Neuromorphic learning method and system of liquid state machine
Through the liquid state machine neuromorphic learning method, a neuromorphic learning system with multiple topological structures is constructed, which solves the problem of single neural network topology structure and high cost of simulated neuronal models in the existing system, and achieves efficient neuromorphic learning.
Patent Information
- Application Number
- CN202411330519.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-24
- Publication Date
- 2024-12-17
AI Technical Summary
The topology of neural networks in existing neuromorphic learning systems is single, and the cost of simulating neuronal models is high, which limits the system's scale promotion and work efficiency.
Using the liquid state machine neuromorphic learning method, by constructing a quadratic pulse neuron model and designing adapted neuron digital circuits in real time, we judge whether the neuron digital circuit can emit pulses based on preset rules, and build an online monitoring learning algorithm. Set the optimal learning rate and backpropagation rate to achieve neuromorphic learning.
The nonlinear quadratic pulse neuron model simulates neuronal discharge behavior, realizes bidirectional regulation of synaptic plasticity, and builds a learning system with multiple topological structures, which improves work efficiency and flexibility of the learning system.
Smart Images
Figure CN119150937A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of artificial intelligence computing, and in particular to a liquid state machine neuromorphic learning method and system. Background Art
[0002] With the development of the times and the advancement of science and technology, artificial intelligence technology has also developed rapidly. Among them, computational neuroscience is an important component of artificial intelligence, which is mainly used to study the principles, structure and function of artificial neural networks.
[0003] Among them, the existing digital implementations of neuromorphic computing mainly include Field Programmable Gate Array (FPGA) and Application Specific Integrated Circuit (ASIC) chips. Furthermore, compared with the existing analog implementation and ASIC implementation, FPGA can be designed and developed in a short time, and has better interaction capabilities with the host computer during optimization, so it has been widely used.
[0004] However, in the process of constructing a neuromorphic learning system with the existing technology, the topological structure of the constructed neural network is relatively simple, and the cost of simulating the neuron model is high, which is not conducive to the large-scale promotion and use of the neuromorphic learning system and reduces work efficiency. Summary of the invention
[0005] Based on this, the purpose of the present invention is to provide a liquid state machine neuromorphic learning method and system to solve the problem that the topological structure of the neural network in the neuromorphic learning system constructed by the prior art is relatively simple and the cost of simulating the neuron model is relatively high.
[0006] The first aspect of the embodiment of the present invention proposes: A liquid state machine neuromorphic learning method, wherein the method comprises: Constructing a quadratic pulse neuron model, and designing a neuron digital circuit adapted to the quadratic pulse neuron model in real time; Based on preset rules, it is determined in real time whether the neuron digital circuit in the quadratic pulse neuron model can emit pulses under specific conditions; If it is determined in real time based on the preset rules that the neuron digital circuit in the quadratic pulse neuron model can emit pulses under specific conditions, the liquid layer neurons and the output layer neurons respectively contained in the preset liquid state machine model are detected in real time, and an online monitoring learning algorithm is constructed based on the liquid layer neurons and the output layer neurons; An optimal learning rate and a back propagation rate corresponding to the online monitoring learning algorithm are set, and based on a preset environment, the neuromorphic learning of the liquid state machine is completed according to the optimal learning rate, the back propagation rate and the online monitoring learning algorithm.
[0007] The beneficial effects of the present invention are as follows: the circuit design and implementation method of the nonlinear quadratic pulse neuron model can simulate the neuron discharge behavior in software and hardware. Furthermore, the synaptic plasticity mechanism based on bidirectional regulation can construct the required online supervised learning algorithm and realize the real-time adjustment of the synaptic weights of all layers, thereby constructing the required learning system, which has a variety of topological structures and improves work efficiency.
[0008] Furthermore, the step of designing a neuron digital circuit adapted to the quadratic pulse neuron model in real time includes: Calculating in real time through a first preset algorithm a parameter calculation pipeline corresponding to the quadratic spiking neuron model; Detecting in real time the parameter operation sequence contained in the parameter calculation pipeline, and designing a neuron digital circuit architecture adapted to the quadratic pulse neuron model according to the parameter operation sequence; The neuron digital circuit is generated according to the parameter calculation pipeline and the neuron digital circuit architecture.
[0009] Furthermore, the expression of the first preset algorithm is:
[0010] Among them, the simulation of digital circuits involves the solution of differential equations. In order to solve the second-order derivative differential equation in digital form, the Euler broken line method is used here to discretize the time step:
[0011] Among them, the duration is the time step, so The state variables at the moment are used to calculate The state variables at the moment, and in this way, the discretized equations suitable for circuit implementation are obtained:
[0012] The parameter calculation pipeline is calculated accordingly, wherein: Represents the real-time changing membrane potential of neurons; represent Current activated or Recovery variables for current inactivation; represents the membrane capacitance; is the membrane potential of the neuron when it is in the resting state; represents the instantaneous threshold potential, when At the instant of , the depolarization of the neuron will lead to the firing of a pulse; represents the excitatory input current; , , , is a dimensionless parameter.
[0013] Furthermore, the step of constructing an online monitoring learning algorithm based on the liquid layer neurons and the output layer neurons includes: Discretely represent the pulse sequences corresponding to the liquid layer neurons and the output layer neurons, wherein when encountering a presynaptic pulse, the decrease in the weight between the liquid layer neurons and the output layer neurons is proportional to the instantaneous value of the presynaptic simulation variable; When encountering a postsynaptic pulse, the increase in the weight between the liquid layer neuron and the output layer neuron is proportional to the postsynaptic simulation variable; A learning rule corresponding to the synaptic weight is formulated in real time, and the online detection learning algorithm is constructed based on the learning rule.
[0014] Furthermore, the step of constructing the online detection learning algorithm based on the learning rule includes: The change of synaptic weight is jointly affected by the pulse sequence of the liquid layer neurons and the output layer neurons; The STDP weight adjustment rule for synaptic weights online can be defined in combination with the Widrow-Hoff rule:
[0015] in represents the learning rate of the algorithm, represents the adjustment rule of synaptic weights, represents the pulse train of the liquid layer neuron, represents the pulse sequence of the output layer neurons, Represents the original pulse sequence; by decomposing the above formula into a time-dependent STDP learning process and an anti-STDP learning process, the formula can be further derived as follows:
[0016] In summary, and It can be described as:
[0017] in , , They correspond to the analog signals converted from the liquid layer, the output layer, and the expected pulse sequence, respectively, and δ represents the conversion coefficient; therefore, the learning rule of the synaptic weights of the output layer is expressed as:
[0018] By summing up all the synaptic weight changes of LTP and LTD in the output layer, we can get the excitatory synaptic weights of the liquid layer and the input layer. and Adjustment rules:
[0019] in represents the back propagation rate, represents the number of neurons in the output layer; In contrast, the weight change caused by LTD Neurons back-propagated to the fluid layer and input layer and At this time and are all positive values; similarly, the inhibitory synaptic weights of the liquid layer and the input layer can be obtained and Adjustment rules:
[0020] The online monitoring learning algorithm is constructed accordingly.
[0021] Furthermore, the step of completing the neuromorphic learning of the liquid state machine based on the preset environment according to the optimal learning rate, the back propagation rate and the online monitoring learning algorithm includes: Build the overall hardware architecture of the learning system and design an input pulse processing module corresponding to the overall hardware architecture; Designing a pulse sequence transmission strategy corresponding to the input pulse processing module, and designing a corresponding pulse sequence receiving circuit and a synaptic weight learning module circuit; A Moore state machine adapted to the salient weight learning module circuit is designed, and the neuromorphic learning is completed based on the Moore state machine.
[0022] Furthermore, the step of completing the neuromorphic learning of the liquid state machine based on the preset environment according to the optimal learning rate, the back propagation rate and the online monitoring learning algorithm also includes: Collecting the learning results of the neuromorphic learning in real time, and comparing the learning results with the classification results of the preset SNN hardware system; Whether the neural morphological learning of the liquid state machine is completed is determined in real time according to the comparison result.
[0023] The second aspect of the embodiment of the present invention proposes: A liquid state machine neuromorphic learning method, wherein the method comprises: A processing module, used to construct a quadratic pulse neuron model and to design a neuron digital circuit adapted to the quadratic pulse neuron model in real time; A judgment module, used for judging in real time based on preset rules whether the neuron digital circuit in the quadratic pulse neuron model can emit pulses under specific conditions; A detection module, for detecting the liquid layer neurons and the output layer neurons respectively contained in the preset liquid state machine model in real time if it is determined in real time based on preset rules that the neuron digital circuit in the quadratic pulse neuron model can emit pulses under specific conditions, and constructing an online monitoring learning algorithm based on the liquid layer neurons and the output layer neurons; A learning module is used to set an optimal learning rate and a back propagation rate corresponding to the online monitoring learning algorithm, and complete the neuromorphic learning of the liquid state machine according to the optimal learning rate, the back propagation rate and the online monitoring learning algorithm based on a preset environment.
[0024] The third aspect of the embodiment of the present invention proposes: A computer comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the liquid state machine neuromorphic learning method as described above is implemented.
[0025] The fourth aspect of the embodiments of the present invention proposes: A readable storage medium stores a computer program thereon, wherein when the program is executed by a processor, the liquid state machine neuromorphic learning method as described above is implemented.
[0026] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS
[0027] Figure 1 A flow chart of a liquid state machine neuromorphic learning method provided by the first embodiment of the present invention; Figure 2 is a calculation pipeline diagram of the neuron model parameters in the present invention; Figure 3 This is a diagram of the quadratic pulse neuron computing system architecture of the present invention; Figure 4 It is the schematic diagram of the top module of the neuron model in the present invention; Figure 5 It is a learning process diagram of the LSM dynamic network in the present invention; Figure 6 It is a schematic diagram of the STDP online supervised learning rule driven by local variables in the present invention; Figure 7 is a schematic diagram of the synaptic weight update mechanism in the present invention; Figure 8 It is a system structure diagram of the hardware implementation in the present invention; Fig. 9 It is the overall architecture diagram of the LSM learning system in the present invention; Fig.10 It is a complete structural diagram of the neuron model element in the present invention; Fig.11 It is a structural diagram of a pulse sequence receiving circuit in the present invention; Fig.12 It is a diagram of the Moore state machine model in the present invention; Fig.13 A structural block diagram of a liquid state machine neuromorphic learning system provided in the third embodiment of the present invention.
[0028] The following specific implementation manner will further illustrate the present invention in conjunction with the above-mentioned drawings. DETAILED DESCRIPTION
[0029] In order to facilitate the understanding of the present invention, the present invention will be described more fully below with reference to the relevant drawings. Several embodiments of the present invention are given in the drawings. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein. On the contrary, the purpose of providing these embodiments is to make the disclosure of the present invention more thorough and comprehensive.
[0030] It should be noted that when an element is referred to as being "fixed to" another element, it may be directly on the other element or there may be a central element. When an element is considered to be "connected to" another element, it may be directly connected to the other element or there may be a central element at the same time. The terms "vertical", "horizontal", "left", "right" and similar expressions used herein are for illustrative purposes only.
[0031] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art to which the present invention belongs. The terms used herein in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The term "and / or" used herein includes any and all combinations of one or more of the related listed items.
[0032] See also Figure 1 , shown is the liquid state machine neuromorphic learning method provided by the first embodiment of the present invention. The liquid state machine neuromorphic learning method provided in this embodiment can construct a learning system with a bidirectional control online supervised learning algorithm, which can handle classification problems of various modes and correspondingly improve the work efficiency of the staff.
[0033] Specifically, this embodiment provides: A liquid state machine neuromorphic learning method specifically comprises the following steps: Step S10, constructing a quadratic pulse neuron model, and designing a neuron digital circuit adapted to the quadratic pulse neuron model in real time; Step S20, judging in real time based on preset rules whether the neuron digital circuit in the quadratic pulse neuron model can emit pulses under specific conditions; Step S30, if it is determined in real time based on the preset rules that the neuron digital circuit in the quadratic pulse neuron model can emit pulses under specific conditions, then the liquid layer neurons and the output layer neurons respectively contained in the preset liquid state machine model are detected in real time, and an online monitoring learning algorithm is constructed based on the liquid layer neurons and the output layer neurons; Step S40, setting an optimal learning rate and a back propagation rate corresponding to the online monitoring learning algorithm, and completing the neuromorphic learning of the liquid state machine according to the optimal learning rate, the back propagation rate and the online monitoring learning algorithm based on a preset environment.
[0034] Specifically, in this example, it should be noted that in order to build a learning system that meets various models, it is necessary to perform corresponding neuromorphic learning on the created liquid state machine. Preferably, the present invention is based on the quadratic pulse neuron model constructed in real time for subsequent processing and learning. Specifically, the present invention will further pre-set the liquid layer neurons and output layer neurons contained in the liquid state machine model, and further construct the required online monitoring learning algorithm. Based on this, the neuromorphic learning of the liquid state machine is finally completed, and the verification of various pattern classification problems is completed accordingly, so as to facilitate subsequent processing.
[0035] Second embodiment For ease of understanding, the present invention further provides a second embodiment. The method provided in the second embodiment of the present invention specifically includes the following steps: Step 1: Analyze the quadratic spiking neuron model. The quadratic spiking neuron model is derived by combining the pulse generation mechanism and the post-pulse reset mechanism. In the field of neuromorphic computing, the coordinate system formed by the membrane potential V representing the instantaneous Na+ current and the recovery variable u is called the phase plane. The V-nullcline can be approximated by a parabola, and the u-nullcline can be approximated by a straight line. The V-nullcline can be approximated by:
[0036] in( , ) is the minimum coordinate value of the phase plane. -The zero line can be approximated as:
[0037] in is the slope, yes As a result, the above two equations can be dynamically approximated to obtain:
[0038] in , They represent the membrane potential and recovery variable at the next moment respectively. To avoid the variable changing to infinity when simulating neurons, an upper limit can be set for the voltage trajectory , and make In reaching After that, it is reset to a new starting value. By appropriately replacing and scaling the variables in the above equation and then deriving from the instantaneous and steady-state current-voltage relationship, we can obtain the quadratic pulse neuron model:
[0039] And the following reset conditions: if ,but
[0040] in, Represents the real-time changing membrane potential of neurons, Represents the membrane potential at the next moment; represent Current activated or The recovery variable of the current inactivation, Represents the recovery variable at the next moment; represents the membrane capacitance; is the membrane potential of the neuron when it is in the resting state; represents the instantaneous threshold potential, when At the instant of , the depolarization of the neuron will lead to the firing of a pulse; represents the excitatory input current; , , , is a dimensionless parameter.
[0041] Step 2: Design the neuron digital circuit. After the analysis of the quadratic pulse neuron model is completed, the design of the neuron digital circuit is divided into the following steps: Step 2.1: Design the neuron digital circuit operation steps. In order to evaluate the accuracy and hardware performance of the quadratic spike neuron, the finite difference approximation method is used to derive an expression that is convenient for digital circuit hardware implementation:
[0042]
[0043] The simulation of digital circuits involves solving differential equations. In order to solve the second-order derivative differential equation in digital form, the Euler broken line method is used here to discretize the time step:
[0044] Among them, the duration is the time step, so The state variables at the moment are used to calculate The state variables at the moment, and in this way, the discretized equations suitable for circuit implementation are obtained:
[0045]
[0046] The parameter calculation pipeline can be deduced from the computational requirements of the above formula. The arithmetic operations in the formula are reasonably allocated to the arithmetic function units and arranged according to the algebraic order of the operations. All data in the neuron model flows directly from input to output through the computational tree, which can maximize the parallelism of the computational tree in time (pipeline) and space (parallel arithmetic unit). The operations in each tree support the maximum throughput of the neuron model in the form of a pipeline.
[0047] Step 2.2: Design the neuron digital circuit architecture. According to the operation sequence presented by the parameter calculation pipeline, the structure of the quadratic pulse neuron model can be divided into four main modules. Through the construction of a reasonable digital circuit operation system, the program design is completed on the Quartus II software using the Verilog HDL language. The top layer of the FPGA presents the calculation process of the membrane potential V and the recovery variable u in separate diagrams (in fact, these two parts are executed in parallel). From the figure, it can be seen that this process instantiates multiple module ports and signal connections, among which the "Mult" multiplier represents the "signed_mult (out, input1, input2)" multiplication calculation function built during the program design process. Considering the case that there are negative numbers in the parameters of the neuron model, multiple register type parameters are defined during program design, and the register width is set to 32 bits to store binary data. When assigning parameters, hexadecimal parameters are used to represent binary values to reduce the complexity of code construction. The first 4 hexadecimal digits represent the integer part of the parameter, and the last 4 hexadecimal digits represent the decimal part of the parameter.
[0048] Step 3: Comparison of neuron software and hardware simulation. Pulse emission is an important way of communication between neurons. Neurons do not generate pulses spontaneously, but are caused by pulse signals transmitted from other neurons. Each biological neuron receives signals from other neurons through the dendritic contacts of the synapse. These signals also change the membrane potential of the neuron in the process of generating transmembrane current. The postsynaptic potential generated thereby is transmitted to other neurons through the axon of the neuron, and the transmitted action potential is the pulse. The present invention uses MATLAB R2016b to simulate the discharge behavior of neurons, and uses the same parameters to simulate the hardware circuit of the neuron in the Modelsim software, and compares the simulation results to detect whether the hardware circuit of the quadratic pulse neuron model can emit pulses under specific conditions.
[0049] Step 4: Design an online supervised learning algorithm for a liquid state machine based on bidirectional control. For the synaptic weights between the liquid layer neurons and the output layer neurons, the present invention adopts the LVDSTDP online pulse mode supervised learning algorithm. During the learning process, the output layer synaptic weights The learning process depends not only on the correlation of the pulse trains between the presynaptic and postsynaptic neurons, but also on the correlation of the teacher neurons. For liquid layer neurons and output layer neurons , the corresponding pulse sequence , The discrete representation of is:
[0050]
[0051] When encountering a presynaptic spike, the weight decreases with the presynaptic analog variable When a postsynaptic spike occurs, the weight increases with the postsynaptic analog variable Therefore, the STDP and anti-STDP driven by local variables have a significant impact on the synaptic weights. The adjustment rules are expressed as follows:
[0052] Implementing an online supervised learning algorithm on the synapses of the LSM output layer requires considering the expected spike train In, by making Gradually approaching To implement the online learning rules of synaptic weights. The changes in , and common impact.
[0053] for The online STDP weight adjustment rule, combined with the Widrow-Hoff rule, can be defined as:
[0054] in Represents the learning rate of the algorithm. By decomposing the above formula into a time-dependent STDP learning process and an anti-STDP learning process, the formula can be further derived as follows:
[0055] In summary, and It can be described as:
[0056] in , , The analog signals corresponding to the liquid layer, output layer, and desired pulse sequence conversion are shown in Figure 2. Output layer synaptic weights The learning rule is expressed as:
[0057] By summing up all the synaptic weight changes of LTP and LTD in the output layer, we can get the excitatory synaptic weights of the liquid layer and the input layer. and Adjustment rules:
[0058] in represents the back propagation rate, Represents the number of neurons in the output layer.
[0059] In contrast, the weight change caused by LTD Neurons back-propagated to the fluid layer and input layer and At this time and are all positive values. Similarly, the inhibitory synaptic weights of the liquid layer and the input layer can be obtained and Adjustment rules:
[0060] Step 5: Test the optimal learning rate and back propagation rate corresponding to the algorithm; secondly, conduct pulse sequence learning experiments on the LSM dynamic network and static network according to the test results. By analyzing the experimental results, the online supervised learning algorithm of the LSM dynamic network proposed in the present invention has lower network errors and smaller learning cycles than the traditional LSM static network algorithm; in LSM, the number of liquid layer neurons, the connection probability between liquid layer neurons and the pulse sequence simulation duration have an impact on the learning accuracy of the two networks. In addition, increasing the number of liquid neurons can enhance the information processing capability of LSM, but it requires more learning cycles; when other network parameters are added, the learning task becomes more complicated, resulting in a gradual decrease in learning accuracy.
[0061] Step 6: Complete the hardware implementation of the liquid state machine learning system. After the design of the online supervised learning algorithm of the liquid state machine based on bidirectional control is completed, the hardware implementation of the liquid state machine learning system is mainly carried out through the following steps: Step 6.1: Build the overall hardware architecture of the learning system. The LSM learning system is described and implemented using Verilog HDL, and is integrated on the Virtex-7 FPGA VC707 development board, where the FPGA is connected to the host through the UART communication interface, receives the input pulse sequence and sends the output final pulse sequence to achieve the purpose of collaborative design. The LSM network architecture is deployed on the FPGA master chip through program design. Note that in order for the quadratic pulse neuron model to receive multiple pulse sequences in the LSM learning system and output pulse sequences through calculation, the interface of the existing neuron hardware circuit needs to be expanded.
[0062] Step 6.2: Design the input pulse processing module.
[0063] Step 6.2.1: Design a pulse sequence transmission strategy. The LSM learning system uses a Universal Asynchronous Receiver-Transmitter (UART) serial port in an asynchronous serial communication mode to receive and send information with the host. A frame of data in the UART communication process consists of a start bit, a data bit, a parity bit, and a stop bit. The data bit is the valid information in a frame of data, and 5-8 bits can be selected. The 0th bit of the data bit is the least significant bit (LSB), and the last bit is the most significant bit (MSB). In the LSM learning system, 6 data bits are used for design. The pulse sequence is an ordered sequence composed of discrete pulse times, so it is very suitable for data transmission implemented by hardware in neuromorphic computing. When generating a pulse sequence suitable for hardware on the host, homogeneous Poisson coding is still used. Assuming that the pulse emission frequency is 40 Hz and the simulation time is 200 ms, the generated pulse sequence can send 2 frames of data through the UART serial port. The LSB and MSB of each frame of data are both identification bits. The start identifier indicates that the binary digit of the LSB is "0", which means that the other data bits of this frame store part of the pulse information; the continuation identifier indicates that the binary digit of the MSB is "1", indicating that the corresponding pulse information is incomplete and it is necessary to wait for the next LSB binary digit to be "1" (that is, the continuation identifier) to be transmitted; the end identifier indicates that the binary digit of the MSB is "0", indicating that the corresponding pulse information has been transmitted. The pulse processing module splices a complete pulse information and converts it into a 2-digit hexadecimal number and saves it in the corresponding register.
[0064] Step 6.2.2: Design the pulse sequence receiving circuit. In the FPGA master chip, all weight registers of the learning system are driven by the STDP clock to achieve pulse synchronization. The Reg register is designed to calculate the pulse sequence time difference. , which can avoid the load and frequent switching of the global clock counter to save power. The pulse train of the incoming synapse, the presynaptic register and postsynaptic registers They are used to store the pulse trains received and fired by neurons respectively. The depth of the presynaptic and postsynaptic registers is determined by the learning window In each STDP clock cycle, the synaptic learning module checks each presynaptic register in serial mode. If there is a valid pulse sequence, the time difference between the presynaptic and postsynaptic pulse sequences can be calculated by extracting the relative position of the "pulse" in the register. .
[0065] Step 6.3: Design the synaptic weight learning module circuit.
[0066] Step 6.3.1: Implement the combinatorial digital logic of STDP. STDP learning window The continuous curve when changing will cause a lot of hardware resource consumption, which will increase the cost of neuromorphic computing hardware implementation. Therefore, when implementing the STDP learning mechanism in the hardware circuit, it is necessary to adjust the synaptic weights and learning window. Here, the STDP learning rule is comprehensively implemented using a combination of digital logic gates, specifically using rectangular functions to create an approximation of the exponential operation in STDP. The STDP timing diagram can be used to design the pulse sequence time difference With learning window The hardware operation method of the discretized STDP is to first set an upper limit on the time difference. To limit ,when The synaptic weights are not updated; secondly, the weights are uniformly discretized into Levels: , Represents the number of binary digits used to define the weight register, the learning window is mapped to the lookup table to complete the discretization. In the continuous change Superscript Represents that the synaptic weight is a discrete value, and the superscript Represents that the synaptic weight is a continuous value.
[0067] Step 6.3.2: Design the weight adjustment method in the hardware circuit. In the hardware circuit, the LVDSTDP learning rule can be simplified to: ,if but
[0068] ,if but
[0069] in represents the time constant, for the liquid layer neurons and output layer neurons The corresponding pulse sequence analog quantities are respectively and In hardware operations, shift calculation can be used to process the multiplication operation between the pulse sequence analog quantity and the decay potential time constant, so the weight change rule applicable to the hardware can be expressed as:
[0070] Here and Indicates the magnitude of weight changes of different sizes. The presynaptic signal (outPre) and the postsynaptic signal (outPre) are regulated by the Moore state machine.
[0071] Step 6.3.3: Design Moore state machine. Finite state machine (FSM) is a control center that can transfer states according to input signals. It can divide a function into several parts for step-by-step processing and switch between states in a pre-designed sequence. It is mainly used to reduce the complexity of nested if statements in parallel computing program design. The output of Moore state machine combinational logic depends only on the current state, which can avoid asynchronous feedback and is safer than Mealy state machine.
[0072] Step 7: Use the system for pattern classification problems for verification. The ability of the LSM learning system to solve nonlinear pattern classification problems is verified through character recognition tasks, including the pulse sequence transmission method of the data set, the design of the host computer software, and the feasibility of using the LSM learning system implemented by the bidirectional control mechanism to recognize various time coding patterns. Finally, the experimental results are compared with the classification performance of the implemented SNN hardware system. It can be seen that the learning system proposed in the present invention has higher operating efficiency, so it can be expanded to large-scale SNNs and used for more complex classification tasks.
[0073] In summary, the liquid state machine neuromorphic learning method and system provided in the above embodiments of the present invention can construct a learning system including an online supervised learning algorithm, and can be used for the verification of various pattern classification problems, thereby improving work efficiency.
[0074] It should be noted that the above modules can be functional modules or program modules, and can be implemented by software or hardware. For modules implemented by hardware, the above modules can be located in the same processor. Alternatively, the above modules can be located in different processors in any combination.
[0075] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by an instruction execution system, device or apparatus (such as a computer-based system, a system including a processor, or other system that can fetch instructions from an instruction execution system, device or apparatus and execute instructions), or in conjunction with such instruction execution systems, devices or apparatuses. For the purposes of this specification, "computer-readable medium" can be any device that can contain, store, communicate, propagate or transmit a program for use by an instruction execution system, device or apparatus, or in conjunction with such instruction execution systems, devices or apparatuses.
[0076] More specific examples of computer-readable media (a non-exhaustive list) include the following: an electrical connection with one or more wires (electronic device), a portable computer disk case (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disk read-only memory (CDROM). In addition, the computer-readable medium may even be a paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, deciphering or, if necessary, processing in another suitable manner, and then stored in a computer memory.
[0077] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware or a combination thereof. In the above-mentioned embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.
[0078] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.
[0079] The above-described embodiments only express several implementation methods of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the present invention. It should be pointed out that, for a person of ordinary skill in the art, several modifications and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the attached claims.
Claims
1. A liquid state machine neuromorphic learning method, characterized in that: The method comprises: Constructing a quadratic pulse neuron model, and designing a neuron digital circuit adapted to the quadratic pulse neuron model in real time; Based on preset rules, it is determined in real time whether the neuron digital circuit in the quadratic pulse neuron model can emit pulses under specific conditions; If it is determined in real time based on the preset rules that the neuron digital circuit in the quadratic pulse neuron model can emit pulses under specific conditions, the liquid layer neurons and the output layer neurons respectively contained in the preset liquid state machine model are detected in real time, and an online monitoring learning algorithm is constructed based on the liquid layer neurons and the output layer neurons; An optimal learning rate and a back propagation rate corresponding to the online monitoring learning algorithm are set, and based on a preset environment, the neuromorphic learning of the liquid state machine is completed according to the optimal learning rate, the back propagation rate and the online monitoring learning algorithm.
2. The liquid state machine neuromorphic learning method according to claim 1, characterized in that: The step of designing a neuron digital circuit adapted to the quadratic pulse neuron model in real time comprises: Calculating in real time through a first preset algorithm a parameter calculation pipeline corresponding to the quadratic spiking neuron model; Detecting in real time the parameter operation sequence contained in the parameter calculation pipeline, and designing a neuron digital circuit architecture adapted to the quadratic pulse neuron model according to the parameter operation sequence; The neuron digital circuit is generated according to the parameter calculation pipeline and the neuron digital circuit architecture.
3. The liquid state machine neuromorphic learning method according to claim 2, characterized in that: The expression of the first preset algorithm is: Among them, the simulation of digital circuits involves the solution of differential equations. In order to solve the second-order derivative differential equation in digital form, the Euler broken line method is used here to discretize the time step: Among them, the duration is the time step, so The state variables at the moment are used to calculate The state variables at the moment are then looped to obtain a set of discretized equations suitable for circuit implementation: The parameter calculation pipeline is calculated accordingly, wherein: Represents the real-time changing membrane potential of neurons; represent Current activated or Recovery variables for current inactivation; represents the membrane capacitance; is the membrane potential of the neuron when it is in the resting state; represents the instantaneous threshold potential, when At the instant of , the depolarization of the neuron will lead to the firing of a pulse; represents the excitatory input current; , , , is a dimensionless parameter.
4. The liquid state machine neuromorphic learning method according to claim 1, characterized in that: The steps of constructing an online supervised learning algorithm using the liquid layer neurons and the output layer neurons include: Discretely represent the pulse sequences corresponding to the liquid layer neurons and the output layer neurons, wherein when encountering a presynaptic pulse, the decrease in the weight between the liquid layer neurons and the output layer neurons is proportional to the instantaneous value of the presynaptic simulation variable; When encountering a postsynaptic pulse, the increase in the weight between the liquid layer neuron and the output layer neuron is proportional to the postsynaptic simulation variable; A learning rule corresponding to the synaptic weight is formulated in real time, and the online detection learning algorithm is constructed based on the learning rule.
5. The liquid state machine neuromorphic learning method according to claim 4, characterized in that: The step of constructing the online detection learning algorithm based on the learning rule comprises: The change of synaptic weight is jointly affected by the pulse sequence of the liquid layer neurons and the output layer neurons; The STDP weight adjustment rule for synaptic weights online can be defined in combination with the Widrow-Hoff rule: in represents the learning rate of the algorithm, represents the adjustment rule of synaptic weights, represents the pulse train of the liquid layer neuron, represents the pulse sequence of the output layer neurons, Represents the original pulse sequence; by decomposing the above formula into a time-dependent STDP learning process and an anti-STDP learning process, the formula can be further derived as follows: In summary, and It can be described as: in , , They correspond to the analog signals converted from the liquid layer, output layer, and expected pulse sequence, respectively, and δ represents the conversion coefficient; therefore, the learning rule of the synaptic weights of the output layer is expressed as: By summing up all the synaptic weight changes of LTP and LTD in the output layer, we can get the excitatory synaptic weights of the liquid layer and the input layer. and Adjustment rules: in represents the back propagation rate, represents the number of neurons in the output layer; In contrast, the weight change caused by LTD Neurons back-propagated to the fluid layer and input layer and At this time and are all positive values; similarly, the inhibitory synaptic weights of the liquid layer and the input layer can be obtained and Adjustment rules: The online monitoring learning algorithm is constructed accordingly.
6. The liquid state machine neuromorphic learning method according to claim 1, characterized in that: The step of completing the neuromorphic learning of the liquid state machine based on the preset environment according to the optimal learning rate, the back propagation rate and the online monitoring learning algorithm includes: Build the overall hardware architecture of the learning system and design an input pulse processing module corresponding to the overall hardware architecture; Design a pulse sequence transmission strategy corresponding to the input pulse processing module, and design a corresponding pulse sequence receiving circuit and a synaptic weight learning module circuit; A Moore state machine adapted to the salient weight learning module circuit is designed, and the neuromorphic learning is completed based on the Moore state machine.
7. The liquid state machine neuromorphic learning method according to claim 6, characterized in that: The step of completing the neuromorphic learning of the liquid state machine based on the preset environment according to the optimal learning rate, the back propagation rate and the online monitoring learning algorithm also includes: Collecting the learning results of the neuromorphic learning in real time, and comparing the learning results with the classification results of the preset SNN hardware system; Whether the neural morphological learning of the liquid state machine is completed is determined in real time according to the comparison result.
8. A liquid state machine neuromorphic learning method, characterized in that: The method comprises: A processing module, used to construct a quadratic pulse neuron model and to design a neuron digital circuit adapted to the quadratic pulse neuron model in real time; A judgment module, used for judging in real time based on preset rules whether the neuron digital circuit in the quadratic pulse neuron model can emit pulses under specific conditions; A detection module, for detecting the liquid layer neurons and the output layer neurons respectively contained in the preset liquid state machine model in real time if it is determined in real time based on preset rules that the neuron digital circuit in the quadratic pulse neuron model can emit pulses under specific conditions, and constructing an online monitoring learning algorithm based on the liquid layer neurons and the output layer neurons; A learning module is used to set an optimal learning rate and a back propagation rate corresponding to the online monitoring learning algorithm, and complete the neuromorphic learning of the liquid state machine according to the optimal learning rate, the back propagation rate and the online monitoring learning algorithm based on a preset environment.
9. A computer comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the liquid state machine neuromorphic learning method according to any one of claims 1 to 7 is implemented.
10. A readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the liquid state machine neuromorphic learning method as described in any one of claims 1 to 7 is implemented.