Adaptive self-feedback LIF neuron circuit
By designing an adaptive self-feedback LIF neuron circuit, the membrane potential intensity and the connection strength between neurons are adjusted using pulse time information, which solves the problem of low accuracy of LIF neuron circuits under extreme signals and achieves higher stability and accuracy.
Patent Information
- Application Number
- CN202411162297.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-23
- Publication Date
- 2024-11-22
- Estimated Expiration
- 2044-08-23
AI Technical Summary
Existing LIF neuron circuits have low accuracy when performing tasks under extreme signal stimulation, lacking an adaptive self-feedback mechanism, resulting in insufficient stability.
An adaptive self-feedback LIF neuron circuit is designed. By combining preneuron circuits, synaptic circuits, and postneuron circuits, the membrane potential intensity is adaptively adjusted using pulse time information, and the connection strength between neurons is enhanced through a self-feedback mechanism to achieve the adaptive self-feedback function.
It improves the accuracy and stability of LIF neuron circuits when processing extreme signals, enhances the biologicality of neurons, enables them to fire pulses more accurately, and adapt to different signal intensities and noise environments.
Smart Images

Figure CN118673978B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of artificial intelligence technology, and in particular to an adaptive self-feedback LIF neuron circuit. Background Art
[0002] In order to make the neural network hardware platform have higher accuracy and stability, research teams are constantly looking for more biological neural circuit models. Figure 1 It is a fully memristive SNN (Spiking Neural Network) circuit platform based on memristor LIF (Leaky Integrate and Fire) neurons. The unsupervised synaptic weight update and pattern classification tasks are demonstrated through experimental tests, proving the feasibility and stability of the system.
[0003] In addition, in order to make the neuron working mode more biological, HH model neuron circuits based on CMOS (Complementary Metal Oxide Semiconductor) have been proposed, such as Figure 2 As shown. The HH neuron includes four ion channels, and the figure shows the functional block diagram of a single neuron model, including the first channel carrying L-type calcium (Ca-L) current, the second channel carrying fast-inactivating T-type voltage-sensitive calcium (Ca-T) current, the third channel carrying voltage-sensitive potassium (K) current, and the fourth channel carrying leak current, which represents the remaining ion flow in the cell. The leak current and K channel current charge the membrane capacitance, and the L-type and T-type Ca channels discharge the membrane capacitance. The membrane voltage is a function of the sum of the ion currents in the four channels flowing into the membrane capacitance Cmem. This model better restores how a single neuron works in the brain.
[0004] However, in these studies, no optimization was performed based on the biological working mechanism of neurons. Although they had relatively outstanding working capabilities, they were powerless under extreme signal stimulation, such as extremely small signals, extremely large signals or extremely complex signals. Summary of the invention
[0005] The present invention provides an adaptive self-feedback LIF neuron circuit, which is used to solve the defect of low accuracy of LIF neuron circuit in executing tasks for extreme signals in the prior art, and improve the accuracy and stability of the LIF neuron circuit in executing tasks.
[0006] The present invention provides an adaptive self-feedback LIF neuron circuit, comprising:
[0007] a preneuron circuit for outputting a first membrane potential value;
[0008] a synaptic circuit, wherein an input end of the synaptic circuit is connected to an output end of the pre-neuron circuit, and the synaptic circuit is used to process the first membrane potential value to obtain a second membrane potential value;
[0009] A post-neuron circuit, wherein the input end of the post-neuron circuit is connected to the output end of the synaptic circuit, the post-neuron circuit is used to perform calculations based on the pulses emitted by itself at the previous moment, input the calculation results into the activation function for processing, perform convolution operation on the processed value and the second membrane potential output by the synaptic circuit at the current moment, obtain the membrane potential value at the current moment according to the convolution operation result, and output the pulse at the current moment according to the membrane potential value.
[0010] According to an adaptive self-feedback LIF neuron circuit provided by the present invention, the post-neuron circuit is specifically calculated according to the pulse emitted by itself at the previous moment by the following formula:
[0011]
[0012] in, is the pulse emitted by the post-neuron circuit at the previous moment, is 0 or 1, n is an increasing arithmetic sequence with an initial value of 1 and a common difference of 1.
[0013] According to an adaptive self-feedback LIF neuron circuit provided by the present invention, the post-neuron circuit obtains the membrane potential value at the current moment according to the convolution operation result by the following formula:
[0014]
[0015] in, is the membrane potential value at time t, is the first membrane potential value output by the front neuron circuit at time t-1, is the pulse emitted by the post-neuron circuit at the previous moment, is the value obtained after the activation function is processed, is the convolution operation, , where C and R are the capacitance and resistance in the basic LIF neuron circuit, is the second membrane potential output by the synaptic circuit at the current moment.
[0016] According to an adaptive self-feedback LIF neuron circuit provided by the present invention, the post-neuron circuit specifically outputs a pulse at the current moment according to the membrane potential value through the following formula:
[0017]
[0018] in, is the pulse emitted by the post-neuron circuit at time t, is the membrane potential value at time t, is the preset threshold.
[0019] An adaptive self-feedback LIF neuron circuit provided by the present invention further includes:
[0020] A timing module, used for providing a clock cycle signal when the post-neuron circuit is working;
[0021] A pulse counter, wherein the input end of the pulse counter is connected to the output end of the timing module and to the output end of the post-neuron circuit, and the pulse counter is used to detect whether the post-neuron circuit issues a pulse within the clock cycle of the post-neuron circuit according to the clock cycle signal;
[0022] A pulse frequency detector, wherein the input end of the pulse frequency detector is connected to the output end of the pulse counter and to the input end of the post-neuron circuit, and the pulse frequency detector is used to detect the moment of pulse emission.
[0023] According to an adaptive self-feedback LIF neuron circuit provided by the present invention, the post-neuron circuit is also used for:
[0024] After the post-neuron circuit outputs a pulse, obtaining the time at which the post-neuron circuit outputs the pulse;
[0025] Determine whether the time when the post-neuron circuit outputs pulses is shorter than the time when surrounding neurons interconnected with the post-neuron circuit output pulses;
[0026] The judgment result is convolved with the second membrane potential currently output by the synaptic circuit, a new membrane potential value is obtained according to the convolution operation result, and a pulse is output again according to the new membrane potential value.
[0027] According to an adaptive self-feedback LIF neuron circuit provided by the present invention, the calculation formula of the judgment result is:
[0028]
[0029] in, is the judgment result, Firing pulses for the post-neuronal circuit The time min is the minimum value between the time when the surrounding neurons interconnected with the post-neuron circuit output pulses and the time when the post-neuron circuit outputs pulses.
[0030] According to an adaptive self-feedback LIF neuron circuit provided by the present invention, the formula for obtaining a new membrane potential value according to the convolution operation result is:
[0031]
[0032] Among them, among them, is the new membrane potential value at time t, is the first membrane potential value output by the front neuron circuit at time t-1, is the pulse emitted by the post-neuron circuit at the previous moment, is the judgment result, is the convolution operation, , where C and R are the capacitance and resistance in the basic LIF neuron circuit, is the second membrane potential output by the synaptic circuit at the current moment.
[0033] According to an adaptive self-feedback LIF neuron circuit provided by the present invention, the post-neuron circuit comprises:
[0034] a first function circuit, wherein an input end of the first function circuit is connected to an output end of the rear neuron cell body, and the first function circuit is used to perform calculations according to pulses emitted by the rear neuron cell body at a previous moment, and input the calculation results into an activation function for processing;
[0035] a post-neuron membrane potential generating circuit, wherein the input end of the post-neuron membrane potential generating circuit is connected to the output end of the first function circuit, and the post-neuron membrane potential generating circuit is used to perform a convolution operation on the value obtained after processing and the second membrane potential outputted by the synaptic circuit at the current moment, and obtain the membrane potential value at the current moment according to the convolution operation result;
[0036] The post-neuron cell body is used to output a pulse at the current moment according to the membrane potential value.
[0037] According to an adaptive self-feedback LIF neuron circuit provided by the present invention, the post-neuron circuit further comprises a second function circuit, the input end of the second function circuit is connected to the output end of the post-neuron cell body;
[0038] The second function circuit is used to obtain the time when the post-neuron circuit outputs the pulse after the post-neuron circuit outputs the pulse, and determine whether the time when the post-neuron circuit outputs the pulse is less than the time when the surrounding neurons interconnected with the post-neuron circuit output the pulse;
[0039] The input end of the post-neuron membrane potential generating circuit is also connected to the output end of the second function circuit, and the post-neuron membrane potential generating circuit is also used to perform a convolution operation on the judgment result and the second membrane potential outputted by the synaptic circuit at the current moment, and obtain a new membrane potential value according to the convolution operation result;
[0040] The post-neuron cell body is also used to output pulses again according to the new membrane potential value.
[0041] The adaptive self-feedback LIF neuron circuit provided by the present invention is optimized based on CMOS LIF neurons, realizes the adaptive self-feedback function through peripheral circuit control, uses the time information of the pulse neural network to output the pulse, and performs convolution operation on the original membrane potential value output by the synaptic circuit according to the value obtained by the pulse signal output at the previous moment, thereby adaptively adjusting the input intensity of its own membrane potential, and then selecting a competition mechanism according to the pulse signal of different categories of neurons at the current moment, and the winner inputs the pulse signal to the second membrane potential again for convolution operation, and finally outputs the pulse at the next moment. It improves the biological nature of neurons, has stronger stability and robustness when processing weaker signals, stronger signals and noise, and makes the pulse more accurately emitted. BRIEF DESCRIPTION OF THE DRAWINGS
[0042] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0043] Figure 1 It is a full-memristor-based SNN circuit platform provided by existing technology;
[0044] Figure 2 It is a CMOS-based HH neuron provided by existing technology;
[0045] Figure 3 It is a schematic diagram of the adaptive mechanism workflow of the adaptive self-feedback LIF neuron circuit provided by the present invention;
[0046] Figure 4 It is a schematic diagram of the ReLU function in the adaptive self-feedback LIF neuron circuit provided by the present invention;
[0047] Figure 5 It is a schematic diagram of the framework of the SFA neuron circuit in the adaptive self-feedback LIF neuron circuit provided by the present invention;
[0048] Figure 6It is a schematic diagram of the circuit structure constructed in the SFA neuron simulation in the adaptive self-feedback LIF neuron circuit provided by the present invention;
[0049] Figure 7 It is a schematic diagram of the self-feedback mechanism working process of the adaptive self-feedback LIF neuron circuit provided by the present invention;
[0050] Figure 8 It is a schematic diagram of the framework of the SFB neuron circuit in the adaptive self-feedback LIF neuron circuit provided by the present invention;
[0051] Fig. 9 It is a schematic diagram of the circuit structure constructed in the SFB neuron simulation in the adaptive self-feedback LIF neuron circuit provided by the present invention;
[0052] Fig.10 It is a schematic diagram of the working process of AF neuron in the adaptive self-feedback LIF neuron circuit provided by the present invention;
[0053] Fig.11 It is a schematic diagram of the framework of the AF neuron circuit in the adaptive self-feedback LIF neuron circuit provided by the present invention. DETAILED DESCRIPTION
[0054] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0055] Brain science and artificial intelligence research are experiencing rapid development in the process of mutual promotion, and the interdisciplinary spiking neural network (SNN) has also attracted widespread attention. In SNN, neurons are mainly responsible for analyzing and processing information and are an important part of the nervous system. Therefore, the study of neuron circuit models has become one of the important research directions in the field of SNN, and is also one of the breakthroughs in realizing the design of low-energy neuromorphic chips.
[0056] At present, the common neuron circuit models mainly include HH model, LIF model, Izhikevich model, etc. In most scenarios, these models are simply mechanically built in the circuit network to realize dynamic characteristics. However, neurons are not based on simplified leakage integration and discharge models, and the complex biological behavior of nerve cells cannot be ignored, which may limit the accuracy of the model and the ability to simulate real neuron behavior, and ultimately affect the overall performance of the network hardware platform.
[0057] In biological systems, neurons can adaptively adjust the membrane potential strength according to input information, thereby adjusting the pulse emission frequency. Based on this feature, the neuron circuit model needs to design an adaptive circuit module so that the neuron circuit can also adaptively adjust the threshold voltage in the integral circuit to simulate the neuron adaptive function. In addition, the competition mechanism between neurons is also very important. It enhances the connection strength between excited neurons through self-feedback and inhibits other neurons at the same time. Therefore, it is also necessary to design a self-feedback circuit so that the neuron circuit can improve the recognition accuracy through the competition mechanism. However, the current common neuron circuit model cannot simulate the adaptive self-feedback working mechanism well. It is necessary to design a new neuron circuit model that can not only simulate the dynamic characteristics of biological neurons, but also realize more complex biological characteristics.
[0058] Combine the following Figure 3 A schematic diagram of the adaptive mechanism workflow of an adaptive self-feedback LIF neuron circuit of the present invention is described, including:
[0059] Preneuron circuit for outputting the first membrane potential value ;
[0060] A synaptic circuit, wherein the input end of the synaptic circuit is connected to the output end of the pre-neuron circuit, and the synaptic circuit is used to Processing to obtain the second membrane potential value ;
[0061] A post-neuron circuit, wherein the input end of the post-neuron circuit is connected to the output end of the synaptic circuit, and the post-neuron circuit is used to generate a pulse according to the pulse emitted by itself at the previous moment. Calculate, input the calculation result into the activation function for processing, convolve the value obtained after processing with the second membrane potential outputted by the synaptic circuit at the current moment, and calculate the value obtained after processing according to the convolution operation result. Get the membrane potential value at the current moment , according to the membrane potential value Output the pulse of the current moment .
[0062] Because there is a delayed timing mechanism in the brain's adaptive mechanism, that is, the pulse emitted at the previous moment Will affect the pulse emitted at the next moment , so this element is included in the design process to make neurons more biological.
[0063] The pulse fired by the neuron at the previous moment At the same time as the output, it will also be used as an input signal to the function The calculation is performed. This embodiment does not limit the type of function. The value is used as input and calculated using an activation function such as the ReLU function, as shown in formula (1):
[0064] (1)
[0065] in, is the ReLU function, that is, when the input When it is greater than 0, the output is equal to the input, such as Figure 4 shown.
[0066] Finally, the value SFA obtained after ReLU function processing is compared with the second membrane potential value output by the synaptic circuit at the current moment. Perform convolution operation.
[0067] According to the convolution operation results, the final membrane potential value at the current moment is obtained For example, according to the first membrane potential value output by the front neuron circuit at the previous moment , the pulses emitted by the neuron at the previous moment , and SFA and the second membrane potential The convolution operation result is calculated to obtain the membrane potential value at the current moment Then output the pulse at this moment according to the LIF mechanism .
[0068] The improved adaptive self-feedback LIF neuron circuit in this embodiment can adapt to different tasks, so that it performs well in tasks such as processing time series data, pattern recognition and classification. Based on the improved adaptive self-feedback LIF neuron circuit in this embodiment, an intelligent chip can be developed to realize image recognition and classification.
[0069] This embodiment is optimized based on CMOS LIF neurons, realizes adaptive functions through peripheral circuit control, utilizes the time information of pulse output by the pulse neural network, and performs convolution operation on the original membrane potential value output by the synaptic circuit according to the value obtained from the pulse signal output at the previous moment, thereby adaptively adjusting the input intensity of its own membrane potential, improving the biological properties of neurons, and having stronger stability and robustness when processing weaker signals, stronger signals and noise, so that pulses are released more accurately.
[0070] On the basis of the above embodiment, the post-neuron circuit in this embodiment is specifically calculated according to the pulse emitted by itself at the previous moment by the following formula (2):
[0071] (2)
[0072] in, is the pulse emitted by the post-neuron circuit at the previous moment, is 0 or 1, n is an increasing arithmetic sequence with an initial value of 1 and a common difference of 1.
[0073] Only in =0, there is valid output, that is .
[0074] On the basis of the above embodiment, the post-neuron circuit in this embodiment specifically obtains the membrane potential value at the current moment according to the convolution operation result through the following formula (3):
[0075] (3)
[0076] in, is the final membrane potential value at time t, is the first membrane potential value output by the front neuron circuit at time t-1, is the pulse emitted by the post-neuron circuit at the previous moment, is the value obtained after the activation function is processed, is the convolution operation, , where C and R are the capacitance and resistance in the basic LIF neuron circuit, is the second membrane potential output by the synaptic circuit at the current moment.
[0077] On the basis of the above embodiment, the post-neuron circuit in this embodiment specifically outputs the pulse at the current moment according to the final membrane potential value through the following formula (4):
[0078] (4)
[0079] in, is the pulse emitted by the post-neuron circuit at time t, is the final membrane potential value at time t, is the preset threshold.
[0080] Based on the above embodiments, Figure 5 As shown, this embodiment also includes:
[0081] A timing module, used for providing a clock cycle signal when the post-neuron circuit is working;
[0082] A pulse counter, wherein the input end of the pulse counter is connected to the output end of the timing module and to the output end of the post-neuron circuit, and the pulse counter is used to detect whether the post-neuron circuit issues a pulse within the clock cycle of the post-neuron circuit according to the clock cycle signal;
[0083] A pulse frequency detector, wherein the input end of the pulse frequency detector is connected to the output end of the pulse counter and to the input end of the post-neuron circuit, and the pulse frequency detector is used to detect the moment of pulse emission.
[0084] Figure 5 The circuit framework diagram is designed according to the working flow diagram and the formulas involved in the above adaptive mechanism. Figure 6 This is the circuit diagram built during simulation. The built circuit corresponds to the circuit module in the framework diagram. In order to facilitate subsequent calculations during the simulation process, an analog-to-digital converter module (ADC) is connected after the neuron to convert the analog signal into a digital signal 0 or 1.
[0085] The timing module is the clock cycle signal when the SFA neuron is working, and the circuit only works within the clock cycle. The spiking counter and the spiking frequency detector play the role of detecting the pulse output signal. The spiking counter detects whether the pulse is released (that is, whether δ is 1), and the spiking frequency detector detects the moment when the pulse is released. The SFA function plays the role of formulas (1) and (2). When the SFA function obtains the result, it is convolved with the original membrane potential of the neuron (the ⊕ symbol in the figure represents the convolution operation). The membrane potential signal finally obtained is input into the neuron again, and the pulse signal at the next moment is output according to the LIF mechanism.
[0086] Based on the above embodiments, Figure 7 As shown, the post-neuron circuit in this embodiment is also used for:
[0087] After the post-neuron circuit outputs a pulse, obtaining the time at which the post-neuron circuit outputs the pulse;
[0088] Determine whether the time when the post-neuron circuit outputs pulses is shorter than the time when surrounding neurons interconnected with the post-neuron circuit output pulses;
[0089] The judgment result is convolved with the second membrane potential currently output by the synaptic circuit, a new membrane potential value is obtained according to the convolution operation result, and a pulse is output again according to the new membrane potential value.
[0090] The neuron self-feedback function is achieved by the interaction of multiple neurons. Each neuron follows the same rules in the self-feedback process. When the pulse signal It will be entered as input The function is processed as shown in formula (5):
[0091] (5)
[0092] Only when hour, Valid, and output the current time T.
[0093] The neurons surrounding the post-neuronal circuit are connected to each other through function Calculated Value. As in the post-neuronal circuit When the value is minimum, the judgment result SFB=1, otherwise it is 0.
[0094] Finally, the judgment result is convolved with the original membrane potential value of the neuron, and a new membrane potential value is obtained based on the convolution operation result. For example, according to the first membrane potential value output by the front neuron circuit at the previous moment , the pulses emitted by the neuron at the previous moment , and SFB and the second membrane potential The convolution operation result is calculated to obtain the new membrane potential value Then output the pulse at this moment according to the LIF mechanism , as shown in formula (4).
[0095] According to the overall process and formula analysis, in the self-feedback mechanism, only the neurons that receive the most synaptic signals can finally output effective signals, effectively filter out interference information, and make more accurate judgments. Adaptive self-feedback neurons can achieve more complex learning tasks by increasing the number of feedback neurons or using different feedback structures.
[0096] This embodiment is optimized based on CMOS LIF neurons, and realizes adaptive self-feedback function through peripheral circuit control. The original membrane potential value output by the synaptic circuit is convolved according to the value obtained by the pulse signal output by the post-neuron, so as to adaptively adjust the input intensity of its own membrane potential, improve the biology of neurons, and have stronger stability and robustness when processing weak signals, strong signals and noise, so that pulses are released more accurately.
[0097] Based on the above embodiment, the calculation formula of the judgment result in this embodiment is formula (6):
[0098] (6)
[0099] in, is the judgment result, Firing pulses for the post-neuronal circuit The time min is the minimum value between the time when the surrounding neurons interconnected with the post-neuron circuit output pulses and the time when the post-neuron circuit outputs pulses.
[0100] The post-neuronal circuit Input to the next level function middle, When the value is minimum, the function It is valid only when SFB=1, otherwise it is 0.
[0101] On the basis of the above embodiment, the formula for obtaining a new membrane potential value according to the convolution operation result in this embodiment is formula (7):
[0102] (7)
[0103] Among them, among them, is the new membrane potential value at time t, is the first membrane potential value output by the front neuron circuit at time t-1, is the pulse emitted by the post-neuron circuit at the previous moment, is the judgment result, is the convolution operation, , where C and R are the capacitance and resistance in the basic LIF neuron circuit, is the second membrane potential output by the synaptic circuit at the current moment.
[0104] Figure 8 In order to design a circuit framework based on the self-feedback mechanism flowchart, formula (5) and formula (6) are put into a circuit module for processing, and this module is called SFB function circuit module. When the neuron emits a pulse, the pulse signal It enters the SFB function circuit module as input for processing, and the processed output value is convolved with the original membrane potential to generate a new membrane potential signal and emit a pulse. Fig. 9 This is the circuit built during the simulation process. Here, 10 neuron outputs are taken as an example. Each neuron affects each other and works according to the designed self-feedback mechanism. At the same time, 10 current sources are used to simulate the synaptic circuit and 10 neurons in cascade. Each neuron output is connected to the SFB function circuit module, and the signal is fed back to the neuron input after being processed in this module.
[0105] On the basis of the above embodiment, the post-neuron circuit in this embodiment includes:
[0106] a first function circuit, wherein an input end of the first function circuit is connected to an output end of the rear neuron cell body, and the first function circuit is used to perform calculations according to pulses emitted by the rear neuron cell body at a previous moment, and input the calculation results into an activation function for processing;
[0107] a post-neuron membrane potential generating circuit, wherein the input end of the post-neuron membrane potential generating circuit is connected to the output end of the first function circuit, and the post-neuron membrane potential generating circuit is used to perform a convolution operation on the value obtained after processing and the second membrane potential outputted by the synaptic circuit at the current moment, and obtain a final membrane potential value at the current moment according to the convolution operation result;
[0108] The post-neuron cell body is used to output a pulse at the current moment according to the final membrane potential value.
[0109] On the basis of the above embodiment, the post-neuron circuit in this embodiment further includes a second function circuit, the input end of the second function circuit is connected to the output end of the post-neuron cell body;
[0110] The second function circuit is used to obtain the time when the post-neuron circuit outputs the pulse after the post-neuron circuit outputs the pulse, and determine whether the time when the post-neuron circuit outputs the pulse is less than the time when the surrounding neurons interconnected with the post-neuron circuit output the pulse;
[0111] The input end of the post-neuron membrane potential generating circuit is also connected to the output end of the second function circuit, and the post-neuron membrane potential generating circuit is also used to perform a convolution operation on the judgment result and the second membrane potential outputted by the synaptic circuit at the current moment, and obtain a new membrane potential value according to the convolution operation result;
[0112] The post-neuron cell body is also used to output pulses again according to the new membrane potential value.
[0113] The adaptive mechanism and the self-feedback mechanism are combined to design an adaptive self-feedback neuron (AF neuron) that has both adaptive and self-feedback mechanisms. The adaptive mechanism and the self-feedback mechanism still use the above processing method, but the workflow needs to be redesigned so that the neuron can continuously complete the two working mechanisms of adaptive and self-feedback. Secondly, the new circuit framework is redesigned according to the workflow to realize the AF neuron circuit design.
[0114] Fig.10 The figure shows the working process of AF neurons. It can be seen from the figure that the adaptive mechanism is to to process, and the self-feedback mechanism is Therefore, the two have a certain order in the working process, that is, the adaptive mechanism works first and then the self-feedback mechanism starts to work. The human brain works in the same way. When a person is suddenly in an environment, the neurons will first make adaptive adjustments to adapt to the signal stimulation of the environment, and then in the subsequent signal reception process, feedback propagation will be carried out to adjust the connection strength of each neuron, and finally the body will respond accordingly.
[0115] First, when the pulse signal of the previous moment When outputting, the AF neuron first performs the adaptive function, that is, according to formula (2), The value of is then used as input to the activation function (ReLU) for calculation, as shown in formula (1). Finally, the output value of the activation function is convolved with the original membrane potential of the neuron according to formula (3) and the pulse signal at this moment is released according to the LIF mechanism of formula (4). , so far the adaptive mechanism is complete.
[0116] Then the self-feedback mechanism starts to work, and the pulse signal after adaptive processing It is processed as self-feedback input again. First, according to formula (5), Function value, then input to the next level function As shown in formula (6), at this time, it is the same as when the self-feedback mechanism works alone. Before the operation, the surrounding interconnected neurons have been passed through the function The calculation is value, only when When the value is minimum, the function It is only valid, that is, SFB = 1, otherwise it is 0. Finally, the obtained value (SFB) is convolved with the neuron membrane potential value, as shown in formula (7), and then the next moment pulse signal is released according to the LIF mechanism of formula (4), completing all the work of the AF neuron.
[0117] like Fig.11 , the circuit framework diagram is designed according to the above workflow. Since the adaptive and self-feedback working mechanism of AF neurons is consistent with the corresponding SFA neurons and SFB neurons, when designing the circuit framework, it is only necessary to adjust the working timing of the two according to the workflow, without changing the core circuit framework during operation.
[0118] Adaptive self-feedback neurons are a concept in the field of neuroscience and play an important role in understanding and simulating brain functions. They can regulate and feedback their own activities according to the intensity and frequency of input signals. This mechanism enables neurons to adapt to different environments and input conditions, improve information processing efficiency and accuracy, and provide a basis for building neural network hardware platforms and intelligent systems.
[0119] This embodiment uses the LIF neuron circuit model as a basis, proposes an adaptive self-feedback working mechanism, and designs a neuron circuit model with an adaptive self-feedback mechanism to achieve more biological functions, so as to be applied in the SNN hardware platform to achieve more complex calculations and achieve higher accuracy and stability.
[0120] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. An adaptive self-feedback LIF neuron circuit, characterized in that: include: a preneuron circuit for outputting a first membrane potential value; a synaptic circuit, wherein an input end of the synaptic circuit is connected to an output end of the pre-neuron circuit, and the synaptic circuit is used to process the first membrane potential value to obtain a second membrane potential value; A post-neuron circuit, wherein the input end of the post-neuron circuit is connected to the output end of the synaptic circuit, the post-neuron circuit is used to perform calculations based on the pulses emitted by itself at the previous moment, input the calculation results into the activation function for processing, perform convolution operation on the value obtained after processing and the second membrane potential output by the synaptic circuit at the current moment, obtain the membrane potential value at the current moment according to the convolution operation result, and output the pulse at the current moment according to the membrane potential value at the current moment; a competition mechanism is selected according to the pulses of different categories of neurons at the current moment, and the winner inputs the pulse into the second membrane potential again for convolution operation, and finally outputs the pulse at the next moment.
2. The adaptive self-feedback LIF neuron circuit according to claim 1, characterized in that: The post-neuron circuit is specifically calculated according to the pulses emitted by itself at the previous moment by the following formula: ; in, is the pulse emitted by the post-neuron circuit at the previous moment, It is an increasing arithmetic sequence of 0 or 1, with the initial value of n being 1 and the common difference being 1.
3. The adaptive self-feedback LIF neuron circuit according to claim 1, characterized in that: The post-neuron circuit specifically obtains the membrane potential value at the current moment according to the convolution operation result through the following formula: ; in, is the membrane potential value at time t, is the first membrane potential value output by the front neuron circuit at time t-1, is the pulse emitted by the post-neuron circuit at the previous moment, is the value obtained after the activation function is processed, is the convolution operation, , C and R are the capacitance and resistance in the basic LIF neuron circuit, is the second membrane potential output by the synaptic circuit at the current moment.
4. The adaptive self-feedback LIF neuron circuit according to claim 1, characterized in that: The post-neuron circuit specifically outputs the pulse at the current moment according to the membrane potential value through the following formula: ; in, is the pulse emitted by the post-neuron circuit at time t, is the membrane potential value at time t, is the preset threshold.
5. The adaptive self-feedback LIF neuron circuit according to claim 1, characterized in that: Also includes: A timing module, used for providing a clock cycle signal when the post-neuron circuit is working; A pulse counter, wherein the input end of the pulse counter is connected to the output end of the timing module and to the output end of the post-neuron circuit, and the pulse counter is used to detect whether the post-neuron circuit issues a pulse within the clock cycle of the post-neuron circuit according to the clock cycle signal; A pulse frequency detector, wherein the input end of the pulse frequency detector is connected to the output end of the pulse counter and to the input end of the post-neuron circuit, and the pulse frequency detector is used to detect the moment of pulse emission.
6. The adaptive self-feedback LIF neuron circuit according to any one of claims 1 to 5, characterized in that: The post-neuron circuit is also used for: After the post-neuron circuit outputs a pulse, obtaining the time at which the post-neuron circuit outputs the pulse; Determine whether the time when the post-neuron circuit outputs pulses is shorter than the time when surrounding neurons interconnected with the post-neuron circuit output pulses; The judgment result is convolved with the second membrane potential currently output by the synaptic circuit, a new membrane potential value is obtained according to the convolution operation result, and a pulse is output again according to the new membrane potential value.
7. The adaptive self-feedback LIF neuron circuit according to claim 6, characterized in that: The calculation formula of the judgment result is: ; in, is the judgment result, Firing pulses for the post-neuronal circuit The time min is the minimum value between the time when the surrounding neurons interconnected with the post-neuron circuit output pulses and the time when the post-neuron circuit outputs pulses.
8. The adaptive self-feedback LIF neuron circuit according to claim 6, characterized in that: The formula for obtaining a new membrane potential value according to the convolution operation result is: ; Among them, among them, is the new membrane potential value at time t, is the first membrane potential value output by the front neuron circuit at time t-1, is the pulse emitted by the post-neuron circuit at the previous moment, is the judgment result, is the convolution operation, , where C and R are the capacitance and resistance in the basic LIF neuron circuit, is the second membrane potential output by the synaptic circuit at the current moment.
9. The adaptive self-feedback LIF neuron circuit according to claim 6, characterized in that: The post-neuron circuit comprises: a first function circuit, wherein an input end of the first function circuit is connected to an output end of the rear neuron cell body, and the first function circuit is used to perform calculations according to pulses emitted by the rear neuron cell body at a previous moment, and input the calculation results into an activation function for processing; a post-neuron membrane potential generating circuit, wherein the input end of the post-neuron membrane potential generating circuit is connected to the output end of the first function circuit, and the post-neuron membrane potential generating circuit is used to perform a convolution operation on the value obtained after processing and the second membrane potential outputted by the synaptic circuit at the current moment, and obtain the membrane potential value at the current moment according to the convolution operation result; The post-neuron cell body is used to output a pulse at the current moment according to the membrane potential value.
10. The adaptive self-feedback LIF neuron circuit according to claim 9, characterized in that: The post-neuron circuit further comprises a second function circuit, the input end of the second function circuit is connected to the output end of the post-neuron cell body; The second function circuit is used to obtain the time when the post-neuron circuit outputs the pulse after the post-neuron circuit outputs the pulse, and determine whether the time when the post-neuron circuit outputs the pulse is less than the time when the surrounding neurons interconnected with the post-neuron circuit output the pulse; The input end of the post-neuron membrane potential generating circuit is also connected to the output end of the second function circuit, and the post-neuron membrane potential generating circuit is also used to perform a convolution operation on the judgment result and the second membrane potential outputted by the synaptic circuit at the current moment, and obtain a new membrane potential value according to the convolution operation result; The post-neuron cell body is also used to output pulses again according to the new membrane potential value.
Citation Information
Patent Citations
Signal processing method for neurons in spiking neural network and network training method
CN113255905A
Low-power-consumption target detection method based on convolutional spiking neural network
CN114926737A
Pulse circuit based on neuron membrane high and low potential sampling
CN116306857A
LIF neuron circuit
CN116523010A
Signal processing method for neuron in spiking neural network and method for training said network
US20230385617A1
