage lasers containing saturated absorbers, photon neurons based on vertical cavity surface emitting laser polarization competition, and photon neurons based on phase change material (PCM) embedded in microring resonators.

8. The analog domain time signal photon spike processing system architecture based on synaptic delay plasticity according to claim 7, characterized in that: The injection locking architecture injects the output of a laser (master laser) into another isolated laser (slave laser) through an optical isolator or circulator, uses an adjustable optical attenuator or optical amplifier to adjust the intensity of the injected light, and a polarization controller to adjust the polarization of the injected light. The laser can be, but is not limited to, a Bragg semiconductor laser (DFB-LD) or a vertical cavity surface emitting laser (VCSEL).

9. The analog domain time signal photon spike processing system architecture based on synaptic delay plasticity according to any one of claims 1 to 3, characterized in that: The synaptic delay adjustment module 3-2 may adopt but is not limited to optical fiber delay line, optical waveguide delay line, slow light delay line or spatial free propagation delay and other components that achieve delay effect through the propagation time of light.

10. The analog domain time signal photon spike processing system architecture based on synaptic delay plasticity according to any one of claims 1-3, characterized in that: The encoding and feature extraction principle of the all-optical spike processing module is based on the time-domain integration characteristics of photon neurons, and encodes signals of different durations, different amplitudes, and different time intervals into spike sequences at different times.

11. The analog domain time signal photon spike processing system architecture based on synaptic delay plasticity according to any one of claims 1 to 3, characterized in that: The learning of delay parameters in the synaptic delay adjustment module is based on the Hebbian rule. The time difference between the target and the actual output peak is used to update the synaptic parameters according to the negative exponential function. The specific rules are: Where Δd is the change in the delay parameter per sample input, Δt is the difference between the ideal output time of the peak and the actual output time, Δt = t o -t d , where t o Refers to the actual output time after the input signal passes through the synapse, t d It is the ideal output time set according to the supervision signal, and m refers to the maximum value of the delay parameter for a single update; when the actual peak output is later than the ideal output time, the delay in the transmission process is reduced; when the actual peak output is earlier than the ideal output time, the delay in the transmission process is increased.

12. A method for processing photon spikes of analog domain time signals based on synaptic delay plasticity, characterized in that: The steps include: a. Receive analog domain time signal input; b. converting the optical analog signal into an electrical analog signal through a detector array; c. Encoding electrical analog signals into photon spikes at different times using a photon neuron spike coding layer; d. regulating the delay of spike transmission through the synaptic delay regulation module; e. outputting the regulated spike signal using the photon neuron spike output layer; f. Combine and threshold the peak signal through the optical decision module to generate a peak response; g. According to the peak response time, the multi-functional output module can complete multi-functional outputs such as target recognition, object classification, and coincidence detection.

Images