Neural network plasticity enhancement method based on meta-learning and dynamic neuron reconfiguration

Abstract

This application relates to the fields of artificial intelligence, deep learning, and continuous learning technologies, and particularly to a method for enhancing the plasticity of neural networks based on meta-learning and dynamic neuron reconstruction. The method includes: first, constructing a multi-dimensional closed-loop monitoring system for neural network plasticity; then, establishing an adaptive two-layer architecture with inter-layer functional differentiation within the monitoring system; next, achieving inter-layer collaborative optimization based on the adaptive two-layer architecture through a task difficulty-weighted meta-learning mechanism; and finally, achieving dynamic regulation of neuron activity by introducing a cross-layer neuron dynamic resetting mechanism based on contribution utility. This application addresses the rigidity problem in continuous learning of neural networks through the synergy of "two-layer decoupling" in physical structure and "meta-learning strategy" in algorithm.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence, deep learning and continuous learning technologies, and in particular to a method for enhancing the plasticity of neural networks based on meta-learning and dynamic neuron reconstruction. Background Technology

[0002] Traditional end-to-end models for autonomous driving typically employ a homogeneous network structure, where all parameters follow a uniform update step size during training. However, this structure suffers from serious "technical flaws" when dealing with non-stationary continuous learning tasks: 1. Structural rigidity in updating dynamics: As the model converges for a specific driving task, the weight matrix gradually sinks into a low-dimensional subspace, resulting in a sharp drop in stable rank. Due to the lack of structural redundancy protection, the gradients for new tasks are overwhelmed by the parameter inertia of old tasks, leading to the so-called "plasticity loss".

[0003] 2. The physical conflict between knowledge transfer and flexible adaptation: Within a single parameter space, the model cannot simultaneously ensure both the stability of the underlying physical laws and the flexibility of the top-level driving style. 3. Low neuron utilization and feature redundancy: In conventional gradient descent optimization, the model often suffers from dead neurons or excessively sparse feature representation, failing to actively utilize the network's flexibility to adapt to new environments, resulting in weak rapid generalization ability.

[0004] Existing techniques (such as Dropout or simple weight resetting) lack modeling of the functional division between layers within the network, making it difficult to induce the model to regain plasticity without destroying existing knowledge. Summary of the Invention

[0005] This application provides a method for enhancing the plasticity of neural networks based on meta-learning and dynamic neuron reconstruction. By combining "double-layer decoupling" in physical structure with "meta-learning strategy" in algorithm, it solves the rigidity problem in the continuous learning of neural networks.

[0006] To address the aforementioned technical problems, this application provides a method for enhancing neural network plasticity based on meta-learning and dynamic neuron reconstruction, comprising the following steps: First, constructing a multi-dimensional closed-loop monitoring system for neural network plasticity; then, establishing an adaptive two-layer architecture with inter-layer functional differentiation within the monitoring system; next, achieving inter-layer collaborative optimization based on the adaptive two-layer architecture through a task difficulty-weighted meta-learning mechanism; and finally, achieving dynamic regulation of neuron activity by introducing a cross-layer neuron dynamic reset mechanism based on contribution utility.

[0007] In some exemplary embodiments, a multi-dimensional closed-loop monitoring system for neural network plasticity is constructed, including: quantifying the learning potential of the network by real-time monitoring of network parameters; wherein, the learning potential is quantified by calculating instance sparsity and stable rank. To achieve parameter activity monitoring, the formula for calculating instance sparsity is:

[0008] in, For instance sparsity This is used to quantify the activation efficiency of a neural network when processing a specific input instance; Indicates the dataset Input Sample The expected value calculation reflects the average sparsity state of this layer in a statistical sense; Represents the input of a specific network layer The characteristic response output; Norm, used to count the number of non-zero neurons in the output vector; The total number of neurons (dimensionality) in this network layer is used to normalize the sparsity to a certain value. Interval; by monitoring this indicator, redundant computing units in the network can be identified, thereby enabling inter-layer differential assessment of the activity of parameters in different branch layers.

[0009] By comparing and monitoring the weight variability of different branch layers To conduct inter-layer differentiation assessment.

[0010] In some exemplary embodiments, an adaptive two-layer architecture with inter-layer functional differentiation is established under the monitoring system, including: under the monitoring system, the layer structure of a single model is divided into two branches with different dynamic characteristics, namely the first layer and the second layer.

[0011] In some exemplary embodiments, the first layer is a general feature extraction layer, used to locate and capture spatiotemporally invariant underlying physical representations; the underlying physical representations include lane topology and dynamic constraints.

[0012] In some exemplary embodiments, the second layer is a fast generalization adaptation layer; the second layer takes over the output of the first layer and is specifically responsible for fine-tuning for the current operating conditions; wherein, the current operating conditions include changes in friction in rainy weather and aggressive driver preferences.

[0013] In some exemplary embodiments, the inter-layer interaction logic between the first layer and the second layer is as follows: the first layer provides a robust basic feature stream to the second layer, and the second layer, based on the meta-gradient of the current task... The output is dynamically compensated; the first layer provides input information for the second layer, as shown in the following formula:

[0014] in, This is the overall output of the model; This refers to the Global Feature Extraction function, used to capture input data. The macroscopic topological structure or overall statistical characteristics; This represents a local relational analysis function (Relationship Graph / Analysis), used to quickly adapt to the interaction patterns and micro-evolutionary logic of a specific scenario.

[0015] In some exemplary embodiments, based on an adaptive two-layer architecture, inter-layer collaborative optimization is achieved through a task difficulty-weighted meta-learning mechanism, including: using meta-learning objectives to guide the model into a parameter region that is easy to adapt to; through prior guidance of the outer loop, enabling the first layer to provide a highly generalized foundation, supporting the first layer to achieve the objective with only minimal parameter displacement when facing new conditions; and then based on the task sampling probability of the loss distribution. Perform dynamic task sampling;

[0016] in, To explore factors, a constant term is introduced to ensure that the model retains the basic search probability for all tasks, preventing search stagnation due to insufficient loss. P i The task sampling probability is based on the loss distribution and is used to guide the model in dynamic task sampling. This is the index number of the candidate task in the task pool; For the current number Loss value of each task This reflects the degree of fit or mastery of the model on the task; For the first task in the task set The loss value of each task; The temperature coefficient is used to adjust the smoothness of the sampling distribution. By changing the weight distribution of the denominator, the model's "utilization" of high-loss tasks and "exploration" of regular tasks are balanced. The first and second layers update the target together.

[0017] In some exemplary embodiments, the first layer and the second layer collaboratively update the target, including: an inner loop and an outer loop; the outer loop is used to find an optimal parameter starting point, so that the first layer can support the second layer to achieve the target with only minimal inter-layer compensation when facing new tasks; the process of the outer loop includes: in the query set Q Joint optimization meta objective The inner loop simulates the instantaneous reaction potential within a single model; the inner loop process includes: freezing the first layer and executing only on the fast generalization adaptation layer. k Step gradient descent:

[0018] in, These are the initial parameters for the second layer of the model; The learning rate for the inner loop; For the parameters computed on the support set The gradient reflects the instantaneous response potential of the first layer to new tasks; For internal circulation The updated parameters are obtained after gradient descent. This is the query set used for outer loop evaluation; These are the parameters of the first layer of the model; To jointly optimize the meta-objective on the query set; through the outer loop The joint optimization aims to find the optimal starting point of parameters so that the model can achieve the goal with only minimal inter-layer compensation when facing new tasks.

[0019] In some exemplary embodiments, the contribution utility formula is as follows:

[0020] in, For the computation layer l The Middle i The combined utility of each neuron; h As the activation value, This represents the static contribution, used to quantify the intensity of neuronal feature expression in forward propagation and identify dormant units; Represents the dynamic potential term; quantifies the sensitivity of neurons to absorbing new information during backpropagation; if neuron weights Larger, but gradient If the value approaches zero, it is determined to be a functionally rigid unit.

[0021] In some exemplary embodiments, if a neuron in the second layer... If the value remains below the threshold, reinitialization is performed using a minimal random perturbation.

[0022] The technical solution provided in this application has at least the following advantages: This application provides a method for enhancing neural network plasticity based on meta-learning and dynamic neuron reconstruction, comprising the following steps: First, constructing a multi-dimensional closed-loop monitoring system for neural network plasticity; then, establishing an adaptive two-layer architecture with inter-layer functional differentiation under the monitoring system; next, achieving inter-layer collaborative optimization through a meta-learning mechanism weighted by task difficulty based on the adaptive two-layer architecture; and finally, achieving dynamic regulation of neuron activity by introducing a cross-layer neuron dynamic reset mechanism based on contribution utility.

[0023] This application addresses the stagnation problem in continuous learning of neural networks by synergistically employing a "two-layer decoupling" in physical structure and a "meta-learning strategy" in algorithm, significantly enhancing the plasticity of neural networks. By constructing a multi-dimensional closed-loop monitoring system and a dynamic reconstruction mechanism, this application's method can accurately identify and dynamically adjust functionally stagnant units in the network, effectively solving the model degradation problem caused by parameter update stagnation during continuous learning. Simultaneously, the adaptive two-layer architecture, combined with a task-difficulty-weighted meta-learning mechanism, enables the model to achieve its goal with minimal parameter shifts when facing new tasks, greatly improving the efficiency of inter-layer collaborative optimization. Furthermore, by combining this with cross-layer neuron dynamic resetting based on contribution utility, it achieves simultaneous optimization of model feature representation capabilities and resource utilization. Attached Figure Description One or more embodiments are illustrated by way of example with reference to the accompanying drawings. These illustrations do not constitute a limitation on the embodiments, and unless otherwise stated, the figures in the drawings are not to be limited by scale.

[0024] Figure 1 This is a flowchart illustrating a method for enhancing the plasticity of neural networks based on meta-learning and dynamic neuron reconstruction, provided in an embodiment of this application.

[0025] Figure 2 This is a meta-optimized architecture diagram provided for embodiments of this application.

[0026] Figure 3 The representation verification graph provided in the meta-training of this application embodiment. Detailed Implementation

[0027] As can be seen from the background technology, existing technologies (such as Dropout or simple weight resetting) lack modeling of the functional division between layers within the network, making it difficult to induce the model to regain plasticity without destroying existing knowledge.

[0028] To address this technical problem, this application provides a method for enhancing neural network plasticity based on meta-learning and dynamic neuron reconstruction, comprising the following steps: First, constructing a multi-dimensional closed-loop monitoring system for neural network plasticity; then, establishing an adaptive two-layer architecture with inter-layer functional differentiation within the monitoring system; next, achieving inter-layer collaborative optimization based on the adaptive two-layer architecture through a task difficulty-weighted meta-learning mechanism; finally, achieving dynamic regulation of neuron activity by introducing a cross-layer neuron dynamic reset mechanism based on contribution utility. This application addresses the rigidity problem in continuous learning of neural networks through the synergy of "two-layer decoupling" in physical structure and "meta-learning strategy" in algorithm.

[0029] The embodiments of this application will now be described in detail with reference to the accompanying drawings. However, those skilled in the art will understand that many technical details have been provided in the embodiments of this application to facilitate a better understanding of the application. However, the technical solutions claimed in this application can be implemented even without these technical details and various variations and modifications based on the following embodiments.

[0030] See Figure 1 This application provides a method for enhancing the plasticity of neural networks based on meta-learning and dynamic neuron reconstruction, comprising the following steps: Step S1: Construct a multidimensional closed-loop monitoring system for neural network plasticity.

[0031] Step S2: Under the monitoring system, establish an adaptive two-layer architecture with functional differentiation between layers.

[0032] Step S3: Based on the adaptive two-layer architecture, inter-layer collaborative optimization is achieved through a meta-learning mechanism that weights task difficulty.

[0033] Step S4: By introducing a cross-layer neuron dynamic reset mechanism based on contribution utility, dynamic regulation of neuronal activity is achieved.

[0034] This application relates to a method and system for constructing a functionally differentiated two-layer architecture within a single neural network and achieving inter-layer collaborative updates and dynamic regulation of neuron activity through a meta-learning mechanism.

[0035] In some embodiments, the construction of a multidimensional closed-loop monitoring system for neural network plasticity in step S1 includes: quantifying its learning potential by real-time monitoring of network parameters; wherein, this is achieved by calculating instance sparsity and stability rank. To achieve parameter activity monitoring, the formula for calculating instance sparsity is:

[0036] in, For instance sparsity , used to characterize the activation efficiency of a network layer when processing a specific input; Indicates in the dataset Expectation operation on; For a specific network layer with respect to input samples The output characteristic response; Norm, used to count the number of non-zero (activated) neurons in the output vector; This represents the total number of neurons in this layer, used to normalize the activation scale; this metric is then combined with the weight variability of different branch layers. This can provide a quantitative basis for subsequent inter-layer collaborative updates.

[0037] By comparing and monitoring the weight variability of different layers This serves as the basis for subsequent inter-layer collaborative updates, enabling inter-layer differentiation assessments.

[0038] In some embodiments, step S2 establishes an adaptive two-layer architecture with inter-layer functional differentiation under the monitoring system, including: under the monitoring system, dividing the layer structure of a single model into two branches with different dynamic characteristics, namely the first layer (General Feature Extraction Layer, GFE) and the second layer (Rapid Generalization Adaptation Layer, RGA).

[0039] In some embodiments, the first layer is a general feature extraction layer used to locate and capture spatiotemporally invariant underlying physical representations, including lane topology and dynamic constraints. The first layer is updated at a low frequency, serving as a "knowledge anchor" for the model.

[0040] In some embodiments, the second layer is a fast generalization adaptation layer; the second layer takes over the output of the first layer and is specifically responsible for fine-tuning for the current operating conditions; wherein, the current operating conditions include changes in friction in rainy weather and aggressive driver preferences.

[0041] In some embodiments, the inter-layer interaction logic between the first layer and the second layer is as follows: the first layer provides a robust basic feature stream to the second layer, and the second layer, based on the meta-gradient of the current task... The output is dynamically compensated; the first layer provides input information for the second layer, as shown in the following formula:

[0042] in, This is the overall output of the model; This refers to the Global Feature Extraction function, used to capture input data. The macroscopic topological structure or overall statistical characteristics; This represents a local relational analysis function (Relationship Graph / Analysis), used to quickly adapt to the interaction patterns and micro-evolutionary logic of a specific scenario.

[0043] To address the rigidity issue caused by the interconnectedness of traditional architectures, this approach employs physical decoupling between layers (GFE and RGA). Meta-learning objectives guide the model into a parameter region that is "easily adaptable," rather than simply minimizing error. Through the "prior" guidance of the outer loop, the GFE layer provides a highly generalizable foundation, enabling the RGA layer to achieve its objectives with minimal parameter shifts when facing new operating conditions.

[0044] Dynamic task sampling: To address the issue of over-convergence of the model in simple scenarios, a task sampling probability based on the loss distribution is introduced. :

[0045] in, To sample the probability of tasks based on the loss distribution, higher sampling weights are assigned to tasks with high loss to ensure that the model remains sensitive to boundary conditions of "low plasticity and high difficulty". As an exploration factor, it is used to introduce a constant offset into the distribution to ensure that all tasks have a basic probability of being selected; This is the traversal index for all candidate tasks in the task set; For the current target task Real-time loss value The higher the loss, the lower the model's plasticity to the task, and the more priority should be given to scheduling learning. For the first task in the task set The loss value for each task is used to normalize the probability through the denominator; The temperature coefficient controls the smoothness of the sampling distribution, adjusting the model's trade-off between "utilizing high-loss tasks" and "exploring global tasks." This mechanism forces the model to evolve in a structured manner, thus addressing the rigidity problem of neural networks. This ensures the model remains sensitive to "low-plasticity, high-difficulty" boundary conditions, prioritizes high-loss (low-plasticity) tasks, forces the model to explore in a structured manner, and simultaneously updates the target collaboratively between the first and second layers.

[0046] In some embodiments, the first layer and the second layer collaboratively update the target, including: an inner loop and an outer loop; the outer loop is used to find an optimal parameter starting point, so that the first layer can support the second layer to achieve the target with minimal inter-layer compensation when facing new tasks; the outer loop process includes: in the query set Q Joint optimization meta objective .

[0047] The inner loop simulates the instantaneous reaction potential within a single model; the inner loop process includes: freezing the first layer and executing only on the fast generalization adaptation layer. k Step gradient descent:

[0048] in, These are the initial parameters for the second layer of the model; The learning rate for the inner loop; For the parameters computed on the support set The gradient reflects the instantaneous response potential of the first layer to new tasks; For internal circulation The updated parameters are obtained after gradient descent. This is the query set used for outer loop evaluation; These are the parameters of the first layer of the model; To jointly optimize the meta-objective on the query set; through the outer loop The joint optimization aims to find the optimal starting point of parameters so that the model can achieve the goal with only minimal inter-layer compensation when facing new tasks.

[0049] In step S4, a cross-layer neuron dynamic resetting (CBP) mechanism based on contribution utility (CU) is introduced. To completely solve the LoP problem and reverse plasticity loss at the physical level, this application specifically performs "metabolic renewal" on stiffened neurons—precisely removing redundant units and re-injecting them into the random representation space.

[0050] To accurately quantify the "value" and "learning potential" of each neuron, and thus determine which should be eliminated and rebuilt, and which should be preserved and passed on, a contribution utility formula—the computational layer—is used. l The Middle i The combined effect of individual neurons : In some embodiments, the contribution utility formula is as follows:

[0051] in, For the computation layer l The Middle i The combined utility of each neuron; h As the activation value, This represents the static contribution, used to quantify the intensity of neuronal feature expression in forward propagation and identify dormant units; Represents the dynamic potential term; quantifies the sensitivity of neurons to absorbing new information during backpropagation; if neuron weights Larger, but gradient If the value approaches zero, it is determined to be a functionally rigid unit.

[0052] In some embodiments, if a neuron in the second layer... If the parameters remain below the threshold, reinitialization is performed using minimal random perturbations. This mechanism ensures that the network always maintains "fresh" parameters, thus providing physical redundancy for learning new scenarios.

[0053] The following section provides a detailed description of the neural network plasticity enhancement method based on meta-learning and dynamic neuron reconstruction provided in this application, with reference to the accompanying drawings.

[0054] Figure 2 This paper demonstrates the working mechanism of the Adaptive Dual Network (ADN) throughout its lifecycle. The core of this architecture lies in the strict decoupling of function and architecture: the network is divided into a Global Feature Extractor (GFE, parameter $\theta$) responsible for preserving general prior knowledge, and a Fast Generalization Adapter (RGA, parameter $\omega$) responsible for capturing task-specific residual differences. In the nested optimization of the inner and outer loops during meta-training, and in the subsequent meta-testing phase, this framework achieves a balance between preventing catastrophic forgetting and enabling rapid assimilation of new scenarios by selectively applying structural masking.

[0055] Inner Loop Adaptation: Unlike the global updates in traditional deep learning, meta-learning is performed on the sampled task distribution $\{Task_i\}$. For the $i$-th support set, the model first copies the initial parameters $\omega_0$ of the RGA for that task. Subsequently, while keeping the GFE frozen, only $\omega$ is updated for $K$ rounds: $\omega_{j+1} \leftarrow \omega_j - \alpha \nabla_{\omega_j} l(f_\theta, \omega_j)$. This step forces the RGA to quickly fit the unique features of the current task.

[0056] Outer-loop meta-update: After the inner-loop trial and error is completed, the updated local parameter set is combined with the frozen network, and the unbiased loss $\nabla_{\theta} l(f_\Theta)$ is calculated on the query set. Subsequently, using this cross-task comprehensive validation gradient, a highly conservative macro-level fine-tuning of $\theta$ is performed on the outer loop: $\theta \leftarrow \theta - \beta_1 \sum \nabla_\theta$. This ensures that GFE absorbs universal, generalized patterns, rather than biases specific to a single task.

[0057] Meta-testing phase: The system aims to achieve rapid and safe adaptation to novel and unknown tasks. To prevent catastrophic forgetting while absorbing new knowledge, the Global Feature Extractor (GFE) is strictly frozen. The system only activates the highly dynamic fast generalization adapter, which computes the $\omega$-specific gradient using the support set of the new task and performs rapid iterative updates to $\omega \leftarrow \omega - \gamma \nabla_\omega$. After each fine-tuning, the system performs validation evaluation on an independent query set; once the generalization loss of the current model falls below a preset target threshold, the loop terminates immediately, indicating that the RGA has successfully captured and adapted to the feature differences of the new scenario, and finally outputs a deployment-ready specialized model.

[0058] To verify the evolution of multi-task feature representations and its impact on model plasticity, Figure 3 The activation values ​​of the feature extraction network at different training stages were visualized. By randomly sampling data from different tasks, these intermediate activation values ​​can be used as an indicator to examine whether the inner loop sampling effectively captures shared information across tasks. Figure 3 The evolution trajectory of representations from 10,000 to 500,000 steps is illustrated using stacked bar charts, and a detailed activation map is presented at the end of meta-training. A significant divergence in optimization dynamics emerges when comparing standard multi-task learning (MTL) with our proposed PlasticMeta. In the MTL baseline model, premature overactivation is observed before 50,000 steps, eventually collapsing into an overly sparse representation dependent on a few dominant neurons. This stems from a sample-centric greedy optimization strategy that triggers a "winner-takes-all" dynamic: once the dominant units are sufficient to explain the aggregate loss, the error gradient flowing to unsaturated neurons weakens drastically, causing these neurons to become dormant due to resource scarcity, resulting in structural rigidity. In stark contrast, PlasticMeta prioritizes a fair and universal parameter initialization strategy, actively utilizing a broader spectrum of neurons to converge their activation values ​​to stable, small, non-zero values. By limiting plasticity degradation to the local network architecture through architectural design and fairly distributing computational load through continuous backpropagation, PlasticMeta effectively protects the protected global feature extractor from representation stagnation and eliminates structural rigidity caused by the dominant unit.

[0059] Based on the above technical solutions, this application provides a method for enhancing the plasticity of neural networks based on meta-learning and dynamic neuron reconstruction, comprising the following steps: First, constructing a multi-dimensional closed-loop monitoring system for neural network plasticity; then, establishing an adaptive two-layer architecture with inter-layer functional differentiation within the monitoring system; next, achieving inter-layer collaborative optimization through a task difficulty-weighted meta-learning mechanism based on the adaptive two-layer architecture; finally, achieving dynamic regulation of neuron activity by introducing a cross-layer neuron dynamic reset mechanism based on contribution utility. This application addresses the rigidity problem in continuous learning of neural networks through the synergy of "two-layer decoupling" in physical structure and "meta-learning strategy" in algorithm.

[0060] Those skilled in the art will understand that the above-described embodiments are specific examples of implementing this application, and in practical applications, various changes in form and detail may be made without departing from the spirit and scope of this application. Any person skilled in the art can make their own modifications and alterations without departing from the spirit and scope of this application; therefore, the scope of protection of this application should be determined by the scope defined in the claims.

Claims

1. A method for enhancing the plasticity of neural networks based on meta-learning and dynamic neuron reconstruction, characterized in that, Includes the following steps: Constructing a multidimensional closed-loop monitoring system for neural network plasticity; Under the monitoring system, an adaptive two-layer architecture with functional differentiation between layers is established; Based on an adaptive two-layer architecture, inter-layer collaborative optimization is achieved through a meta-learning mechanism that weights tasks by difficulty. By introducing a cross-layer neuron dynamic reset mechanism based on contribution utility, dynamic regulation of neuronal activity can be achieved.

2. The neural network plasticity enhancement method based on meta-learning and dynamic neuron reconstruction according to claim 1, characterized in that, Constructing a multidimensional closed-loop monitoring system for neural network plasticity, including: The learning potential of a network is quantified by monitoring its parameters in real time; this includes calculating instance sparsity and stable rank. To achieve parameter activity monitoring, the formula for calculating instance sparsity is: in, Indicates instance sparsity, used to measure the sparsity of neuron activation; Indicates in the dataset Mathematical expectation on; This indicates that the network layer is at the input Lower activation value Norm, which is the number of non-zero activated neurons; d This represents the total dimension or total number of neurons in that neural layer; the weight variability of different layers is monitored by comparison. To conduct inter-layer differentiation assessment.

3. The neural network plasticity enhancement method based on meta-learning and dynamic neuron reconstruction according to claim 1, characterized in that, Within the monitoring system, an adaptive two-tier architecture with functional differentiation between layers is established, including: Under the monitoring system, the layer structure of a single model is divided into two branches with different dynamic characteristics, namely the first layer and the second layer.

4. The neural network plasticity enhancement method based on meta-learning and dynamic neuron reconstruction according to claim 3, characterized in that, The first layer is a general feature extraction layer, used to locate and capture spatiotemporally invariant underlying physical representations; The underlying physical representation includes lane topology and dynamic constraints.

5. The neural network plasticity enhancement method based on meta-learning and dynamic neuron reconstruction according to claim 1, characterized in that, The second layer is a fast generalization and adaptation layer; it receives the output from the first layer and is specifically responsible for fine-tuning for the current working conditions; among them, Current operating conditions include changes in friction due to rain and aggressive driver preferences.

6. The neural network plasticity enhancement method based on meta-learning and dynamic neuron reconstruction according to claim 1, characterized in that, The interaction logic between the first layer and the second layer is as follows: The first layer provides a robust base feature flow for the second layer, which in turn is based on the meta-gradient of the current task. The output is dynamically compensated, and the two layers achieve heterogeneous target tasks through different optimization methods; The first layer provides input information to the second layer, as shown in the following formula: in, This is the overall output of the model; This represents the global feature extraction function, used to capture input data. The macroscopic topological structure or overall statistical characteristics; This represents a local relational analysis function, used to quickly adapt to the interaction patterns and micro-evolutionary logic of specific scenarios.

7. The neural network plasticity enhancement method based on meta-learning and dynamic neuron reconstruction according to claim 3, characterized in that, Based on an adaptive two-layer architecture, inter-layer collaborative optimization is achieved through a task difficulty-weighted meta-learning mechanism, including: Meta-learning objectives are used to guide the model into a parameter region that is easy to adapt to; Through prior guidance from the outer loop, the first layer can provide a highly generalized base, supporting the first layer to achieve the goal with only minimal parameter displacement when facing new working conditions; the first layer and the second layer collaboratively update the goal; Then, based on the task sampling probability of the loss distribution Perform dynamic task sampling; in, To explore factors, a constant bias is introduced to ensure the base sampling rate and avoid the model getting trapped in local optima; For the first The sampling probability of each task; For the first The current loss value for each task is used to measure the training difficulty or mastery of that task. The index of the task set; This is the set of loss values ​​for each task in the task pool; The temperature coefficient is used to smooth the probability distribution and control the balance between "exploration" and "exploitation" in the sampling strategy. Through this dynamic sampling mechanism, the model can prioritize weak links with higher losses and achieve adaptive adjustment of task weights.

8. The neural network plasticity enhancement method based on meta-learning and dynamic neuron reconstruction according to claim 7, characterized in that, The first and second layers collaboratively update the target, including: Inner and outer loops; the outer loop is used to find an optimal starting point for parameters, so that the first layer can support the second layer to achieve its goal with minimal inter-layer compensation when facing new tasks; the outer loop process includes: jointly optimizing the meta-objective on the query set Q. ; The inner loop simulates the instantaneous reaction potential within a single model; the inner loop process includes: freezing the first layer and executing only on the fast generalization adaptation layer. k Step gradient descent: in, These are the initial parameters for the second layer of the model; The learning rate of the inner loop is used to control the step size for gradient updates on the support set. This indicates the parameter on the support set. The gradient reflects the model's instantaneous responsiveness to the current specific task; These are the new parameters obtained after rapid adaptation via the inner loop; The representative query set is used to evaluate the generalization performance of the adapted parameters on the task. These are the parameters of the first layer of the model; The loss function is applied to the query set; this is achieved by adjusting the loss function in the outer loop. Joint optimization aims to find an optimal starting point for parameters, so that the model can quickly adapt to new tasks with minimal inter-layer compensation.

9. The neural network plasticity enhancement method based on meta-learning and dynamic neuron reconstruction according to claim 1, characterized in that, The formula for contribution utility is as follows: in, For the computation layer l The Middle i The combined utility of each neuron; h As the activation value, This represents the static contribution, used to quantify the intensity of neuronal feature expression in forward propagation and identify dormant units; Represents the dynamic potential term; quantifies the sensitivity of neurons to absorbing new information during backpropagation; if neuron weights Larger, but gradient If the value approaches zero, it is determined to be a functionally rigid unit.

10. The neural network plasticity enhancement method based on meta-learning and dynamic neuron reconstruction according to claim 9, characterized in that, If a neuron in the second layer... If the value remains below the threshold, reinitialization is performed using a minimal random perturbation.

Images