Multi-valued spiking neuron processing method based on dynamic threshold feedback reset

Abstract

The invention provides a multi-valued spiking neuron processing method based on dynamic threshold feedback reset, and relates to the technical field of brain-like intelligent calculation, and the method comprises the steps: introducing a dynamic threshold in explicit association with an input current, a multi-valued issuing mechanism, and combining the feedback reset adjustment of a historical pulse behavior; the neurons have higher information expression ability and adaptive state regulation and control ability in a single time step; meanwhile, a substitution gradient coupled with a dynamic threshold value is introduced for multi-valued distribution to guarantee gradient stable propagation in the training stage, and multi-valued pulses are equivalently converted into sparse binary events in the reasoning stage, so that the low energy consumption characteristic of event-driven calculation is kept while the expression ability is improved. According to the invention, neuron processing of multi-valued pulse expression, threshold and reset adaptive adjustment, training stage gradient stability and inference stage event-driven energy efficiency maintenance problems can be simultaneously solved on the whole.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of brain-like intelligent computing, and in particular to a multi-value pulse neuron processing method based on dynamic threshold feedback reset. BACKGROUND

[0002] Pulse neural network is a kind of neural computing model inspired by the information transmission mechanism of biological nervous system, which transmits information through discrete pulse events rather than continuous numerical values, has the characteristics of event-driven and time-series modeling, and is widely concerned in the field of low-power computing and neuromorphic data processing. Most existing pulse neural networks are constructed based on the leaky integrate-and-fire model, which integrates the input current over time and combines with a fixed threshold to determine the pulse firing behavior, so that the spatial information and temporal information are unified in the membrane potential. When the membrane potential exceeds the preset threshold, the neuron generates a pulse and performs a membrane potential reset, thereby realizing the state evolution across time steps. To improve the flexibility of the model, some studies introduce a learnable threshold or a soft reset mechanism, so that the threshold parameter or the reset amplitude can be adjusted during training, thereby improving the expression ability and training performance of the model to some extent. At the same time, some studies also try to extend the time step length or stack the network depth to alleviate the problem of insufficient expression ability of binary pulse.

[0003] However, the existing technology still has the problem of being difficult to simultaneously consider information expression ability, training stability and inference energy efficiency as a whole. On the one hand, traditional pulse neurons usually use binary pulses for information transmission, and the continuously changing membrane potential is compressed into a single event result during firing, resulting in irreversible loss of rich spatial information contained in the membrane potential amplitude. Even if multi-value pulses are introduced to improve the information carrying capacity of single firing, multiple implicit thresholds will be introduced during training, making it more difficult for gradient propagation based on non-differentiable firing functions, and traditional surrogate gradient methods are difficult to adapt stably, which may cause problems such as gradient interval mismatch or training divergence. On the other hand, the existing threshold and reset mechanism mostly adopts a fixed or weak adaptive strategy, which fails to fully adjust the input state and historical firing behavior of the neuron, which may easily lead to long-term over-activation or long-term inhibition of some neurons, thereby affecting the dynamic balance and generalization performance of the network as a whole. In addition, multi-value pulses often need to participate in calculation with integer amplitude during inference, which destroys the original sparse event-driven characteristics of pulse neural networks, making the calculation process degenerate into high-energy multiplication and addition operation, and weakening the advantages of pulse neural networks in energy efficiency. Therefore, the existing technology urgently needs a neuron processing method that can simultaneously solve the problems of multi-value pulse expression, adaptive adjustment of threshold and reset, gradient stability during training, and event-driven energy efficiency during inference as a whole. SUMMARY

[0004] The application provides a multi-value pulse neuron processing method based on dynamic threshold value feedback reset, which can solve the problems of multi-value pulse expression, adaptive adjustment of threshold value and reset, gradient stability in training stage and event-driven energy efficiency maintenance in inference stage.

[0005] In a first aspect of the application, a multi-value pulse neuron processing method based on dynamic threshold value feedback reset is provided, and the method comprises the following steps: At each time step, a preset static threshold value is subjected to nonlinear mapping based on the input current received by the neuron, and a dynamic threshold value corresponding to the current input state of the neuron is generated in combination with a threshold adjustment parameter; Based on the dynamic threshold value, a pulse state variable is introduced to determine whether a pulse event occurs in the current time step, and a learnable substitute gradient bound to the pulse state variable is constructed according to the characteristics of the pulse state variable at the pulse trigger position; An instantaneous feedback quantity is generated based on the input current and the pulse state variable, and a cumulative feedback quantity with time memory characteristics is introduced to weight and fuse the feedback quantities of historical time steps and the instantaneous feedback quantity of the current time step, so as to form a cumulative feedback result for adjusting the reset strength of the membrane potential; The cumulative feedback result and a soft reset voltage determined based on the dynamic threshold value jointly act on the membrane potential reset process to obtain a feedback reset voltage; After the membrane potential is updated by leaky integration and the feedback reset voltage is applied, a multi-value firing function composed of a truncation function and a floor function is used to map the relationship between the membrane potential and the dynamic threshold value into a discrete multi-value pulse output; Based on the membrane potential update equation, the dynamic threshold value, the multi-value firing function and the feedback reset voltage, a discrete dynamic model of a dynamic threshold value feedback reset leaky integration multi-value firing neuron is constructed; In the training stage, a substitute gradient function associated with the dynamic threshold value is constructed for a plurality of implicit threshold values corresponding to the multi-value firing neuron, in the inference stage, the multi-value pulse output is converted into an equivalent binary pulse representation, and the bit-level redundancy is reduced through an integer-to-binary conversion mode, so that the calculation process based on the multi-value pulse is remapped into an event-driven operation mainly based on sparse accumulation.

[0006] In a second aspect of the application, a multi-value pulse neuron processing device based on dynamic threshold value feedback reset is provided, which is used to execute any one of the multi-value pulse neuron processing methods based on dynamic threshold value feedback reset described above, and the device comprises an acquisition module, a processing module and an output module, wherein: The acquisition module is configured to perform, at each time step, a non-linear mapping on a preset static threshold value based on an input current received by a neuron, and generate a dynamic threshold value corresponding to a current input state of the neuron in combination with a threshold adjustment parameter; The processing module is configured to introduce a pulse state variable based on the dynamic threshold value to determine whether a pulse event occurs at the current time step for the neuron, and construct a learnable substitute gradient bound to the pulse state variable according to a characteristic of the pulse state variable at a pulse trigger position; The processing module is configured to generate an instantaneous feedback quantity based on the input current and the pulse state variable, and form a cumulative feedback result for adjusting a reset strength of a membrane potential by introducing a cumulative feedback quantity with a time memory characteristic to perform weighted fusion of the feedback quantity at a historical time step and the instantaneous feedback quantity at the current time step. The processing module is configured to cause the cumulative feedback result and a soft reset voltage determined based on the dynamic threshold value to jointly act on a membrane potential reset process to obtain a feedback reset voltage. The processing module is configured to, after the membrane potential is updated by a leaky integration and the feedback reset voltage is applied, map a relationship between the membrane potential and the dynamic threshold value to a discrete multi-valued pulse output by using a multi-valued firing function composed of a truncation function and a floor function. The processing module is configured to construct a discrete dynamic model of a dynamic threshold value feedback reset leaky integration multi-valued firing neuron based on a membrane potential update equation, the dynamic threshold value, the multi-valued firing function, and the feedback reset voltage. The output module is configured to, in a training phase, construct a substitute gradient function associated with the dynamic threshold value for a plurality of implicit threshold values corresponding to the multi-valued firing neuron, convert the multi-valued pulse output into an equivalent binary pulse representation in an inference phase, and reduce bit-level redundancy by an integer-to-binary conversion manner, so that a calculation process based on the multi-valued pulse is remapped to an event-driven operation mainly based on sparse accumulation.

[0007] In a third aspect of the present application, an electronic device is provided, which includes a processor, a memory, a user interface, and a network interface, the memory is configured to store instructions, the user interface and the network interface are configured to communicate with other devices, and the processor is configured to execute the instructions stored in the memory to enable the electronic device to perform the method according to any one of the preceding aspects.

[0008] In a fourth aspect of the present application, a non-transitory computer-readable storage medium is provided, which stores instructions that, when executed, perform the method according to any one of the preceding aspects.

[0009] To sum up, the one or more technical solutions provided in the embodiments of the present application have at least the following technical effects or advantages: 1. The present application introduces a dynamic threshold explicitly associated with the input current at the neuron level, allowing the threshold to vary independently in space and time and participate in the firing and resetting process as a unified decision scale, thereby improving the information expression capability of single firing without increasing the time step or network depth; at the same time, by decoupling the multi-value firing decision from the pulse state variable and configuring a learnable alternative gradient for the pulse state variable explicitly associated with the dynamic threshold, the multiple implicit thresholds corresponding to the multi-value pulse all have stable, limited and self-adaptive gradient channels in the training stage, fundamentally alleviating the gradient mismatch problem caused by the non-differentiable firing function; further, the introduction of a feedback reset mechanism combining input state and historical pulse behavior allows the membrane potential reset strength to have memory in time and adaptively adjust with the neuron state, effectively suppressing the over-activation and long-term inhibition phenomenon and maintaining network dynamic balance; in the inference stage, by equivalently splitting the multi-value pulse amplitude into binary pulse events and introducing an integer-to-binary conversion method, the multi-value information is remapped to sparse event-driven operation while maintaining numerical equivalence, thereby improving the expression capability while avoiding the energy degradation caused by integer multiplication and addition, thus overall balancing multi-value pulse expression, threshold and reset adaptive adjustment, training stability and inference energy efficiency.

[0010] 2. By introducing the input current into the nonlinear mapping and coupling it with the threshold adjustment parameter to generate a dynamic threshold, the threshold is no longer a global constant, but can adaptively change with the input state at the neuron granularity and time granularity, thereby significantly enhancing the heterogeneity between neurons and the information discrimination ability of single firing without increasing the complexity of network structure, avoiding the homogenization of firing behavior of different neurons under the same threshold constraint, and facilitating the improvement of overall expression capability and feature discrimination capability.

[0011] 3. By explicitly introducing a pulse state variable on top of the multi-value firing result and constructing a learnable alternative gradient for the pulse state variable, the sign determination semantics of the pulse event remains unchanged in forward propagation, while a continuous controllable gradient path is restored in backward propagation, thereby effectively solving the gradient interruption problem caused by non-differentiable decision under multi-value firing conditions, allowing the firing boundaries corresponding to multiple implicit thresholds to be stably trained, and significantly improving the convergence and stability of multi-value pulse neurons in the training stage.

[0012] 4. By jointly mapping the input current and the pulse state variable into the instantaneous feedback quantity, and further introducing the cumulative feedback quantity updated across time steps, the reset strength of the membrane potential is simultaneously constrained by the current input driving state and the historical pulse behavior trend, thereby introducing a memory regulation capability in the time dimension for the reset mechanism, effectively suppressing the problem of long-term overactivation or long-term inhibition of neurons, and enhancing the dynamic balance and generalization stability of the network as a whole.

[0013] 5. By using the dynamic threshold as a unified firing scale, the continuous proportional relationship between the effective membrane potential and the threshold is truncated and rounded down to map the discrete multi-value pulse output, so that the firing result in a single time step can express different degrees of threshold information in the form of integer levels, thereby significantly improving the spatial information carrying capacity of a single pulse while maintaining the discreteness of the pulse, and avoiding the severe compression of membrane potential amplitude information by traditional binary firing.

[0014] 6. By fusing the cumulative feedback result and the soft reset voltage corresponding to the dynamic threshold to generate a feedback reset voltage, the membrane potential reset process includes both the basic reset information related to the current threshold state and the modulation information related to the historical pulse behavior, thereby realizing adaptive adjustment of the reset strength, avoiding the dynamic imbalance problem caused by fixed or single-scale reset mechanism, and further improving the stability and controllability of neuron state evolution.

[0015] 7. By constructing an alternative gradient function explicitly associated with the dynamic threshold for multiple implicit thresholds under the multi-value firing structure during the training phase, the gradient is only propagated within the effective threshold neighborhood, fundamentally alleviating the gradient interval mismatch problem caused by multi-value firing; and during the inference phase, the multi-value pulse is converted into a binary pulse set and combined with the integer-to-binary conversion method, the original calculation process dependent on the integer amplitude is remapped to sparse event-driven accumulation operation, thereby significantly reducing the number of synaptic operations and overall energy consumption while maintaining numerical equivalence and inference accuracy, restoring the low-power advantage of pulse neural networks during the inference phase. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 is a flowchart of a multi-value pulse neuron processing method based on dynamic threshold feedback reset disclosed by an embodiment of the present application; Figure 2 is a module diagram of a multi-value pulse neuron processing device based on dynamic threshold feedback reset disclosed by an embodiment of the present application; Figure 3 is a structural diagram of an electronic device disclosed by an embodiment of the present application.

[0017] Explanation of reference signs: 201, acquisition module; 202, processing module; 203, output module; 301, processor; 302, communication bus; 303, user interface; 304, network interface; 305, memory. DETAILED DESCRIPTION

[0018] In order to make the technical solution in the specification better understood by the skilled in the art, the technical solution in the specification will be clearly and completely described below in combination with the drawings in the embodiments of the specification. Obviously, the described embodiments are only some of the embodiments of the present application, not all the embodiments.

[0019] In the description of the embodiments of the present application, the words such as "for example" or "for instance" are used to represent an example, illustration or description. Any embodiment or design scheme described as "for example" or "for instance" in the embodiments of the present application should not be interpreted as more preferred or more advantageous than other embodiments or design schemes. Rather, the words such as "for example" or "for instance" are intended to present the relevant concept in a specific way.

[0020] In the description of the embodiments of the present application, the term "a plurality of" means two or more. For example, a plurality of systems means two or more systems, and a plurality of screen terminals means two or more screen terminals. In addition, the terms "first" and "second" are used for description purposes only, and should not be interpreted as indicating or implying relative importance or implicitly indicating the indicated technical features. Therefore, the features defined with "first" and "second" can explicitly or implicitly include one or more of the features. The terms "include", "contain", "have" and their variants mean "include but are not limited to", unless otherwise specifically emphasized.

[0021] Although the existing pulse neural network has the advantages of event-driven and timing modeling, it is still difficult to balance the information expression ability, training stability and inference efficiency in practical application: the neurons based on the traditional leaky integrate-and-fire model usually use binary pulses for information transmission, and the continuous membrane potential is compressed into a single event result during the firing process, resulting in serious loss of spatial information; although the introduction of multi-value pulses can improve the information carrying capacity of single firing, it introduces multiple implicit thresholds in the training stage, making the gradient propagation under the non-differentiable firing function more complex, and the existing alternative gradient method is difficult to adapt stably, which can easily cause unstable training; at the same time, the existing threshold and reset mechanism is mostly fixed or weakly adaptive, which cannot be adjusted jointly with the input state and historical firing behavior of the neuron, which can easily cause over-activation or long-term inhibition of the neuron, and destroy the dynamic balance of the network; in addition, the multi-value pulse usually participates in the calculation with an integer amplitude in the inference stage, which destroys the original sparse event-driven characteristics of the pulse neural network, and makes the calculation process degenerate into high-energy multiplication and addition operation, thereby weakening its low-power advantage, so it is urgent to develop a neuron processing method that can coordinate the multi-value pulse expression, adaptive threshold and reset adjustment, gradient stability in the training stage, and event-driven efficiency in the inference stage as a whole.

[0022] The embodiment discloses a multi-value pulse neuron processing method based on dynamic threshold feedback reset, referring to Figure 1 , comprising the following steps S110-S170: S110, at each time step, performing nonlinear mapping on a preset static threshold based on the input current received by the neuron, and generating a dynamic threshold corresponding to the current input state of the neuron in combination with a threshold adjustment parameter.

[0023] The multi-value pulse neuron processing method based on dynamic threshold feedback reset disclosed by the embodiment of the application is applied to a server. The server includes but is not limited to electronic devices such as mobile phones, tablet computers, wearable devices, PC (Personal Computer), and the like, and can also be a background server running a multi-value pulse neuron processing method based on dynamic threshold feedback reset. The server can be realized by an independent server or a server cluster composed of multiple servers.

[0024] In one possible implementation, a preset static threshold value is subjected to a nonlinear mapping based on an input current received by a neuron, and a dynamic threshold value corresponding to a current input state of the neuron is generated in combination with a threshold adjustment parameter, specifically including: introducing the input current into a preset nonlinear mapping function to perform amplitude compression and sign preservation processing to obtain a threshold modulation intermediate quantity limited to a continuous bounded interval; introducing a threshold adjustment parameter corresponding to the neuron, and performing weighted modulation on the threshold modulation intermediate quantity using the threshold adjustment parameter to obtain a threshold modulation quantity; and fusing the threshold modulation quantity with the preset static threshold value to generate a dynamic threshold value corresponding to the input state of the neuron at a current time step.

[0025] Specifically, in generating the dynamic threshold value, first, the threshold modulation intermediate quantity is constructed around the input current, so that the intermediate quantity can not only reflect the change trend of the input current, but also avoid unstable threshold modulation caused by too large amplitude span of the input current. In a specific implementation, an input current corresponding to a target neuron spatial position is obtained at each time step, and the input current is sent into a preset nonlinear mapping function to perform amplitude compression and sign preservation processing, where the amplitude compression is used to map the input current from a possible unbounded or wide range distribution to a continuous and bounded range, so that the threshold modulation will not be uncontrollably amplified for an abnormally large current, and the sign preservation is used to retain the positive and negative directions of the input current, so that the threshold modulation can distinguish the driving effects of the enhancement firing tendency and the inhibition firing tendency. The preset nonlinear mapping function can adopt a hyperbolic tangent function, so that the mapping output is continuous, monotonic and bounded in value, and has high sensitivity to small amplitude input and gradually saturates to large amplitude input, thereby stably providing the threshold modulation intermediate quantity at the initial training stage and the input fluctuation stage. Wherein, the input current refers to the synaptic convergence current or equivalent driving quantity received by the neuron at the current time step, the preset nonlinear mapping function refers to a nonlinear transformation form fixedly selected and consistently used in the model structure for all time steps, the amplitude compression refers to mapping the amplitude of the input current to a limited interval to suppress the influence of extreme values, the sign preservation refers to keeping the mapping result consistent with the positive and negative directions of the input current, and the threshold modulation intermediate quantity refers to an intermediate variable obtained through nonlinear mapping and not yet introduced into the learnable modulation intensity, which is used for subsequent weighted modulation and threshold fusion.

[0026]

[0027] wherein, represents the threshold modulation intermediate quantity, which is obtained by nonlinear mapping of the input current and is limited to a continuous bounded interval; represents the input current at the current time step, which is obtained by converging the pulses of the upper layer through the synaptic weights and corresponds to the spatial position of the target neuron; represents the hyperbolic tangent function, which realizes amplitude compression and sign direction preservation through saturation characteristics, so that the threshold modulation intermediate quantity is stable in value and maintains a smooth response to input changes After obtaining the threshold modulation intermediate quantity, a threshold adjustment parameter is introduced to weight modulate it, so that different neurons or different network layers can form differentiated threshold modulation strength, thereby realizing the heterogeneous threshold mechanism at the neuron level. In specific implementation, the threshold adjustment parameter is configured for each neuron or for each network layer, and the threshold adjustment parameter is involved in training update as a learnable parameter, so that it can adaptively adjust the threshold modulation amplitude according to the backpropagation signal of the loss function. In terms of parameter constraints, in order to suppress the threshold from swinging sharply due to excessive threshold modulation, the threshold adjustment parameter can be constrained in a value range, so that it always falls within the preset interval during training. The result of weighted modulation forms the threshold modulation quantity, which maintains a same shape corresponding relationship with the threshold modulation intermediate quantity, thereby ensuring that the dynamic threshold can be generated point by point independently in space and consistent with the arrangement of neurons. Among them, the threshold adjustment parameter refers to a learnable parameter for adjusting the contribution strength of the nonlinear mapping result, the weighted modulation refers to scaling the threshold modulation intermediate quantity by the threshold adjustment parameter to obtain the threshold modulation quantity, the threshold modulation quantity refers to the information quantity that has contained the learnable modulation strength, and is used to fuse with the static threshold to form the dynamic threshold.

[0028]

[0029] Among them, The threshold modulation quantity is obtained by weighting modulating the threshold modulation intermediate quantity by the threshold adjustment parameter. The threshold adjustment parameter can be limited to to constrain the threshold modulation amplitude and improve the training stability; The threshold modulation intermediate quantity is obtained by nonlinear mapping of the input current and maintains the characteristics of sign direction and continuous boundedness After obtaining the threshold modulation amount, the threshold modulation amount is fused with a preset static threshold to form a dynamic threshold, so that the dynamic threshold has both adaptability to changes in input state and numerical stability around a fixed reference. In a specific implementation, the preset static threshold is set as a constant reference value, and is regarded as a base item of the dynamic threshold. The dynamic threshold of the current time step is obtained by fusing the threshold modulation amount, and the fusion manner can adopt additive fusion, so that the threshold modulation amount produces positive or negative deviation to the static threshold, thereby realizing up-regulation or down-regulation of the threshold. Under this mechanism, when the input current after nonlinear mapping is positive and the threshold adjustment parameter is positive, the dynamic threshold is improved relative to the static threshold, so that the pulse triggering is more strict to suppress the over-activation trend. When the input current after nonlinear mapping is negative or the threshold adjustment parameter is negative, the dynamic threshold is reduced relative to the static threshold, so that the pulse triggering is easier to reduce the long-term dormancy risk. And the dynamic threshold is continuously called as a unified threshold constraint in subsequent multi-value firing function and soft reset voltage generation, so as to ensure that the firing judgment scale and the reset reference scale are consistent in the same time step. The preset static threshold refers to a threshold reference that is fixed in the model structure and does not change with time steps. Fusion refers to the operation rule of combining the threshold modulation amount and the static threshold to generate the dynamic threshold. The dynamic threshold refers to the threshold value determined by the static threshold and the threshold modulation amount at the current time step and the current neuron position. It changes with the input current in the time dimension and changes with the neuron position in the space dimension.

[0030]

[0031] wherein, represents the dynamic threshold, the value of which is obtained by fusing the static threshold and the threshold modulation amount and corresponds to the input state of the current time step; represents the preset static threshold, the value of which is used as a threshold reference and can be set as a constant 1 to stabilize the threshold level; represents the threshold modulation amount, the value of which is obtained by weighting modulation of the threshold adjustment parameter to the threshold modulation intermediate value and is used to realize adaptive deviation of the threshold S120, on the basis of the dynamic threshold, a pulse state variable is introduced to determine whether a pulse event occurs in the neuron at the current time step, and a learnable alternative gradient bound to the pulse state variable is constructed according to the characteristics of the pulse state variable at the pulse triggering position.

[0032] In a possible implementation, on the basis of the dynamic threshold, a pulse state variable is introduced to determine whether a neuron generates a pulse event at a current time step, and a learnable surrogate gradient bound to the pulse state variable is constructed for the characteristics of the pulse state variable at the pulse trigger position, specifically including: taking the dynamic threshold as a unified threshold constraint at the current time step, and performing a multi-value firing function calculation based on the comparison relationship between the membrane potential and the dynamic threshold to obtain a multi-value pulse output of the neuron at the current time step; introducing a pulse state variable corresponding to the multi-value pulse output, identifying the non-differentiable characteristics introduced by the sign determination of the pulse state variable at the pulse trigger position; constructing a learnable surrogate gradient bound to the pulse state variable for the non-differentiable characteristics; configuring the learnable surrogate gradient to keep the event determination semantics of the pulse state variable unchanged in the forward propagation process, and continuously approximating the gradient of the pulse state variable in the backward propagation process to restore the gradient propagation path; by binding the learnable parameters in the learnable surrogate gradient to the pulse state variable, the learnable parameters are adaptively updated according to the demand of the sensitive interval of event determination of the loss function in the training process.

[0033] Specifically, in the current time step, the dynamic threshold is taken as a unified threshold constraint, which is directly reflected in the discrete mapping of the multi-value pulse output to the relative relationship between the membrane potential and the dynamic threshold. Specifically, first, the membrane potential is obtained and taken as a representativeness of the excitability at the current time step, and the dynamic threshold is obtained and taken as a threshold reference at the current time step, so that whether the membrane potential exceeds the threshold and the degree of exceeding the threshold are both quantified under the same threshold reference system. Subsequently, the core independent variable of the multi-value firing function is constructed based on the ratio relationship between the membrane potential and the dynamic threshold, so that the membrane potential is normalized in the threshold scale before discretization, thereby ensuring that the threshold changes of different neurons and different time steps will not destroy the comparability of the firing levels. Among them, the truncation in the multi-value firing function is used to limit the firing level within the preset maximum level, and the floor function is used to map the continuous value to the discrete integer level, and finally the multi-value pulse output of the neuron at the current time step is obtained, and the multi-value pulse output is taken as the common leading quantity for the subsequent pulse state variable calculation and feedback reset voltage calculation. The unified threshold constraint refers to that the firing determination, firing level and reset reference in the same time step are all driven by the same dynamic threshold. The membrane potential refers to the voltage variable representing the internal state of the neuron after leakage integration. The multi-value firing function refers to a discrete mapping function composed of truncation and floor function. The multi-value pulse output refers to the pulse amplitude expressed in integer level to represent the firing strength of a single time step. The expression of the multi-value firing and the dynamic threshold is:

[0034]

[0035] wherein, represents the first the multi-valued pulse output of the layer at time step t, which takes discrete integer values and is used to represent the firing level; the membrane potential of the l-th layer at time step t, which is determined by the leaky integral and the reset; the membrane potential of the l-th layer at time step t, which is determined by the leaky integral and the reset; the dynamic threshold of the layer at time step t, which is used as a unified threshold reference quantity to participate in the firing level determination; n represents the maximum firing value of the pulse; the composite mapping of first truncation and then rounding down, which limits the continuous response to (0, n) by truncation and obtains discrete integer levels by rounding down, so as to realize the discrete output of multi-valued firing.

[0036] After obtaining the multi-valued pulse output, the purpose of introducing the pulse state variable is to separate the event semantics of whether a pulse event occurs from the size of the firing level, so that the subsequent feedback modulation and reset logic can work stably at the event granularity without being affected by the ambiguity introduced by the multi-valued level change. Specifically, the pulse state variable is established in one-to-one correspondence with the multi-valued pulse output, and the multi-valued pulse output is mapped to a binary event indicator by symbolic judgment, so that the pulse state variable indicates that a pulse event occurs when the multi-valued pulse output is non-zero, and the pulse state variable indicates that a pulse event does not occur when the multi-valued pulse output is zero. Symbolic judgment refers to the discrete discrimination of the positive-zero relationship of numerical values and the output of event indication results. The pulse trigger position refers to the boundary point where the multi-valued pulse output jumps from zero to non-zero or from non-zero to zero, which corresponds to the segmented switching position of symbolic judgment. The non-differentiable characteristic refers to the nonexistence or zero of the gradient of symbolic judgment at the boundary point, which causes the interruption of the backpropagation link. The expression of the pulse state variable is:

[0037]

[0038] wherein, the pulse state variable of the l-th layer at time step t, which takes values used to indicate whether a pulse event occurs; the pulse state variable of the l-th layer at time step t, which takes values used to indicate whether a pulse event occurs; the multi-valued pulse output; the symbolic judgment function, which is used to map the multi-valued pulse output to an event indicator, so that the pulse event determination is consistent with the representation of the multi-valued firing level.

[0039] A learnable surrogate gradient bound to the spike state variable is constructed for the non-differentiable property, the core of which is to provide a trainable and controllable approximation channel for the gradient of the multi-valued spike output with respect to the spike state variable, so as to avoid the gradient interruption caused by the sign determination in backpropagation. Specifically, the learnable surrogate gradient is designed as a gradient proxy bound to the spike state variable, which substitutes the real gradient of the sign determination in backpropagation, and suppresses the gradient explosion or vanishing by imposing a value range constraint on the proxy, so that the training process can still converge stably under the condition of the discreteness introduced by the multi-valued firing. The learnable surrogate gradient refers to the gradient approximation form used for backpropagation, which contains learnable parameters and updates with training. The binding refers to the gradient approximation form that is specially served for the gradient propagation position of the spike state variable, and the corresponding relationship is fixed in the computational graph. The expression of the learnable surrogate gradient is:

[0040]

[0041] wherein, represents the gradient channel of the spike state variable with respect to the multi-valued spike output, which is used to provide gradient transmission in backpropagation; represents the learnable surrogate gradient parameter of the i-th layer, which is updated in the training process and used to adjust the intensity of the gradient channel; the value constraint of the learnable surrogate gradient parameter is:

[0042] wherein, The value range of is realized by interval constraint to ensure that the gradient channel remains bounded in backpropagation and improve the training stability, and the initial value is 0.5 to provide moderate gradient transmission capability in the early stage of training.

[0043] The implementation of the learnable surrogate gradient that keeps the event decision semantics unchanged in the forward propagation and provides continuous approximation gradients in the backward propagation is a decoupling configuration of the forward values and the backward gradients in the computational graph. Specifically, the forward propagation path strictly adopts the sign decision results of the spike state variable, so that the spike state variable always maintains the binary event indication semantics of whether a spike event occurs, and ensures that the subsequent branches such as feedback quantity calculation, cumulative feedback quantity update, and feedback reset voltage generation obtain consistent control signals at the event granularity. The backward propagation path does not follow the non-differentiable gradient of the sign decision, but provides a continuous and bounded gradient approximation through the learnable surrogate gradient parameters, so that the optimization signal of the loss function to the firing behavior can be transmitted back to the membrane potential related parameters, the dynamic threshold related parameters and the upstream synaptic parameters across the spike trigger position. The continuous approximation refers to the fact that the gradient is numerically continuous and controllable through the learnable surrogate gradient parameters. The gradient propagation path recovery refers to the fact that the available gradient channel is re-established at the sign decision, so that the time-unfolded backward propagation can cover the node where the spike state variable is located and keep the link intact.

[0044] By binding the learnable parameters in the learnable surrogate gradient to the spike state variable, adaptive update of the sensitive interval of the event decision during the training process is realized, which is specifically embodied in that the learnable surrogate gradient parameters are included in the optimization target of the loss function as trainable variables and are updated with the backward propagation. Since the spike state variable is located at a key coupling position of the multi-value pulse output to the feedback reset and energy consumption control, the optimization of the loss function to the classification error, the firing rate constraint or the energy consumption proxy will jointly act on the gradient demand at this position, thereby driving the learnable surrogate gradient parameters to adaptively adjust so as to increase the gradient strength when the firing and non-firing need to be distributed in a more sensitive interval, and to decrease the gradient strength when unstable oscillation needs to be suppressed. The sensitive interval refers to the region where the optimization demand of the loss function near the event decision boundary is stronger. The adaptive update refers to the fact that the learnable surrogate gradient parameters are automatically adjusted according to the backward propagation gradient in the training iteration without the need for manual setting of a fixed slope or a fixed window, so that the spike state variable has trainability while maintaining the event decision semantics and forms a consistent coupling optimization closed loop with the multi-value firing, the dynamic threshold and the feedback reset mechanism.

[0045] In S130, an instantaneous feedback quantity is generated based on the input current and the spike state variable, and by introducing a cumulative feedback quantity with time memory characteristics, the feedback quantities of the historical time steps are weighted and fused with the instantaneous feedback quantity of the current time step to form a cumulative feedback result for adjusting the reset strength of the membrane potential.

[0046] In one possible implementation, the instantaneous feedback quantity is generated based on the input current and the pulse state variable, and the cumulative feedback quantity with time memory characteristics is introduced to weight and fuse the feedback quantity of the historical time step and the instantaneous feedback quantity of the current time step to form the cumulative feedback result for adjusting the membrane potential reset strength, specifically including: taking the input current as a continuous state quantity representing the external driving strength of the current time step, and taking the pulse state variable as an event indication quantity representing whether a pulse event occurs in the current time step; performing conditional mapping processing on the input current based on the pulse state variable, so that the input current generates an instantaneous feedback quantity pointing to the pulse punishment direction when the pulse state variable indicates that a pulse event occurs, and generates an instantaneous feedback quantity pointing to the pulse excitation direction when the pulse state variable indicates that no pulse event occurs, so as to obtain an instantaneous feedback quantity consistent with the current input state and the pulse event state of the neuron; after generating the instantaneous feedback quantity, a cumulative feedback quantity bound to the neuron and recursively updated across time steps is introduced as a feedback state quantity with time memory characteristics; the cumulative feedback quantity of the last time step is subjected to a time decay weight to obtain a historical feedback component, and the historical feedback component is weighted and fused with the instantaneous feedback quantity to obtain an updated cumulative feedback quantity.

[0047] Specifically, when the input current is defined as a quantity representing the external driving strength of the neuron at the current time step, and the pulse state variable is defined as an event indication quantity representing whether a pulse event occurs in the neuron at the current time step, the input current is used to depict the continuous state difference of the same neuron at different time steps, and the pulse state variable is used to depict the discrete bifurcation of the occurrence of a pulse event and the non-occurrence of a pulse event within the same time step, so that the subsequent feedback construction has both a continuously adjustable amplitude basis and a discrete explicit event semantics. Among them, the input current is a tensor-valued number that can be updated neuron by neuron and time step by time step, and can be positive or negative to reflect the direction difference between excitatory drive and inhibitory drive, and the pulse state variable is a binary indication quantity that can be updated neuron by neuron and time step by time step, and its value of 0 represents no pulse event, and its value of 1 represents a pulse event, so as to form a binding structure of continuous state quantity and event indication quantity at the same neuron level to support conditional feedback mapping.

[0048] When performing conditional mapping processing on the input current based on the pulse state variable, in order to form an instantaneous feedback quantity pointing to the pulse punishment direction when a pulse event occurs, and an instantaneous feedback quantity pointing to the pulse excitation direction when no pulse event occurs, the input current needs to be normalized and compressed by a sigmoid function first, so as to be mapped to the interval (0, 1) to avoid the input current positive and negative and amplitude span interfering with the feedback direction determination, and an intensity factor is introduced to learn and adjust the feedback strength; the expression of the pulse feedback quantity is:

[0049] Among them, This represents the impulse-based feedback quantity of the l-th layer neuron at time step t. This represents the intensity factor, which is a learnable parameter used to adjust the overall amplitude of the feedback quantity and can be set by layer or by channel. Represents the pulse state variable and its value range is ,when hour So that the feedback quantity acts in the direction of punishment, when hour So that the feedback quantity acts in the direction of the excitation. Represents the sigmoid function and input current. Compressing to the (0,1) interval enables the differentiation of cases where both events generate impulses but have different input drive strengths, thereby ensuring that the instantaneous feedback quantity is consistent with the neuron's current input state and impulse event state.

[0050] After generating the instantaneous feedback quantity, when introducing the cumulative feedback quantity, which is bound to the neuron and recursively updated across time steps, the cumulative feedback quantity is defined as a feedback state quantity with time memory properties. This is used to recursively retain the feedback trend in the penalty or incentive direction from historical time steps, thus avoiding the reset adjustment jittering over time due to relying solely on the instantaneous feedback quantity of the current time step. The time memory property is reflected in the fact that the cumulative feedback quantity is not an equal-weighted sum of all historical feedback quantities, but rather a cumulative feedback quantity with a time decay weight applied, giving a higher proportion of feedback history closer to the current time step and a lower proportion of feedback history from older times. This time decay weight is configured as a learnable parameter to adaptively fit the temporal dependency strength of different layers and neurons during training. When applying the time decay weight to the cumulative feedback quantity of the previous time step and weighting and fusing it with the instantaneous feedback quantity to obtain the updated cumulative feedback quantity, the expression for the cumulative feedback quantity is:

[0051]

[0052] in, Indicates the first The cumulative feedback amount of layer neurons at time step t. This represents the cumulative feedback amount at the previous time step. This represents the time decay weight, which is a learnable parameter used to characterize the degree of memory of the accumulated feedback from the previous time step and to achieve recursive decay. This represents the pulse-based feedback quantity at the current time step; the principle of this recursive structure lies in... The historical cumulative feedback is retained through exponential decay, so that the cumulative feedback includes both the historical feedback trend and the current instantaneous feedback correction, thereby forming a stable feedback memory representation that can be used for subsequent adjustment of the membrane potential reset intensity.

[0053] S140, the cumulative feedback result and the soft reset voltage determined based on the dynamic threshold value jointly act on the membrane potential reset process to obtain a feedback reset voltage.

[0054] In a possible implementation, the cumulative feedback result and the soft reset voltage determined based on the dynamic threshold value jointly act on the membrane potential reset process to obtain a feedback reset voltage, and specifically includes: taking the cumulative feedback result as a feedback constraint term of the reset strength of the membrane potential, and determining the soft reset voltage corresponding to the dynamic threshold value in the current time step; performing normalization and amplitude shaping processing on the cumulative feedback result, so that the cumulative feedback result is constrained to a continuous bounded interval and the sign direction is kept unchanged, to obtain a normalized cumulative feedback result superimposed with the soft reset voltage in the numerical scale; performing strength modulation on the normalized cumulative feedback result by using a feedback strength factor bound to the neuron, to obtain a modulated feedback result; and superimposing and fusing the modulated feedback result and the soft reset voltage to generate the feedback reset voltage containing the basic reset information corresponding to the dynamic threshold value and the cumulative constraint information of the historical pulse behavior.

[0055] Specifically, when the cumulative feedback result is taken as the feedback constraint term of the reset strength of the membrane potential, the cumulative feedback result is used to fix the trend of the pulse event of the historical time step in a recursive form as the reset modulation basis of the current time step, so that the reset strength is no longer determined only by the threshold state of the current time step, but is also constrained by the historical firing behavior. Meanwhile, when the soft reset voltage corresponding to the dynamic threshold value is determined in the current time step, the soft reset voltage is not determined by a fixed constant, but is directly determined by the dynamic threshold value and the multi-value pulse output, so that the reset reference is consistent with the threshold reference system of the current time step, and it is ensured that the higher the multi-value firing level is, the greater the threshold deduction is, so that the reset amplitude and the firing strength maintain a monotonic consistent relationship in the same time step. The feedback constraint term refers to the historical behavior constraint quantity used to adjust the reset strength, the soft reset voltage refers to the threshold-related reset quantity deducted from the membrane potential when the pulse event occurs, the dynamic threshold value refers to the threshold reference quantity generated by the static threshold value and the input current modulation and independently changed at the neuron granularity, and the multi-value pulse output refers to the discrete integer firing level output by the multi-value firing function. The expression of the soft reset voltage is:

[0056]

[0057] wherein, Vreset l(t) represents the soft reset voltage of the lth layer at the time step t, which is used to provide the basic reset information consistent with the dynamic threshold value; Vth l(t) represents the dynamic threshold value of the lth layer at the time step t, which is used as a unified threshold reference quantity to determine the soft reset reference; represents the multi-valued pulse output of the l-th layer at time step t, which takes discrete integer values and is used to represent the firing level, so that the soft reset voltage is linearly amplified with the firing level, thereby realizing the consistent constraint that the stronger the firing, the stronger the basic reset.

[0058] When performing normalization and amplitude shaping processing on the cumulative feedback result, the normalization and amplitude shaping processing are used to map the cumulative feedback result from a possibly unbounded recursive accumulation state to a continuous bounded interval, so that it can be stably superimposed with the soft reset voltage in the numerical scale, and avoid the cumulative feedback result from causing the reset modulation term to dominate the reset process due to amplitude drift under long time step recursion. At the same time, the processing needs to keep the sign direction unchanged to ensure that the adjustment direction of the cumulative feedback result to the reset strength has consistent semantics in time evolution, so that the cumulative feedback result will not reverse direction due to numerical scaling when expressing the trend that reset inhibition should be strengthened or reset should be weakened to avoid dormancy. Among them, normalization refers to compressing the cumulative feedback result to a pre-set limited range for superposition operation, amplitude shaping refers to suppressing the impact of extreme values on reset modulation through smooth bounded mapping, and continuous bounded interval refers to mapping output that is continuous in numerical value and has upper and lower bounds to ensure training and inference stability. The expression of the normalized cumulative feedback result is:

[0059]

[0060] wherein, represents the normalized cumulative feedback result of the l-th layer at time step t, which takes values in a continuous bounded interval and is used for subsequent reset modulation; represents the cumulative feedback result of the l-th layer at time step t, which is obtained by recursively fusing the historical cumulative feedback result and the current time step feedback quantity; represents the sigmoid function, which maps the cumulative feedback result to the interval (0, 1) through bounded compression, thereby realizing amplitude shaping and suppressing amplitude drift caused by recursive accumulation.

[0061] When intensity modulation is performed on the normalized cumulative feedback result by using the feedback intensity factor bound to the neuron, the feedback intensity factor is used to control the effective intensity of the cumulative feedback result participating in the reset modulation, so that different layers or different neurons can form differentiated reset feedback sensitivity. In the feedback recursive structure of the document, the feedback intensity factor is fixed as a learnable parameter through the generation process of the cumulative feedback result, so that it is updated adaptively according to the comprehensive demand of the firing rate, stability and accuracy of the loss function in the training process, so as to embed the control quantity of how strong the reset modulation needs into the evolution link of the cumulative feedback result itself, and complete the intensity modulation before entering the reset modulation link. At the same time, in order to ensure that the reset modulation only works when the pulse event occurs, the normalized cumulative feedback result is event-gated by using the pulse state variable, so that the normalized cumulative feedback result only participates in the reset modulation when the pulse state variable indicates that the pulse event occurs, thereby avoiding unnecessary reset disturbance to the membrane potential in the unfired state. Among them, the feedback intensity factor refers to the learnable parameter for adjusting the generation intensity of the cumulative feedback result, the intensity modulation refers to changing the contribution ratio of feedback to reset through the learnable parameter, and the event gating refers to using the pulse state variable to control whether the feedback modulation item takes effect. The recursive expression of the cumulative feedback result is:

[0062]

[0063] wherein, represents the cumulative feedback result of the lth layer at time step t; represents a memory factor, whose value range is (0, 1) and is used to apply a time memory weight to the cumulative feedback result of the last time step, so that the historical feedback trend is preserved in a recursive manner; represents the cumulative feedback result of the last time step; represents a feedback intensity factor, which is used to adjust the action intensity of the input current in feedback generation; represents the input current of the lth layer at time step t; is used for bounded compression of the input current to stabilize the instantaneous feedback amplitude; represents a pulse state variable, whose value is and the pulse event is mapped to the feedback direction by so that the pulse event and the non-pulse event correspond to different directions of the feedback trend.

[0064] When the modulation feedback result is superimposed with the soft reset voltage to generate the feedback reset voltage, the superimposition is used to integrate the dynamic threshold corresponding basic reset information and the historical pulse behavior cumulative constraint information into a unified reset control quantity in the same time step, so that the membrane potential reset not only inherits the residual membrane potential reservation characteristics of the soft reset, but also introduces the inhibition ability of the cumulative feedback result to the abnormal excessive activation. In this integration structure, the soft reset voltage provides a basic deduction item with consistent threshold scale, the normalized cumulative feedback result provides a bounded reset modulation item, and the pulse state variable is used to gate the reset modulation item, so that the reset modulation item only participates in superposition when the pulse event occurs, thereby ensuring that the reset modulation semantics is consistent with the pulse event semantics. Among them, the feedback reset voltage refers to the reset voltage containing the basic reset information and the historical behavior constraint information, and the superimposition refers to the additive combination of the soft reset voltage and the modulation feedback result to form a single reset control quantity. The expression of the feedback reset voltage is:

[0065]

[0066] wherein, represents the feedback reset voltage of the lth layer at time step t; represents the soft reset voltage determined by the dynamic threshold and the multi-value pulse output, and is used to provide the basic reset information; represents the normalized cumulative feedback result, and is used to provide the bounded reset modulation information; represents the pulse state variable, and is used to event gate the reset modulation item, so that the reset modulation item only takes effect when the pulse event occurs, thereby making the feedback reset voltage be numerically constrained by the current threshold state and the historical pulse behavior at the same time and used for subsequent membrane potential reset process.

[0067] S150, after the membrane potential is updated by the leak integration and the feedback reset voltage is applied, a multi-value firing function composed of a truncation function and a floor function is used to map the relationship between the membrane potential and the dynamic threshold to a discrete multi-value pulse output.

[0068] In a possible implementation, after the membrane potential is updated by leaky integration and a feedback reset voltage is applied, a multi-value firing function composed of a clipping function and a floor function is used to map the relationship between the membrane potential and the dynamic threshold to a discrete multi-value pulse output, specifically including: performing leaky integration update on the membrane potential in the current time step, and applying a feedback reset voltage to the updated membrane potential when a pulse event occurs, to obtain an effective membrane potential representing the effective state of the neuron in the current time step; taking the dynamic threshold as a unified firing decision scale, and constructing a continuous firing response quantity based on the proportional relationship between the effective membrane potential and the dynamic threshold; introducing the continuous firing response quantity into the clipping function to perform amplitude limiting processing, so as to constrain the continuous firing response quantity in a preset multi-value firing range, to obtain a clipped firing response quantity; introducing the clipped firing response quantity into the floor function to perform discretization mapping, to obtain a multi-value pulse output in the form of a discrete integer representing the firing intensity of the neuron in the current time step, wherein the numerical value of the multi-value pulse output is used to represent the superthreshold degree of the effective membrane potential relative to the dynamic threshold.

[0069] Specifically, when leaky integration update is performed on the membrane potential in the current time step, the reset membrane potential of the previous time step is first obtained, and then decayed by a leak factor to reflect the natural fall of the membrane potential over time, and then the input current of the current time step is superimposed to reflect the charging effect of the external drive on the membrane potential, so as to obtain the updated membrane potential. The leaky integration update is used to unify the historical state memory and the current drive response into the same membrane potential state, the reset membrane potential is used to accept the effective state after the reset operation of the previous time step, the leak factor is used to control the retention degree of the historical state in the current time step, and the input current is used to represent the external drive intensity obtained by synaptic weighting convergence. When a pulse event occurs, a feedback reset voltage is applied to the updated membrane potential, and the effective membrane potential is obtained by performing a threshold-related and history-feedback-constrained deduction operation on the membrane potential, so that the effective membrane potential becomes the final membrane potential state before entering the multi-value firing decision, thereby ensuring that the state transition of charging first, resetting second, and firing third in the same time step is consistent and can be propagated to the next time step. The expression of the membrane potential update is:

[0070]

[0071] wherein, represents the updated membrane potential of the lth layer at time step t, which is used as the voltage state for subsequent reset and firing calculation; represents a leak factor, the value of which is used to control the influence of the membrane potential of the previous time step on the membrane potential of the current time step, and the greater the value, the stronger the retention of the historical state; represents the reset membrane potential of the lth layer at time step t-1, which is used to accept the effective state after the reset operation of the previous time step; represents the input current of the l-th layer at time step t, which is used to depict the charging contribution of the external drive to the membrane potential at the current time step. The expression of the membrane potential reset is:

[0072]

[0073] wherein, represents the effective membrane potential of the l-th layer at time step t, which is the voltage state after applying the feedback reset voltage on the basis of the updated membrane potential; represents the updated membrane potential; represents the feedback reset voltage, which contains both the basic reset information corresponding to the dynamic threshold and the historical pulse behavior constraint information corresponding to the cumulative feedback result, and is used to adjust the reset strength and suppress the risk of abnormal over-activation and long-term dormancy.

[0074] When the dynamic threshold is taken as a unified firing decision scale, the dynamic threshold simultaneously assumes the roles of firing level decision reference quantity and voltage normalization reference quantity within the current time step, so that the membrane potential states of different neurons at different time steps can be comparably mapped under the same threshold reference system. In specific implementation, the effective membrane potential and the dynamic threshold are first obtained, and a continuous firing response quantity is constructed through the proportional relationship between the two, so that the effective membrane potential is normalized to a relative scale space with the dynamic threshold as the reference, so that the temporal and spatial independent changes of the dynamic threshold will not destroy the relative meaning of the firing level. Among them, the continuous firing response quantity is used to retain the continuous information of the superthreshold degree before discretization, the dynamic threshold is used to provide adaptive threshold reference at the neuron granularity, and the proportional relationship is used to convert the voltage state into a dimensionless relative superthreshold strength representation, so as to realize multi-level discrete firing through subsequent truncation and downward rounding. The expression of the continuous firing response quantity is:

[0075]

[0076] wherein, represents the continuous firing response quantity of the l-th layer at time step t, which is used to represent the relative strength of the effective membrane potential with respect to the dynamic threshold; represents the effective membrane potential; represents the effective membrane potential; represents the dynamic threshold, which is generated by the static threshold and the input current modulation and independently changes at the neuron granularity, and is used to provide a unified firing decision scale.

[0077] When the continuous firing response quantity is introduced into the clipping function to perform the amplitude limiting processing, the clipping function is used to limit the continuous firing response quantity in a preset multi-value firing range, so as to avoid that the abnormal fluctuation of the effective membrane potential causes the firing level to exceed the boundary and triggers the instability of training and reasoning. In a specific implementation, the lower limit of the continuous firing response quantity is limited to 0 to avoid generating a negative level firing, the upper limit of the continuous firing response quantity is limited to a maximum pulse firing value to ensure that the multi-value pulse output is represented in an integer coding range, and the result after clipping is used as a clipped firing response quantity for subsequent integer discrete. The multi-value firing range refers to a continuous interval corresponding to a set of allowed integer firing levels, the maximum pulse firing value refers to the highest integer level that can be reached by the multi-value pulse output, and the clipped firing response quantity refers to the continuous response quantity after the upper and lower limit constraints, which is still a continuous value but meets the level range requirement of subsequent discretization. The expression of the clipped firing response quantity is:

[0078]

[0079] wherein, represents the clipped firing response quantity of the lth layer at the time step t; represents the continuous firing response quantity; n represents the maximum pulse firing value, which is determined by the upper limit of the multi-value coding level set by the network; represents the clipping function, which limits the amplitude of the continuous firing response quantity by the lower limit 0 and the upper limit n, so that the output falls within the preset multi-value firing range.

[0080] When the clipped firing response quantity is introduced into the floor function to perform the discrete mapping, the floor function is used to map the clipped firing response quantity from the continuous domain to the integer domain, so as to obtain the multi-value pulse output in the form of discrete integers, and make the multi-value pulse output strictly correspond to the superthreshold degree of the effective membrane potential with respect to the dynamic threshold in the integer level. In a specific implementation, the clipped firing response quantity is located in the interval [0, n], which is mapped to an integer in by the floor function, so that when the effective membrane potential just exceeds the dynamic threshold, a smaller integer level is output, and when the effective membrane potential significantly exceeds the dynamic threshold, a larger integer level is output, so that richer spatial information is expressed in the form of integer amplitude in a single time step; wherein the discrete mapping refers to the quantization process from continuous to integer, the multi-value pulse output refers to the integer firing level of the neuron at the current time step, and the superthreshold degree refers to the relative deviation intensity between the effective membrane potential and the dynamic threshold, which is stably mapped to the integer level by a proportional relationship and the clipped and discrete again. The expression of the multi-value pulse output is:

[0081]

[0082] wherein, denotes the multi-valued pulse output of the l-th layer at time step t, which takes discrete integer values and is used to represent the firing intensity level; denotes the truncated firing response quantity; denotes the floor function, which realizes the continuous-to-discrete mapping by taking the maximum integer not greater than the input, so that the numerical value of the multi-valued pulse output monotonically represents the superthreshold degree of the effective membrane potential relative to the dynamic threshold.

[0083] S160, based on the membrane potential update equation, the dynamic threshold, the multi-valued firing function and the feedback reset voltage, a discrete dynamic model of the dynamic threshold feedback reset leaky integral multi-valued firing neuron is constructed.

[0084] In constructing the discrete dynamic model of the dynamic threshold feedback reset leaky integral multi-valued firing neuron, the dynamic threshold generation, the membrane potential leaky integral update, the multi-valued firing determination, the pulse event indication, the cumulative feedback recursion, the feedback reset voltage generation and the reset membrane potential update in the same time step are solidified into a closed-loop recursion structure, so that the reset membrane potential of the last time step and the cumulative feedback result jointly constrain the membrane potential evolution of the current time step, and at the same time, the input current of the current time step adjusts the firing and reset through the dynamic threshold and the feedback link, thereby forming a trainable, inferable and recursive neuron dynamic behavior on the discrete time axis.

[0085] At the beginning of each time step, the dynamic threshold is first generated based on the input current, so that the dynamic threshold changes with the input current at the neuron granularity and serves as a unified threshold reference quantity throughout the firing determination and reset reference generation of the current time step, and the expression of the dynamic threshold is:

[0086] wherein, denotes the dynamic threshold of the l-th layer at time step t, which takes values that change with the input current and are independently formed at the neuron granularity; denotes the static threshold, which is used to provide a threshold reference and is fixed at 1 in implementation; denotes the threshold adjustment parameter, whose value range is limited to (-1, 1) to constrain the threshold modulation amplitude and is updated adaptively in training; denotes the input current of the l-th layer at time step t, which takes values obtained by converging the pulse of the last layer through the synaptic weight and corresponds to the neuron spatial position one by one; denotes the hyperbolic tangent function, which is used to map the input current to a continuous bounded interval and maintain the sign direction, so that the dynamic threshold maintains a smooth response to input changes and avoids threshold oscillation.

[0087] After obtaining the dynamic threshold, a leaky integration update is performed on the membrane potential to form an updated membrane potential at the current time step, which retains the decay memory of the reset membrane potential at the previous time step and superimposes the driving contribution of the input current at the current time step, and the expression of the membrane potential update equation is:

[0088] wherein, represents the updated membrane potential of the lth layer at time step t, which is used as the voltage state of firing and resetting; represents the membrane time constant, which is used to control the retention degree of the state at the previous time step at the current time step; represents the reset membrane potential of the lth layer at time step t-1, which is used to undertake the effective state after resetting at the previous time step and as a state quantity for cross-time step recursion; represents the input current, which is used to represent the external driving strength at the current time step and directly participates in the charging of the membrane potential.

[0089] After obtaining the updated membrane potential, a multi-value pulse output is generated based on the dynamic threshold and the multi-value firing function, so that the super-threshold degree of the membrane potential with respect to the dynamic threshold is mapped to a discrete integer level, so that higher density of firing information is expressed in a single time step in the multi-value pulse output, and the expression of the multi-value pulse output is:

[0090] wherein, represents the multi-value pulse output of the lth layer at time step t, which is a discrete integer and is used to represent the firing strength level; represents a continuous response quantity obtained by normalizing the membrane potential in the dynamic threshold scale, so that the firing levels at different threshold states remain comparable; n represents the maximum firing value of the pulse, which is determined by the upper limit of the multi-value coding set by the network; represents a composite mapping composed of truncation and downward rounding, which first limits the continuous response quantity to [0, n] to prevent level overflow, and then obtains a discrete integer level through downward rounding, so as to realize the discretization of multi-value firing.

[0091] After obtaining the multi-value pulse output, a pulse state variable is introduced to provide an event indication quantity of whether a pulse event occurs, so that the subsequent feedback link and the reset link obtain consistent gating signals in the event granularity, and identify the problem that the sign determination is not differentiable at the pulse trigger position, causing the interruption of the backpropagation gradient, and the expression of the pulse state variable is:

[0092] wherein, ​denotes the pulse state variable of the l-th layer at time step t, which takes value 0 to represent no pulse event and takes value 1 to represent a pulse event; denotes the multi-valued pulse output; denotes the sign determination function, which is used to map the multi-valued pulse output to the event indicator, thereby decoupling the event semantics from the multi-valued level semantics and being used to gate the subsequent reset modulation term.

[0093] In the training phase, in order to enable the pulse state variable to participate in back propagation, the gradient of the pulse state variable in back propagation is provided with a continuous approximation by a learnable surrogate gradient, the forward propagation keeps the event semantics of the sign determination unchanged, and the back propagation restores the available gradient channel, so that the threshold adjustment parameter, the leakage factor, the feedback intensity factor and other parameters can obtain stable updates through time-expanding back propagation. The expression of the learnable surrogate gradient is:

[0094] wherein, denotes the gradient channel of the pulse state variable with respect to the multi-valued pulse output; denotes the learnable surrogate gradient parameter, which is adaptively updated in the training process and is used to adjust the gradient intensity, so that the optimization requirement of the loss function on the event determination boundary can pass through the non-differentiable position; the value range of the learnable surrogate gradient parameter is:

[0095] wherein, the value range is used to ensure that the gradient channel is bounded and improve the training stability, and by initializing to 0.5, the training has moderate gradient transmission ability at the initial stage.

[0096] After generating the pulse state variable, a feedback link is constructed to jointly map the input current and the pulse event state into an instantaneous feedback quantity, and time memory is introduced through cumulative feedback recursion, so that the reset adjustment is simultaneously constrained by the historical pulse behavior trend and the current time step input drive. The expression of the pulse-based feedback quantity is:

[0097] wherein, denotes the pulse-based feedback quantity of the l-th layer at time step t, which is used to represent the feedback direction and the feedback intensity at the current time step; denotes the feedback intensity factor, which is a learnable parameter and is used to adjust the feedback amplitude; participates in the sign mapping as an event direction selection quantity in the feedback mapping, so that numerically forms a selection result of the punishment direction or the incentive direction, thereby mapping the pulse event state into the feedback direction; sigmoid function, which is used to bound the input current to stabilize the feedback amplitude and make different input drive strengths distinguishable in the feedback strength; represents the input current and is used to characterize the external drive strength at the current time step.

[0098] On this basis, the cumulative feedback result is introduced as a feedback state variable with time memory characteristics, and the cumulative feedback result at the previous time step is time-weighted through a memory factor and fused with the feedback at the current time step to form a cumulative feedback result that can be used for reset modulation, so that the feedback adjustment to the reset strength has recursive stability, and the expression of the cumulative feedback result is:

[0099] wherein, represents the cumulative feedback result at time step t; represents the cumulative feedback result at the previous time step; represents the memory factor, which has a value range of (0, 1) and is used to control the retention degree of the historical cumulative feedback result at the current time step, so that the feedback memory is recursively propagated over time and can be updated adaptively in training; represents the feedback strength factor and is used to adjust the contribution strength of the input drive in feedback generation; is used to compress the input current to a bounded interval to stabilize the instantaneous feedback component; represents the pulse state variable and is determined by determines the feedback direction, so that the occurrence of a pulse event and the non-occurrence of a pulse event correspond to different feedback directions, thereby injecting event semantics into the cumulative feedback recursion.

[0100] After obtaining the cumulative feedback result, it is used together with the soft reset voltage determined based on the dynamic threshold to reset the membrane potential to generate a feedback reset voltage, so that the reset control quantity contains both the basic reset information corresponding to the dynamic threshold and the cumulative constraint information of the historical pulse behavior, and the expression of the feedback reset voltage is:

[0101] wherein, represents the feedback reset voltage of the lth layer at time step t; represents the soft reset voltage determined by the dynamic threshold and the multi-value pulse output, which is used to provide basic reset information and make the higher the firing level, the stronger the basic reset; represents the normalized cumulative feedback result obtained by bounded mapping of the cumulative feedback result, which is used to provide reset modulation information and suppress the amplitude drift caused by the cumulative feedback recursion; is used for event gating of the reset modulation term, so that the reset modulation term only takes effect when a pulse event occurs, thereby ensuring that the reset modulation semantics are consistent with the pulse event semantics.

[0102] Finally, the feedback reset voltage is applied to the updated membrane potential to obtain the reset membrane potential, and the reset membrane potential is taken as the state carrying quantity of the next time step membrane potential update equation, thereby closing the discrete dynamics recursion loop and making the dynamic threshold-multiple value firing-feedback accumulation-feedback reset-state carrying form a consistent association in time evolution, and the expression of the reset membrane potential is:

[0103] wherein, represents the reset membrane potential of the lth layer at time step t, which enters the next time step membrane potential update as a state quantity across time step recursion; represents the updated membrane potential; represents the feedback reset voltage and is used to perform a threshold-related and history feedback-constrained deduction on the membrane potential after the occurrence of a pulse event, so that the neuron can simultaneously suppress abnormal excessive activation and reduce the risk of long-term dormancy during time evolution, thereby forming a stable and complex encoding-capable discrete dynamics behavior.

[0104] S170, in the training phase, for multiple implicit thresholds corresponding to the multiple value firing neuron, build a substitute gradient function associated with the dynamic threshold, in the inference phase, convert the multiple value pulse output into equivalent binary pulse representation, and reduce bit level redundancy through integer to binary conversion method, so that the calculation process based on multiple value pulse is remapped to event driven operation mainly based on sparse accumulation.

[0105] In a possible implementation, after constructing the discrete dynamic model of the dynamic threshold feedback reset leaky integrate-and-fire neuron based on the membrane potential update equation, the dynamic threshold, the multi-value firing function, and the feedback reset voltage, the method further includes: in the training phase, identifying, based on the discrete mapping characteristics of the multi-value firing function, that the dynamic threshold constraint corresponds to multiple implicit threshold positions associated with firing levels; constructing, for the implicit threshold positions, a substitute gradient function that is explicitly associated with the dynamic threshold, so that the substitute gradient function maintains the discrete firing decision semantics of the multi-value firing function unchanged in the forward propagation process; configuring the substitute gradient function to provide a gradient response within an effective interval determined by the dynamic threshold; setting a gradient adjustment parameter in the substitute gradient function as a learnable parameter, so that the gradient adjustment parameter is updated adaptively in the training process according to the demand of the loss function for the multi-value pulse sensitive region; in the inference phase, performing equivalent conversion processing on the multi-value pulse output to map the multi-value pulse output to a set of binary pulses that are consistent in the accumulation result, so that the amplitude information of the multi-value pulse is split into multiple binary pulse events; introducing an integer-to-binary conversion mode to map the integer amplitude representation corresponding to the multi-value pulse output to a binary bit-level representation; constructing a parallel binary pulse channel based on the binary bit-level representation; and in the neuron or synapse calculation process, performing an accumulation operation on the pulse event occurrence position based on the binary pulse channel, so that the original calculation process based on the multi-value pulse is remapped to event-driven operation.

[0106] Specifically, when identifying the implicit threshold positions using the discrete mapping characteristics of the multi-value firing function in the training phase, the dynamic threshold is regarded as a unified firing decision scale, and each discrete firing level of the multi-value pulse output is regarded as a counting result of the effective membrane potential crossing different level boundaries. Since the multi-value firing function maps the continuous voltage state to the integer level through proportional normalization, truncation, and rounding down, when the multi-value pulse output jumps from level 0 to level 1, from level 1 to level 2, and so on to the maximum level, there are multiple sets of trigger boundaries corresponding to the effective membrane potential in the voltage domain, and these trigger boundaries are in an integer multiple relationship with the dynamic threshold in the voltage domain, thereby forming multiple implicit threshold positions associated with the firing levels one by one, and making the multi-value discrete expression equivalent to the structural feature of multi-threshold triggering in the training. The implicit threshold positions are used for the substitute gradient function to locate the gradient sensitive region in the back propagation and avoid treating the multi-value discrete jump as a single threshold jump.

[0107] When constructing a surrogate gradient function explicitly associated with the dynamic threshold for the implicit threshold position, the surrogate gradient function is used to provide a continuous gradient channel for the non-differentiable transition of the multi-valued firing function during backpropagation, while still maintaining the discrete decision semantics of the multi-valued firing function on the relationship between the effective membrane potential and the dynamic threshold during forward propagation, so that the training process does not change the event firing behavior of the neuron but only changes the gradient propagation method. The surrogate gradient function regards each implicit threshold position as a local gradient window center and regards the dynamic threshold as a scale reference for these window centers, so that the gradient window is synchronized with the change of the dynamic threshold, thereby ensuring that when the dynamic threshold presents spatial and temporal independent changes at the neuron granularity, the gradient sensitive region can still align with the real level boundary and maintain the training stability. The expression of the surrogate gradient function is:

[0108]

[0109] wherein, represents the result of approximating the gradient channel of the multi-valued pulse output to the firing decision internal variable by the surrogate gradient, which is used for backpropagation; represents the multi-valued pulse output of the lth layer at time step t; represents the updated membrane potential, which is used as the voltage state of the multi-valued firing; represents the abstract representation of the discrete mapping process formed by normalization, truncation and down rounding according to the dynamic threshold, which is used to bind the surrogate gradient to the discrete mapping in mathematics; b represents the implicit threshold number, which takes a value from 1 to n and corresponds to each firing level boundary; n represents the maximum firing value of the pulse; represents the threshold reference quantity, which is used as the scale reference of each implicit threshold center in the surrogate gradient construction, so that the implicit threshold position can be represented as ; a represents the gradient window width parameter, which is used to determine the size of the voltage neighborhood that allows the generation of gradient response near each implicit threshold; is used to map the inequality decision to an indication result, so that when falls into the window centered at with a width of a, the output is 1, otherwise the output is 0, thereby limiting the gradient response within the effective interval; the symbol is used to express the joint action of multiple window constraints on the gradient response, so that the multiple implicit threshold positions corresponding to the multi-valued firing jointly determine the available gradient area.

[0110] When the surrogate gradient function is configured to provide a gradient response within the effective interval determined by the dynamic threshold, the effective interval is defined as a window centered at each implicit threshold position a local window interval of the partial derivative of the threshold reference quantity, and the center of the window interval moves synchronously with the threshold reference quantity, so that the gradient response is always concentrated in the voltage neighborhood where the level transition may occur instead of spreading in the saturation region far from the threshold, thereby avoiding the problem of gradient interval expansion caused by multi-value firing. In implementation, the inequality judgment is used to construct the window indication and force the gradient outside the window to be 0, so that the back propagation only passes through the window The gradient response of a fixed amplitude is provided, thereby converting the training difficulty of multi-value discrete transition into local training of multiple small windows and maintaining the boundedness and sparsity of the gradient.

[0111] When the gradient adjustment parameter in the surrogate gradient function is set as a learnable parameter, the gradient adjustment parameter is used to adjust the effective range of the gradient window or the gradient response strength, so that the training process can adaptively change which threshold boundary needs to be finely optimized and how strong the gradient near the boundary should be according to the needs of the loss function for the sensitive area of the multi-value pulse. In combination with the explicit dependence of the multi-value firing function on the dynamic threshold, the surrogate gradient further transmits the gradient from the multi-value pulse output to the updated membrane potential and the dynamic threshold through the chain relationship in the back propagation, so that the dynamic threshold related parameters can be stably optimized under the multi-threshold structure. The chain relationship is explicitly derived in the form of the voltage state and the threshold reference quantity in the document, thereby ensuring that the multi-value firing, the dynamic threshold and the surrogate gradient form a consistent coupling update path in the training graph. The expression of the gradient relationship of the multi-value pulse output to the updated membrane potential is:

[0112]

[0113] wherein, represents the gradient of the multi-value pulse output to the updated membrane potential, which is used to transmit the optimization signal of the loss function to the membrane potential evolution related parameters; represents the local gradient response given by the surrogate gradient.

[0114] The expression of the gradient relationship of the multi-value pulse output to the threshold reference quantity is:

[0115] In the inference phase, when performing equivalent conversion processing on the multi-value pulse output, the integer amplitude of the multi-value pulse output is interpreted as the superposition result of multiple unit pulse events in the same time step, and a binary pulse set is constructed to make the binary pulse set consistent with the original multi-value pulse output in the cumulative sense, thereby splitting the integer amplitude information into multiple binary pulse events. In implementation, for each neuron position, the case of multi-value pulse output k is mapped to a binary pulse set of length n, where the first k elements are set to 1 and the remaining elements are set to 0, so that the bit-by-bit cumulative result of the binary pulse set is equal to k. Thus, the multi-value drive is converted into multiple binary drives without changing the downstream cumulative result, and the subsequent synapse calculation can restore the event-driven sparse accumulation characteristic with binary pulse events as the basic unit.

[0116] When introducing the conversion mode from integer to binary, the integer amplitude representation corresponding to the multi-value pulse output is directly mapped to the binary bit-level representation, so that the binary pulse set no longer uses the redundant representation of padding by unit pulses, but uses the compact representation of binary bit decomposition, thereby reducing bit-level redundancy and reducing the number of subsequent parallel channels. In implementation, for each neuron position, the integer amplitude is converted into several binary bits, and whether each bit is 1 or not is used as an indication of the binary pulse event corresponding to the bit, so that the binary bit-level representation is numerically equivalent to the integer amplitude in terms of reconfigurability, and provides conditions for subsequent incorporation of weight constants into the convolution kernel or into the synapse mapping, thereby avoiding the introduction of additional high-overhead multiplication operations during inference.

[0117] Based on the binary bit-level representation, a parallel binary pulse channel is constructed, and the binary event matrix of each bit is used as an independent channel, and the spatial size and time alignment relationship consistent with the original pulse matrix are maintained at the channel level, so that each channel can directly participate in synapse mapping or convolution mapping. In implementation, the number of binary pulse channels is determined by the number of binary bits, and each channel corresponds to a fixed bit weight constant, so that the channel output can recover the equivalent integer drive contribution when numerically reconstructed through the bit weight constant. At the same time, since the bit weight constant is a fixed constant in the inference phase, the bit weight constant can be re-parameterized into the convolution kernel or the synapse weight to avoid explicit multiplication, so that the parallel binary pulse channel not only maintains computational equivalence but also reduces computational overhead as much as possible.

[0118] ​In the process of neuron or synapse calculation, the accumulation operation on the position of pulse event is performed based on the binary pulse channel. Each binary pulse channel triggers accumulation only at the element where the event position is 1, thereby re-mapping the product accumulation process originally requiring the participation of integer amplitude to event-driven operation mainly based on sparse accumulation. In implementation, synapse mapping is performed on each binary pulse channel respectively, and channel-level accumulation is performed on the mapping result, so that the final accumulation result is consistent in value with the calculation result directly using the multi-value pulse output for integer-driven calculation. At the same time, due to the sparsity of the binary pulse channel, a large number of positions do not need to trigger accumulation, thereby reducing the number of synapse operations and improving the energy efficiency in the reasoning stage.

[0119] The embodiment also discloses a multi-value pulse neuron processing device based on dynamic threshold value feedback reset, referring to Figure 2 , comprising an acquisition module 201, a processing module 202 and an output module 203, and the device is used for performing any one of the dynamic threshold value feedback reset based multi-value pulse neuron processing methods described above, wherein: The acquisition module 201 is used for performing nonlinear mapping on a preset static threshold value based on the input current received by the neuron in each time step, and generating a dynamic threshold value corresponding to the current input state of the neuron in combination with a threshold adjustment parameter; The processing module 202 is used for introducing a pulse state variable to determine whether a pulse event occurs in the neuron in the current time step on the basis of the dynamic threshold value, and constructing a learnable substitute gradient bound to the pulse state variable in view of the characteristics of the pulse state variable at the pulse trigger position; The processing module 202 is used for generating an instantaneous feedback quantity based on the input current and the pulse state variable, and fusing the feedback quantity of the historical time step and the instantaneous feedback quantity of the current time step by introducing a cumulative feedback quantity with time memory characteristics, so as to form a cumulative feedback result used for adjusting the reset strength of the membrane potential; The processing module 202 is used for applying the cumulative feedback result and a soft reset voltage determined based on the dynamic threshold value to the membrane potential reset process to obtain a feedback reset voltage; The processing module 202 is used for mapping the relationship between the membrane potential and the dynamic threshold value into a discrete multi-value pulse output by using a multi-value firing function composed of a truncation function and a floor function after the membrane potential is updated by leaky integration and the feedback reset voltage is applied; The processing module 202 is used for constructing a discrete dynamic model of the dynamic threshold value feedback reset leaky integration multi-value firing neuron based on the membrane potential update equation, the dynamic threshold value, the multi-value firing function and the feedback reset voltage; The output module 203 is configured to, in the training phase, construct a substitute gradient function associated with the dynamic threshold for the multiple implicit thresholds corresponding to the multi-value firing neuron, in the inference phase, convert the multi-value pulse output into an equivalent binary pulse representation, and reduce bit-level redundancy through integer-to-binary conversion, so as to remap the multi-value pulse-based calculation process into an event-driven operation mainly based on sparse accumulation.

[0120] It should be noted that: the device provided in the above embodiment is only exemplified by the above division of functional modules when realizing its functions. In actual application, the above functions can be completed by different functional modules according to needs, that is, the internal structure of the device is divided into different functional modules to complete all or part of the above-described functions. In addition, the device and method embodiments provided in the above embodiments belong to the same concept, and the specific implementation process is detailed in the method embodiments, which will not be described here.

[0121] The embodiment further discloses an electronic device, which refers to Figure 3 The electronic device can include at least one processor 301, at least one communication bus 302, a user interface 303, a network interface 304, and at least one memory 305.

[0122] The communication bus 302 is configured to realize the connection and communication between the components.

[0123] The user interface 303 can include a display screen (Display) and a camera (Camera). Optionally, the user interface 303 can further include a standard wired interface and a wireless interface.

[0124] The network interface 304 can optionally include a standard wired interface and a wireless interface (such as a WI-FI interface).

[0125] The processor 301 can include one or more processing cores. The processor 301 connects various parts within the server through various interfaces and lines, performs various functions of the server and processes data by running or executing instructions, programs, code sets or instruction sets stored in the memory 305, and calling data stored in the memory 305. Alternatively, the processor 301 can be implemented in at least one of a hardware form of a digital signal processing (DSP), a field-programmable gate array (FPGA), and a programmable logic array (PLA). The processor 301 can integrate a combination of one or more of a central processing unit (CPU), a graphics processing unit (GPU), and a modem. Among them, the CPU mainly processes operating systems, user interfaces, and application programs; the GPU is responsible for rendering and drawing the content to be displayed on the display screen; and the modem is used for processing wireless communication. It can be understood that the above-mentioned modem can also not be integrated into the processor 301, but can be realized by a separate chip.

[0126] The memory 305 can include a random access memory (RAM) and a read-only memory (ROM). Optionally, the memory includes a non-transitory computer-readable storage medium. The memory 305 can be used to store instructions, programs, codes, code sets or instruction sets. The memory 305 can include a program storage area and a data storage area, wherein the program storage area can store instructions for implementing an operating system, instructions for at least one function (such as a touch function, a sound playing function, an image playing function, etc.), instructions for implementing the above-mentioned various method embodiments, etc.; the data storage area can store data involved in the above-mentioned various method embodiments, etc. The memory 305 can also be at least one storage device located away from the aforementioned processor 301. As a computer storage medium, the memory 305 can include an operating system, a network communication module, a user interface 303 module, and an application program of a multi-value pulse neuron processing method based on dynamic threshold feedback reset.

[0127] In Figure 3In the electronic device shown, the user interface 303 is mainly used to provide an interface for user input, and obtain data input by the user; and the processor 301 can be used to invoke an application program of a multi-value pulse neuron processing method based on dynamic threshold feedback reset stored in the memory 305, which, when executed by one or more processors 301, causes the electronic device to perform the method of one or more of the above embodiments.

[0128] It should be noted that, for the foregoing method embodiments, in order to simply describe, they are all described as a series of action combinations, but those skilled in the art should know that the present application is not limited to the order of the actions described, because according to the present application, certain steps can be performed in other order or at the same time. Secondly, those skilled in the art should know that the embodiments described in the specification all belong to preferred embodiments, and the actions and modules involved are not necessarily essential to the present application.

[0129] In the above embodiments, the description of each embodiment has its own focus, and the parts not described in detail in a certain embodiment can be referred to the related description of other embodiments.

[0130] In several embodiments provided by the present application, it should be understood that the disclosed device can be implemented in other ways. For example, the device embodiments described above are only schematic. The division of the units is only a logical function division. There can be another division manner in actual implementation, for example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the displayed or discussed mutual couplings or direct couplings or communication connections between different parts can be indirect couplings or communication connections through some interfaces, devices or units, and can be electrical or other forms.

[0131] The units described as separate components can or can not be physically separate, and the components displayed as units can or can not be physical units, that is, they can be located in one place, or can be distributed on a plurality of network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the embodiment scheme.

[0132] In addition, each functional unit in each embodiment of the present application can be integrated into a processing unit, or each unit can exist physically independently, or two or more units can be integrated into one unit. The integrated unit can be realized in the form of hardware, or in the form of a software functional unit.

[0133] The integrated unit, if implemented in the form of a software function unit and sold or used as an independent product, can be stored in a computer readable memory. Based on such understanding, the technical solutions of the present application, essentially or in other words, the part that contributes to the prior art or the whole or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a memory 305 and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server or a network device, etc.) to execute all or part of the steps of the embodiments of the present application. The aforementioned memory 305 includes: a U disk, a mobile hard disk, a magnetic disk or an optical disk, and various media that can store program codes.

[0134] The present application also discloses a non-transitory computer-readable storage medium, which stores instructions. When executed by one or more processors 301, the instructions cause an electronic device to perform one or more methods as described in the above embodiments.

[0135] The above are only exemplary embodiments of the present disclosure, and cannot limit the scope of the present disclosure. That is, any equivalent changes and modifications made in accordance with the teachings of the present disclosure are still within the scope of the present disclosure. Other embodiments of the present disclosure will be readily apparent to those skilled in the art upon considering the specification and practicing the true principles of the present disclosure. The present application is intended to cover any variations, uses or adaptive changes of the present disclosure that follow the general principles of the present disclosure and include common knowledge or conventional technical means in the technical field not described in the present disclosure. The specification and examples are only considered as exemplary, and the scope and spirit of the present disclosure are defined by the claims.

Claims

1. A method for processing multi-valued spiking neurons based on dynamic threshold feedback reset, characterized in that, The method includes: Within each time step, a nonlinear mapping is performed on a preset static threshold based on the input current received by the neuron, and a dynamic threshold corresponding to the current input state of the neuron is generated by combining the threshold adjustment parameters. Based on the dynamic threshold, a pulse state variable is introduced to determine whether the neuron has a pulse event at the current time step, and a learnable alternative gradient bound to the pulse state variable is constructed based on the characteristics of the pulse state variable at the pulse trigger position. Instantaneous feedback quantity is generated based on the input current and the pulse state variable. By introducing a cumulative feedback quantity with time memory characteristics, the feedback quantity of the historical time step and the instantaneous feedback quantity of the current time step are weighted and fused to form a cumulative feedback result for adjusting the membrane potential reset intensity. The cumulative feedback result and the soft reset voltage determined based on the dynamic threshold are applied together to the membrane potential reset process to obtain the feedback reset voltage; After the membrane potential is updated by leakage integral and the feedback reset voltage is applied, a multi-valued output function composed of a cutoff function and a floor function is used to map the relationship between the membrane potential and the dynamic threshold into discrete multi-valued pulse output. Based on the membrane potential update equation, the dynamic threshold, the multi-valued firing function, and the feedback reset voltage, a discrete dynamic model of the dynamic threshold feedback reset leakage integral multi-valued firing neuron is constructed. During the training phase, for the multiple implicit thresholds corresponding to the multivalued firing neurons, alternative gradient functions associated with the dynamic thresholds are constructed. During the inference phase, the multivalued pulse output is converted into an equivalent binary pulse representation, and bit-level redundancy is reduced by converting integers to binary, so that the computation process based on the multivalued pulses is remapped into event-driven computation mainly based on sparse accumulation.

2. The method for processing multi-valued spiking neurons based on dynamic threshold feedback resetting according to claim 1, characterized in that, The process of performing a nonlinear mapping on a preset static threshold based on the input current received by the neuron, and generating a dynamic threshold corresponding to the current input state of the neuron by combining threshold adjustment parameters, specifically includes: The input current is introduced into a preset nonlinear mapping function to perform amplitude compression and symbol preservation processing to obtain a threshold modulation intermediate quantity that is limited to a continuous bounded interval. A threshold adjustment parameter corresponding to the neuron is introduced, and the threshold modulation intermediate quantity is weighted and modulated using the threshold adjustment parameter to obtain the threshold modulation quantity; The threshold modulation amount is fused with the preset static threshold to generate a dynamic threshold corresponding to the input state of the neuron at the current time step.

3. The method for processing multi-valued spiking neurons based on dynamic threshold feedback resetting according to claim 1, characterized in that, Based on the dynamic threshold, a pulse state variable is introduced to determine whether the neuron has a pulse event at the current time step. Furthermore, based on the characteristics of the pulse state variable at the pulse triggering location, a learnable alternative gradient bound to the pulse state variable is constructed, specifically including: The dynamic threshold is used as a unified threshold constraint for the current time step, and a multi-valued firing function is calculated based on the comparison between the membrane potential and the dynamic threshold to obtain the multi-valued pulse output of the neuron at the current time step. A pulse state variable corresponding to the multi-valued pulse output is introduced, and the non-differentiable characteristic of the pulse state variable introduced by the sign determination at the pulse trigger position is identified. A learnable alternative gradient is constructed and bound to the impulse state variable for the non-differentiable property; The learnable alternative gradient is configured to maintain the event-determining semantics of the impulse state variable during forward propagation and to continuously approximate the gradient of the impulse state variable during backward propagation in order to recover the gradient propagation path. By binding the learnable parameters in the learnable alternative gradient to the impulse state variable, the learnable parameters are adaptively updated during training according to the requirements of the loss function for the sensitive interval of event determination.

4. The method for processing multi-valued spiking neurons based on dynamic threshold feedback resetting according to claim 1, characterized in that, The process of generating an instantaneous feedback quantity based on the input current and the pulse state variable, and then weighting and fusing the feedback quantity of the historical time step with the instantaneous feedback quantity of the current time step by introducing a cumulative feedback quantity with time memory characteristics to form a cumulative feedback result for adjusting the membrane potential reset intensity, specifically includes: The input current is used as a continuous state quantity characterizing the external driving intensity of the current time step, and the pulse state variable is used as an event indicator quantity characterizing whether a pulse event has occurred in the current time step. Based on the pulse state variable, conditional mapping processing is performed on the input current, so that the input current generates an instantaneous feedback quantity pointing to the pulse penalty direction when the pulse state variable indicates that a pulse event has occurred, and generates an instantaneous feedback quantity pointing to the pulse excitation direction when the pulse state variable indicates that no pulse event has occurred, thereby obtaining an instantaneous feedback quantity that is consistent with the current input state of the neuron and the pulse event state. After generating the instantaneous feedback quantity, a cumulative feedback quantity that is bound to the neuron and recursively updated across time steps is introduced as a feedback state quantity with time memory characteristics. A time decay weight is applied to the cumulative feedback quantity of the previous time step to obtain a historical feedback component, and the historical feedback component is weighted and fused with the instantaneous feedback quantity to obtain the updated cumulative feedback quantity.

5. The method for processing multi-valued spiking neurons based on dynamic threshold feedback resetting according to claim 1, characterized in that, After the membrane potential is updated by leakage integration and the feedback reset voltage is applied, a multi-valued output function composed of a cutoff function and a floor function is used to map the relationship between the membrane potential and the dynamic threshold into a discrete multi-valued pulse output, specifically including: The membrane potential is updated by leakage integral within the current time step, and the feedback reset voltage is applied to the updated membrane potential when a pulse event occurs, so as to obtain the effective membrane potential characterizing the effective state of the neuron at the current time step. The dynamic threshold is used as a unified firing determination criterion, and a continuous firing response quantity is constructed based on the proportional relationship between the effective membrane potential and the dynamic threshold. The continuous firing response is introduced into the truncation function to perform amplitude limiting processing, so as to constrain the continuous firing response within a preset multi-value firing range, thus obtaining the truncated firing response; The truncated firing response is incorporated into a floor function to perform discretization mapping, thereby obtaining a multi-valued pulse output in discrete integer form representing the firing intensity of the neuron at the current time step, wherein the magnitude of the multi-valued pulse output is used to characterize the degree to which the effective membrane potential exceeds the dynamic threshold.

6. The method for processing multi-valued spiking neurons based on dynamic threshold feedback resetting according to claim 1, characterized in that, The step of applying the accumulated feedback result and the soft reset voltage determined based on the dynamic threshold together to the membrane potential reset process to obtain the feedback reset voltage specifically includes: The cumulative feedback result is used as a feedback constraint term for the membrane potential reset strength; Determine the soft reset voltage corresponding to the dynamic threshold within the current time step; Normalization and amplitude shaping are performed on the cumulative feedback result to constrain the cumulative feedback result to a continuous bounded interval while keeping the sign direction unchanged, so as to obtain a normalized cumulative feedback result that is superimposed on the soft reset voltage on the numerical scale. The normalized cumulative feedback result is intensity modulated using a feedback intensity factor bound to the neuron to obtain a modulated feedback result; The modulation feedback result is superimposed and fused with the soft reset voltage to generate a feedback reset voltage that simultaneously contains basic reset information corresponding to the dynamic threshold and historical pulse behavior cumulative constraint information.

7. The method for processing multi-valued spiking neurons based on dynamic threshold feedback resetting according to claim 1, characterized in that, After constructing a discrete dynamic model of the dynamic threshold feedback reset leakage integral multivalued firing neuron based on the membrane potential update equation, the dynamic threshold, the multivalued firing function, and the feedback reset voltage, the method further includes: During the training phase, based on the discrete mapping characteristics of the multivalued firing function, it is identified that the multivalued firing neuron has multiple implicit threshold positions associated with firing levels under the dynamic threshold constraint. For the implicit threshold position, an alternative gradient function is constructed that is explicitly associated with the dynamic threshold, so that the alternative gradient function maintains the discrete firing decision semantics of the multi-value firing function during forward propagation; By configuring the alternative gradient function to provide a gradient response within an effective range determined by the dynamic threshold; The gradient adjustment parameter in the alternative gradient function is set as a learnable parameter, so that the gradient adjustment parameter is adaptively updated during training according to the requirements of the loss function for the multi-valued impulse sensitive region. During the inference phase, an equivalent transformation process is performed on the multi-valued pulse output to map the multi-valued pulse output to a set of binary pulses that maintain consistency in the accumulation result, so that the amplitude information of the multi-valued pulse is split into multiple binary pulse events. An integer-to-binary conversion method is introduced to map the integer amplitude representation corresponding to the multi-value pulse output to a binary bit-level representation; Parallel binary pulse channels are constructed based on the aforementioned binary bit-level representation; During neuron or synapse computation, an accumulation operation is performed on the location of the pulse event based on the binary pulse channel, so that the original computation process based on multi-valued pulses is remapped to the event-driven operation.

8. A multi-value spiking neuron processing device based on dynamic threshold feedback reset, characterized in that, The device is used to execute a multi-value spiking neuron processing method based on dynamic threshold feedback reset as described in any one of claims 1-7. The device includes an acquisition module, a processing module, and an output module, wherein: The acquisition module is used to perform a nonlinear mapping on a preset static threshold based on the input current received by the neuron at each time step, and generate a dynamic threshold corresponding to the current input state of the neuron by combining the threshold adjustment parameters. The processing module is used to introduce a pulse state variable based on the dynamic threshold to determine whether the neuron has a pulse event at the current time step, and to construct a learnable alternative gradient bound to the pulse state variable based on the characteristics of the pulse state variable at the pulse trigger position. The processing module is used to generate an instantaneous feedback quantity based on the input current and the pulse state variable, and to weight and fuse the feedback quantity of the historical time step with the instantaneous feedback quantity of the current time step by introducing a cumulative feedback quantity with time memory characteristics, so as to form a cumulative feedback result for adjusting the membrane potential reset intensity. The processing module is used to apply the accumulated feedback result and the soft reset voltage determined based on the dynamic threshold together to the membrane potential reset process to obtain the feedback reset voltage. The processing module is used to map the relationship between the membrane potential and the dynamic threshold into discrete multi-valued pulse outputs after the membrane potential is updated by leakage integral and the feedback reset voltage is applied, using a multi-valued output function composed of a cutoff function and a floor function. The processing module is used to construct a discrete dynamic model of the dynamic threshold feedback reset leakage integral multi-valued firing neuron based on the membrane potential update equation, the dynamic threshold, the multi-valued firing function, and the feedback reset voltage. The output module is used to construct alternative gradient functions associated with the dynamic thresholds for multiple implicit thresholds corresponding to the multivalued firing neurons during the training phase, and to convert the multivalued pulse output into an equivalent binary pulse representation during the inference phase, and to reduce bit-level redundancy by converting integers to binary, so that the calculation process based on the multivalued pulses is remapped into event-driven operations mainly based on sparse accumulation.

9. An electronic device, characterized in that, The device includes a processor, a communication bus, a user interface, a network interface, and a memory. The memory is used to store instructions. The user interface and the network interface are both used to communicate with other devices. The communication bus is used to enable communication between the components within the electronic device. The processor is used to execute the instructions stored in the memory to cause the electronic device to perform the method as described in any one of claims 1-7.

10. A non-transitory computer-readable storage medium, characterized in that, The computer-readable storage medium stores instructions that, when executed, perform the method as described in any one of claims 1-7.

Images