A neuron allocation strategy to achieve large-scale model continuous learning method and system

Abstract

The present invention relates to the field of model training technology, and specifically to a method and system for realizing continuous learning of a large model using a neuron allocation strategy, wherein the method comprises the following steps: the large model adopts a multi-level knowledge storage structure, and when updating knowledge, the knowledge unit is protected by a selective protection mechanism based on importance; when the large model is continuously learning, the performance and computing resources of the large model are monitored in real time, and the task resource allocation scheme is dynamically optimized and executed based on the real-time performance monitoring information and resource status information; by evaluating the importance of neurons and introducing the performance contribution of neurons to the current task, the computing resources and update frequency of each neuron are dynamically allocated in layers according to the importance of each neuron. The present invention can significantly improve the efficiency of system resource utilization through a dynamic resource management mechanism and a resource allocation strategy based on importance, thereby achieving optimal adjustment of the overall efficiency.

Description

Technical Field

[0001] The present invention relates to the field of model training technology, and in particular to a method and system for implementing continuous learning of large models through a neuron allocation strategy. Background Art

[0002] The main challenges faced by large language models during continuous learning include cold start problems, concept drift, and catastrophic forgetting. Existing technical solutions primarily address parameter tuning and architectural optimization, but they still have significant shortcomings in cognitive resource allocation and knowledge representation. Traditional deep learning models often experience catastrophic forgetting when learning new tasks, meaning they acquire new knowledge while significantly losing previously acquired capabilities. This problem is particularly prominent in the field of large language models. For example, when a model updates its knowledge through incremental training, it often causes a significant degradation of its original capabilities, severely limiting its practical application value.

[0003] To address this issue, researchers have proposed various continuous learning methods. In 2017, DeepMind proposed the elastic weight consolidation method, which protects key knowledge by calculating parameter importance. In 2020, Facebook AI Research proposed the progressive neural network, which dynamically expands the network structure to adapt to new tasks. However, these methods all have significant limitations: they either require significant computing resources or struggle to maintain stable performance in complex scenarios.

[0004] Especially in the continuous learning scenarios of large language models, the models need to maintain general capabilities while constantly adapting to new domain knowledge and task requirements. Traditional methods are often unable to cope with such complex learning objectives. On the one hand, the model needs to effectively manage the learning process of hundreds of billions of parameters; on the other hand, it needs to balance the acquisition and retention of new and old knowledge within limited computing resources. The resource classification schemes of existing technologies are generally preset resource allocation schemes with a high degree of manual participation. As a result, it is impossible to dynamically allocate the computing resources required for each part according to the training process, and it is impossible to achieve optimal adjustment of overall efficiency. The limitations are quite obvious. Summary of the Invention

[0005] The technical problem to be solved by the present invention is as follows: Existing model training methods have certain limitations, and catastrophic forgetting often occurs when learning new tasks, which can easily lead to significant degradation of original capabilities during continuous learning. At the same time, existing resource allocation methods cannot balance the acquisition and retention of new and old knowledge under limited computing resources, resulting in the inability to dynamically allocate the computing resources required for each part according to the training process, and unable to achieve optimal adjustment of overall efficiency.

[0006] To solve the above technical problems, the first aspect of the present invention adopts the following technical solution: a neuron allocation strategy to implement a large model continuous learning method, comprising the following steps:

[0007] The large model adopts a multi-level knowledge storage structure. When updating knowledge, it protects knowledge units through a selective protection mechanism based on importance.

[0008] When a large model is continuously learning, the performance and computing resources of the large model are monitored in real time. Based on the real-time performance monitoring information and resource status information, the task resource allocation plan is dynamically optimized and executed;

[0009] By evaluating the importance of neurons and introducing the degree of contribution of neurons to the performance of the current task, the computing resources and update frequency of each neuron are dynamically allocated in layers according to the importance of each neuron.

[0010] When the present invention works, it can significantly improve the efficiency of system resource utilization and effectively reduce computing overhead through dynamic resource management mechanism and resource allocation strategy based on importance. Through real-time monitoring of large model performance and computing resources, it can dynamically adjust resource allocation according to the actual effect of training, thereby achieving optimal adjustment of overall efficiency. At the same time, the knowledge protection mechanism based on importance and the progressive knowledge integration method can significantly enhance the transfer learning effect of the model between tasks in different fields, realize asymmetric response to changes in system performance, shorten the adaptation time to new tasks, and have stronger generalization ability.

[0011] Preferably, the multi-level knowledge storage structure includes a basic knowledge layer for storing general extraction capabilities of large models, a task-specific layer for storing professional knowledge related to specific tasks, and a cross-task relationship layer for recording knowledge transfer relationships between different tasks.

[0012] Preferably, when protecting knowledge units through a selective protection mechanism based on importance, the following steps are adopted to obtain the verification performance, usage frequency and timeliness of the knowledge units, and output the importance of the knowledge units through a weighted summation method. When the importance of the knowledge unit is greater than a preset threshold, the knowledge unit is protected by adjusting the learning rate and update threshold.

[0013] When the present invention works, it quantifies the importance index of knowledge units through a selective protection mechanism based on importance, and gives stronger protection to knowledge units with higher importance by adjusting the learning rate and update threshold. When facing complex and dynamic multi-task continuous learning scenarios, it can balance the acquisition and retention of new and old knowledge, thereby avoiding the forgetting of key knowledge.

[0014] Preferably, when real-time monitoring of the performance and computing resources of the large model is performed, the following steps are adopted: by monitoring the resource status data stream during the continuous learning of the large model, a number of resource status parameters corresponding to each task are obtained, and real-time resource status information is obtained by calculation, wherein the resource status parameters include at least one of memory usage, processor load, and energy consumption indicators;

[0015] By monitoring the performance data flow of the large model during continuous learning, several performance evaluation indicators corresponding to the large model are obtained, and real-time performance monitoring information is obtained through calculation. The performance evaluation indicators include at least one of accuracy, recall rate, task completion time and resource utilization.

[0016] Preferably, the following steps are used to dynamically optimize the task resource allocation plan and execute it based on real-time performance monitoring information and resource status information:

[0017] A1: Obtain real-time performance monitoring information, set dynamic adjustment coefficients based on the acquired resource status parameters and a preset coefficient adjustment mechanism, and calculate resource consumption information using a weighted summation method.

[0018] A2: Calculate the efficiency of the current task resource allocation plan based on the performance monitoring information. If the efficiency of the current task resource allocation plan is lower than a preset threshold, proceed to step A3.

[0019] A3: Adopting a performance-constrained optimization method, the task resource allocation scheme is dynamically optimized by solving the optimization problem so that the task performance is greater than the preset minimum performance threshold.

[0020] When the present invention works, it comprehensively reflects the operating status of the system by constructing a multi-dimensional performance evaluation system. Based on the loss aversion principle and the optimal solution of performance constraints, it can effectively prevent a significant performance degradation during the continuous learning process of the large model, provide reliable performance guarantee for the continuous learning process of the large model, and also provide an important decision-making basis for the subsequent neuron allocation strategy.

[0021] As a preference, by evaluating the importance of neurons and introducing the degree of contribution of neurons to the performance of the current task, the following steps are adopted when dynamically allocating the computing resources and update frequency of each neuron in a hierarchical manner according to the importance of each neuron:

[0022] B1: By monitoring the performance data stream of the large model during continuous learning, the activation frequency of neurons, the strength of connection weights, and the task relevance factor that reflects the contribution of neurons to the current task are obtained, and the importance of neurons is calculated;

[0023] B2: Divide the range according to the preset importance and set the neurons to match the resource allocation priority;

[0024] B3: Based on the resource allocation priority, the preset adjustment mechanism is called to hierarchically allocate the computing resources and update frequency of each neuron. The adjustment mechanism includes a performance elasticity adjustment mechanism based on elastic coefficient allocation, a task characteristic adjustment mechanism that sets allocation templates for different task types, a periodic allocation adjustment mechanism that reserves a preset proportion of exploration time for periodic resource experiments, and an intra-level allocation adjustment mechanism for implementing refined allocation within each importance division range.

[0025] When the present invention is working, it quantifies the contribution of neurons to the current task, realizes dynamic resource allocation based on neuron importance evaluation, ensures that key neurons can obtain sufficient resources in a timely manner, realizes accurate allocation and dynamic adjustment of neuron resources, and ensures that the system maintains efficient resource utilization during the continuous learning process.

[0026] Preferably, the following steps are also included: constructing a current scenario vector, triggering and / or actively performing an associated query in a preset experience memory library, and outputting and executing matching response strategies under similar historical scenarios based on scenario similarity calculation. In the specific implementation, performance scenario similarity and / or resource scenario similarity are included, which can be adjusted according to actual needs.

[0027] Preferably, the experience memory bank is established by the following steps: obtaining historical scenarios, corresponding response strategies and subsequent actual effects, and storing the experience memory bank through a preset hybrid data storage solution.

[0028] When the present invention works, it integrates the experience memory library and constructs the current scenario vector to trigger and / or actively perform associated query work in the preset experience memory library, thereby calculating and outputting matching response strategies under similar historical scenarios based on scenario similarity, which can achieve more intelligent resource allocation and improve the efficiency of system resource utilization. By accumulating historical experience through the establishment of an experience memory library, it can fully apply historical experience while also predicting potential risks, thereby further making forward-looking adjustments and achieving more intelligent resource allocation effects.

[0029] Preferably, when the large model continues to learn, the following steps are also included: representative historical samples are obtained through the importance sampling method, a dynamically updated experience pool is established, and samples are selected from the experience pool for playback training through a preset memory playback cycle to enhance the retention of important knowledge.

[0030] When the present invention is working, in order to avoid catastrophic forgetting, a dynamic experience pool is maintained to store representative historical samples, and a memory replay mechanism is used to select samples from the experience pool for replay training, which can strengthen the retention of important knowledge and at the same time ensure the consistency of knowledge updating to avoid hallucination problems.

[0031] To solve the above technical problems, the second aspect of the present invention adopts the following technical solution: a neuron allocation strategy is used to implement a large-model continuous learning system, and a neuron allocation strategy as described in any of the above aspects is used to implement a large-model continuous learning method, characterized by comprising:

[0032] Cognitive resource management module, used for comprehensive real-time monitoring and output of decision-making basis;

[0033] The performance protection mechanism adjustment module is used to monitor the performance of large models in real time and output optimization suggestions through an asymmetric response mechanism to make timely adjustments to large model performance changes;

[0034] Experience memory module, used to store historical experience fragments for retrieval when formulating resource allocation plans;

[0035] Neuron resource dynamic allocation module, used to generate and execute resource allocation plans based on decision-making basis and optimization suggestions;

[0036] The knowledge integration module is used for continuous learning of large models and the effective integration of new and old knowledge while ensuring resource constraints;

[0037] The cognitive resource management module, the performance protection mechanism adjustment module, the experience memory library module and the neuron resource dynamic allocation module are data-connected, and the monitoring end of the cognitive resource management module and the monitoring end of the performance protection mechanism adjustment module are both connected to the knowledge integration module to monitor their resource status data flow and performance data flow.

[0038] When the present invention works, it adopts a systematic information transmission scheme including experience feedback, and introduces structured experience data flow on the basis of the original performance data flow, resource status flow, scheduling instruction flow, and knowledge status flow, so as to ensure the coordinated operation of each link of the system, ensure the effective realization of the overall function and the improvement of intelligence, and at the same time, it can perform intelligent prediction and adaptation based on experience. Through the collaborative work of each module, the system has high intelligence, robustness and foresight in optimizing resource allocation in a dynamic environment, providing a more reliable guarantee for the long-term stable and efficient operation of the system.

[0039] The beneficial technical effects of the present invention include:

[0040] 1. The present invention can significantly improve the efficiency of system resource utilization and effectively reduce computing overhead through a dynamic resource management mechanism and a resource allocation strategy based on importance. By real-time monitoring of large model performance and computing resources, resource allocation can be dynamically adjusted according to the actual training effect, thereby achieving optimal adjustment of overall efficiency. At the same time, the knowledge protection mechanism based on importance and the progressive knowledge integration method can significantly enhance the transfer learning effect of the model between tasks in different fields, achieve an asymmetric response to changes in system performance, shorten the adaptation time to new tasks, and have stronger generalization capabilities.

[0041] 2. The present invention quantifies the importance index of knowledge units through a selective protection mechanism based on importance, and gives stronger protection to knowledge units with higher importance by adjusting the learning rate and update threshold. When facing complex and dynamic multi-task continuous learning scenarios, it can balance the acquisition and retention of new and old knowledge, thereby avoiding the forgetting of key knowledge.

[0042] 3. The present invention comprehensively reflects the operating status of the system by constructing a multi-dimensional performance evaluation system. Based on the loss aversion principle and the optimal solution of performance constraints, it can effectively prevent a significant performance degradation during the continuous learning process of the large model, provide reliable performance guarantees for the continuous learning process of the large model, and also provide an important decision-making basis for the subsequent neuron allocation strategy.

[0043] 4. The present invention quantifies the contribution of neurons to the current task, realizes dynamic resource allocation based on neuron importance evaluation, ensures that key neurons can obtain sufficient resources in a timely manner, realizes accurate allocation and dynamic adjustment of neuron resources, and ensures that the system maintains efficient resource utilization during the continuous learning process.

[0044] 5. The present invention integrates the experience memory library and constructs the current scenario vector to trigger and / or actively perform associated queries in the preset experience memory library, thereby outputting matching response strategies under similar historical scenarios based on scenario similarity calculations, thereby achieving more intelligent resource allocation and improving system resource utilization efficiency. By accumulating historical experience through the establishment of an experience memory library, it is possible to fully apply historical experience while also predicting potential risks, thereby further making forward-looking adjustments and achieving more intelligent resource allocation effects.

[0045] 6. In order to avoid catastrophic forgetting, the present invention maintains a dynamic experience pool to store representative historical samples, and adopts a memory replay mechanism to select samples from the experience pool for replay training, which can strengthen the retention of important knowledge, while also ensuring the consistency of knowledge updating and avoiding hallucination problems.

[0046] 7. The present invention adopts a systematic information transmission scheme that includes experience feedback, and introduces structured experience data flow on the basis of the original performance data flow, resource status flow, scheduling instruction flow, and knowledge status flow, to ensure the coordinated operation of all links in the system, ensuring the effective realization of the overall function and the improvement of intelligence. At the same time, it can perform intelligent prediction and adaptation based on experience, and through the collaborative work of various modules, the system has high intelligence, robustness and foresight in optimizing resource allocation in a dynamic environment, providing a more reliable guarantee for the long-term stable and efficient operation of the system.

[0047] Other features and advantages of the present invention will be disclosed in detail in the following specific embodiments and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] The present invention will be further described below with reference to the accompanying drawings:

[0049] Figure 1 A schematic diagram of the structure of a large-scale continuous learning system for a neuron allocation strategy;

[0050] Figure 2 A workflow diagram for dynamically optimizing task resource allocation and executing a neuron allocation strategy in a large-scale continuous learning method;

[0051] Figure 3 A workflow diagram for a neuron allocation strategy that dynamically allocates computing resources of neurons based on their importance in a large-scale continuous learning method;

[0052] Figure 4 A workflow diagram for a neuron allocation strategy that implements a large-model continuous learning method based on scenario similarity calculation to output matching response strategies under similar historical scenarios. DETAILED DESCRIPTION

[0053] The following is an explanation and description of the technical solutions of the embodiments of the present invention in conjunction with the drawings of the embodiments of the present invention. However, the following embodiments are only preferred embodiments of the present invention and are not exhaustive. Based on the embodiments in the implementation manner, other embodiments obtained by those skilled in the art without creative work are all within the scope of protection of the present invention.

[0054] In the following description, terms such as "inside", "outside", "up", "down", "left", "right", etc. that indicate directions or positional relationships are only used to facilitate the description of the embodiments and simplify the description, and do not indicate or imply that the device or element referred to must have a specific direction, be constructed and operate in a specific direction. Therefore, they should not be understood as limiting the present invention.

[0055] Example 1:

[0056] See also Figure 1 This embodiment discloses a method for implementing continuous learning of a large model using a neuron allocation strategy, including the following steps:

[0057] The large model adopts a multi-level knowledge storage structure. When updating knowledge, it protects knowledge units through a selective protection mechanism based on importance.

[0058] When a large model is continuously learning, the performance and computing resources of the large model are monitored in real time. Based on the real-time performance monitoring information and resource status information, the task resource allocation plan is dynamically optimized and executed;

[0059] By evaluating the importance of neurons and introducing the degree of contribution of neurons to the performance of the current task, the computing resources and update frequency of each neuron are dynamically allocated in layers according to the importance of each neuron.

[0060] When this example is working, the dynamic resource management mechanism and the resource allocation strategy based on importance can significantly improve the efficiency of system resource utilization and effectively reduce computing overhead. Through real-time monitoring of large model performance and computing resources, resource allocation can be dynamically adjusted according to the actual training effect, thereby achieving optimal adjustment of overall efficiency. At the same time, the knowledge protection mechanism based on importance and the progressive knowledge integration method can significantly enhance the transfer learning effect of the model between tasks in different fields, realize asymmetric response to changes in system performance, shorten the adaptation time to new tasks, and have stronger generalization ability.

[0061] In specific implementation, the multi-level knowledge storage structure includes a basic knowledge layer for storing the general extraction capabilities of large models, a task-specific layer for storing professional knowledge related to specific tasks, and a cross-task relationship layer for recording the knowledge transfer relationship between different tasks. In specific implementation, the basic knowledge layer can be recorded in a sparse coding manner, the task-specific layer can be indexed using task identifiers, and the cross-task association layer can represent the knowledge relevance between different tasks through an association matrix. Of course, any other suitable storage method can also be used. When updating knowledge, the verification performance, usage frequency, and timeliness of the knowledge unit are obtained, and the importance of the knowledge unit is output through a weighted summation method. When working, the following formula is used for calculation:

[0062] ;

[0063] in: is the importance of the knowledge unit, For the corresponding weight coefficient, when working, it can be customized according to the actual training requirements, or it can be dynamically adjusted based on historical experience. is the verification performance of the knowledge unit, is the frequency of use of the knowledge unit, To ensure the timeliness of knowledge units, when the importance of a knowledge unit is greater than a preset threshold, the learning rate and update threshold are adjusted to protect the knowledge unit, thereby giving stronger protection to knowledge units with higher importance.

[0064] When working in this example, a selective protection mechanism based on importance is used to quantify the importance index of knowledge units. By adjusting the learning rate and update threshold, knowledge units with higher importance are given stronger protection. When facing complex and dynamic multi-task continuous learning scenarios, it can balance the acquisition and retention of new and old knowledge, thereby avoiding the forgetting of key knowledge.

[0065] Preferably, a knowledge dependency management method based on a graph structure can be used to maintain the consistency of knowledge, track and maintain the dependency relationships between knowledge, and ensure the coherence of knowledge updates.

[0066] As a further improvement of this embodiment, in order to avoid knowledge conflicts, it is also necessary to perform preset conflict resolution work when a conflict is detected between new and old knowledge. In the specific implementation, it is necessary to first evaluate the degree of conflict and the scope of impact, and select a retention strategy based on the verification performance. When necessary, it is necessary to establish a conditional mapping relationship to realize the contextual application of knowledge, thereby realizing the effective accumulation and updating of large model knowledge. The knowledge integration mechanism not only ensures the protection of important knowledge, but also provides reliable guarantees for the continuous evolution of the model, and ultimately achieves a stable improvement in the continuous learning ability of the large model.

[0067] Preferably, the preset integration strategy set can be called to perform knowledge update based on the similarity between the new knowledge and the existing knowledge units. Specifically, the following steps are adopted: receiving a new learning task, traversing the existing knowledge units, calculating and outputting the similarity between the new knowledge and the existing knowledge units, and calling the matching integration strategy in the preset integration strategy set based on the similarity between the new knowledge and the existing knowledge units to perform knowledge update or fusion. The integration strategy set includes at least one of an incremental update integration strategy for realizing knowledge fusion by fine-tuning existing parameters, a partial parameter sharing integration strategy for introducing new features based on the original knowledge, and an independent integration strategy for allocating independent parameter spaces and establishing necessary relationship mappings. The appropriate integration strategy is selected through knowledge of different similarities. This progressive knowledge integration method can reduce system pressure and improve the efficiency of knowledge updating.

[0068] Preferably, when real-time monitoring of large model performance and computing resources is performed, the following steps are adopted: by monitoring the resource status data stream during continuous learning of the large model, a number of resource status parameters corresponding to each task are obtained, and real-time resource status information is obtained by calculation. The resource status parameters include at least one of memory usage, processor load, and energy consumption indicators. The resource consumption indicator can be calculated using the following formula:

[0069] ;

[0070] in: is the resource consumption indicator, is the dynamic adjustment coefficient, For memory usage, is the processor load, For energy consumption indicators, when working, the resource saturation adaptive mechanism can be used to increase the coefficient when approaching the threshold, or the marginal utility dynamic adjustment can be used to adjust according to the change in performance gain per unit of resources, and the task occupancy ratio limit can be used to exponentially increase the coefficient when a single task exceeds, for example, 20% of the total resources. Through these mechanisms, we can jointly ensure that resource allocation follows the principle of diminishing marginal utility, prevent excessive concentration of resources, reflect the diminishing marginal utility characteristics of resource use, and effectively prevent excessive concentration of resources on specific tasks.

[0071] By monitoring the performance data stream of the large model during continuous learning, several performance evaluation indicators corresponding to the large model are obtained, and real-time performance monitoring information is obtained through calculation. The performance evaluation indicators include at least one of accuracy, recall rate, task completion time, and resource utilization rate. The performance evaluation indicators can be calculated using the following formula:

[0072] ;

[0073] in: is the performance evaluation index, is the value of the i-th performance indicator at time t, The weight coefficient corresponding to each performance indicator can fully reflect the system operation status.

[0074] In specific implementation, in order to better realize the performance protection mechanism, a multi-threshold adaptive adjustment mechanism can be adopted, and a performance monitoring mechanism based on sliding windows can be established when adjusting the strategy in real time to timely discover performance change trends. Once a possible performance degradation trend is detected, preventive adjustment measures can be immediately initiated. Differentiated adjustment strategies can be implemented according to the type of performance change, such as slow decline, sudden decline, specific task decline, resource pressure decline, etc. For example, knowledge replay frequency and regularization strength can be adjusted during slow decline, checkpoint rollback and problem isolation can be triggered during sudden decline, dedicated resources can be allocated and relevant knowledge can be reactivated when the performance of a specific task declines, and intelligent caching and batch optimization can be initiated when the performance decline is caused by resource pressure. Furthermore, in order to avoid overreaction, an adjustment threshold mechanism can be set, a tolerance range for performance changes can be set, and the adjustment process can be triggered only when the range is exceeded. Long-term performance monitoring can also be performed, and a cumulative effect evaluation mechanism can be introduced to identify long-term decline trends by calculating the cumulative value of performance changes in N consecutive time windows, and initiate more in-depth adjustment strategies.

[0075] In the specific implementation, the following steps are also included: obtaining representative historical samples through the importance sampling method, establishing a dynamically updated experience pool, selecting samples from the experience pool for playback training through a preset memory playback cycle, and enhancing the retention of important knowledge.

[0076] When this example is working, in order to avoid catastrophic forgetting, a dynamic experience pool is maintained to store representative historical samples, and a memory replay mechanism is used to select samples from the experience pool for replay training. This can strengthen the retention of important knowledge, while also ensuring the consistency of knowledge updates and avoiding hallucination problems.

[0077] Example 2:

[0078] This embodiment provides a neuron allocation strategy to implement a large model continuous learning method. The similarities with other embodiments are not repeated here, and the differences are described in detail below.

[0079] See also Figure 2 In this embodiment, the following steps are used to dynamically optimize the task resource allocation plan and execute it based on real-time performance monitoring information and resource status information:

[0080] A1: Obtain real-time performance monitoring information, set dynamic adjustment coefficients based on the acquired resource status parameters and a preset coefficient adjustment mechanism, and calculate resource consumption information using a weighted summation method.

[0081] A2: Calculate the efficiency of the current task resource allocation plan based on the performance monitoring information. If the efficiency of the current task resource allocation plan is lower than a preset threshold, proceed to step A3.

[0082] A3: Adopting a performance-constrained optimization method, the task resource allocation scheme is dynamically optimized by solving the optimization problem so that the task performance is greater than the preset minimum performance threshold.

[0083] When working in this example, a multi-dimensional performance evaluation system is constructed to comprehensively reflect the operating status of the system. Based on the loss aversion principle and the optimal solution of performance constraints, it can effectively prevent significant performance degradation during the continuous learning process of the large model, provide reliable performance guarantees for the continuous learning process of the large model, and also provide important decision-making basis for subsequent neuron allocation strategies.

[0084] During specific implementation, in step A3, an optimization method based on performance constraints is adopted. By solving the optimization problem so that the task performance is greater than the preset minimum performance threshold to dynamically optimize the task resource allocation scheme, the following steps are adopted: determining the task objectives and constraints, modeling the problem, constructing the optimization function through the Lagrange multiplier method, solving the optimal conditions, using binary search combined with iterative approximation to accelerate the solution, obtaining candidate solutions, and making an optimal judgment on the candidate solutions based on the preset optimal judgment basis. The optimal judgment basis includes at least a judgment basis on whether the marginal benefit ratio of resources is close to the threshold, a judgment basis on whether the performance index is close to the minimum standard, and a judgment basis on the stability of the scheme.

[0085] During the actual execution process, the resource management mechanism completes the dynamic adjustment of resources through the steps of continuous monitoring, efficiency calculation, triggering reallocation, and executing new plans. The system calculates the amount of resource allocation by combining methods such as performance gap ratio method, task-specific learning curve prediction and segmented compensation strategy to ensure that resource allocation is accurately matched with performance improvement requirements. The selection of these methods or the adjustment of parameters can be adjusted based on the effective historical strategies recommended for similar performance gaps or task types in the experience memory library.

[0086] Preferably, to ensure the stability of resource management, a buffer mechanism can be introduced to avoid frequent adjustments by setting adjustment intervals and amplitude limits, and to keep reserved resources to cope with sudden demands.

[0087] As a further improvement of this embodiment, when processing concurrent tasks, a priority queue can also be used to manage resource requests, and priorities can be assigned based on a multi-dimensional scoring mechanism such as task importance, urgency, and expected benefits. Dynamic priority decay can be combined to prevent starvation, and resource channels and resource occupancy feedback mechanisms can be used to ensure that critical tasks obtain resources first. The weight of the task priority and the parameters of the decay function can be fine-tuned based on historical experience to optimize the overall throughput or critical task response time under specific mixed workloads.

[0088] See also Figure 3 , preferably, by evaluating the importance of neurons and introducing the degree of contribution of neurons to the performance of the current task, the computing resources and update frequency of each neuron are dynamically allocated in layers according to the importance of each neuron. The following steps are adopted:

[0089] B1: By monitoring the performance data stream of the large model during continuous learning, the activation frequency of neurons, the strength of connection weights, and the task relevance factor that reflects the contribution of neurons to the current task are obtained, and the importance of neurons is calculated;

[0090] B2: Divide the range according to the preset importance, for example, high: 50-60%, medium: 30-40%, low: 10-20%, and set neurons to match the resource allocation priority;

[0091] B3: Based on the resource allocation priority, the preset adjustment mechanism is called to hierarchically allocate the computing resources and update frequency of each neuron. The adjustment mechanism includes a performance elasticity adjustment mechanism based on elastic coefficient allocation, a task characteristic adjustment mechanism that sets allocation templates for different task types, a periodic allocation adjustment mechanism that reserves a preset proportion of exploration time for periodic resource experiments, and an intra-level allocation adjustment mechanism for implementing refined allocation within each importance division range.

[0092] When this example is working, by quantifying the contribution of neurons to the current task, dynamic resource allocation is achieved based on neuron importance evaluation, ensuring that key neurons can obtain sufficient resources in a timely manner, achieving accurate allocation and dynamic adjustment of neuron resources, and ensuring that the system maintains efficient resource utilization during the continuous learning process.

[0093] Example 3:

[0094] This embodiment provides a neuron allocation strategy to implement a large model continuous learning method. The similarities with other embodiments are not repeated here, and the differences are described in detail below.

[0095] In this embodiment, the following steps are also included: constructing a current scenario vector, triggering and / or actively performing an associated query in a preset experience memory library, and outputting and executing matching response strategies under similar historical scenarios based on scenario similarity calculation.

[0096] Preferably, the experience memory bank is established by adopting the following steps to obtain historical scenarios, corresponding response strategies and subsequent actual effects, and the experience memory bank is stored through a preset hybrid data storage solution.

[0097] In specific implementation, historical experience fragments are structuredly stored through the experience memory module 3. When storing, a hybrid data storage solution can be adopted. For example, key-value storage is used to quickly index and retrieve simple scenarios based on task IDs or status codes, and a time series database or graph database is used to store detailed experience fragments containing complex state transitions and multi-dimensional features. An example is (T_features, D_features, S_pre_vector, A_params, P_delta_metrics, S_post_vector), where T_features includes task type coding, data scale, etc. D_features is used to describe the attributes or feature sets of the data processed by the task, and is used to quantify the characteristics, scale, distribution or complexity of the data. A_params is the parameters and configurations used when executing the action. P_delta_metrics represents the performance change index, which is used to quantify the impact of the action on the system performance. S_pre_vector and S_post_vector It is a quantified system state feature vector. During update and maintenance, a storage strategy based on importance and redundancy is implemented. It can only store experience fragments that lead to significant performance changes, that is, experience fragments that exceed the threshold, or experience fragments that represent novel situations, that is, experience fragments whose similarity with existing situations in the library is lower than the threshold. At the same time, a time window combined with importance sampling mechanism can be used to regularly archive or remove old, low-importance records to control storage overhead.

[0098] During retrieval, in order to achieve efficient retrieval, it is necessary to provide an efficient retrieval interface that supports context vector similarity, for example, using cosine similarity or k-NN algorithm to search in the context feature space, or structured query, for example, querying the historical best allocation strategy A_params under a specific task type.

[0099] When this example is working, by integrating into the experience memory library and constructing the current scenario vector, it triggers and / or actively performs associated queries in the preset experience memory library, thereby outputting matching response strategies under similar historical scenarios based on scenario similarity calculations, which can achieve more intelligent resource allocation and improve the efficiency of system resource utilization. By accumulating historical experience through the establishment of an experience memory library, it is possible to fully apply historical experience while also predicting potential risks, thereby further making forward-looking adjustments and achieving more intelligent resource allocation effects.

[0100] Preferably, when constructing the current scenario vector and triggering and / or actively performing an associated query in a preset experience memory library, the following steps are adopted: obtaining current state information, constructing a standardized scenario vector, where the current state information includes key dimensions such as task characteristics, system resource status, and current performance indicators, actively querying the experience memory library, and using a predefined scenario similarity function to retrieve the most relevant historical experience fragments. In specific implementation, a k-NN search based on the Faiss library or hash matching in a specific scenario can be used to achieve efficient similarity calculation, and at the same time output a similarity score for each historical experience fragment.

[0101] See also Figure 4 In specific implementation, triggering and / or actively performing the associated query work in the preset experience memory library, outputting the matching response strategies under similar historical scenarios based on scenario similarity calculation and executing them, the following steps are adopted:

[0102] C1: Obtain the activation frequency of the current neuron, the connection weight strength, and the task relevance factor used to reflect the contribution of the neuron to the current task, and calculate the importance of the neuron. In specific implementation, in step C1, the following steps are also included: obtain the task relevance factor of the current neuron and the historical accumulated importance, and correct the importance of the current neuron to obtain the final neuron importance, which can achieve further correction of the neuron and facilitate subsequent decision fusion work;

[0103] C2: Analyzes the retrieved historical experience fragments and predicts the performance changes corresponding to different candidate allocation strategies. During implementation, a preset pre-trained lightweight prediction model can be called to output the predicted performance change results for each candidate allocation strategy.

[0104] C3: Obtain the allocation strategy that has worked best in similar situations in history as a strong recommendation, and output the resource scheduling plan based on the importance of the current neuron and execute it.

[0105] Preferably, in step C3, the allocation strategy with the best effect in similar situations in history is obtained as a strong suggestion, and the resource scheduling plan is output and executed in combination with the importance of the current neuron. The following steps are adopted to obtain the allocation strategy with the best effect in similar situations in history as a strong suggestion, and perform decision fusion in combination with the importance of the current neuron. In specific implementation, the fusion mechanism can adopt clear rules or a weighted model. For example, when the prediction result of the current allocation strategy is significantly better than the default strategy and the prediction execution degree is high, the current allocation strategy can be adopted and the hierarchical division based on the importance of the current neuron can be adjusted accordingly. In the reverse direction, the allocation is mainly based on the importance of the current neuron, but the hierarchy of the boundary neurons can also be fine-tuned according to the prediction result corresponding to the current allocation strategy. When working, the fusion function used in the fusion can be a rule-based system, for example, a system that prioritizes high-confidence predictions or a learned function. For cold start or encountering a completely new situation, that is, when the similarity is low, the fusion mechanism automatically reduces its dependence on the memory library, focuses on the current evaluation, and outputs the fused strategy as the final resource scheduling plan and executes it.

[0106] Preferably, in step C3, the following steps are also included, recording the complete decision-making process and storing it in the experience memory library, the storage content includes the current situation, retrieved historical experience fragments and corresponding prediction results, resource scheduling plans and subsequent actual effects.

[0107] Preferably, the dynamic adjustment coefficient when calculating resource consumption indicators and the weight coefficient when calculating performance evaluation indicators can be initialized or constrained based on successful experiences of similar historical tasks or system load scenarios retrieved from the experience memory library. Furthermore, the performance tolerance range can also be dynamically adjusted based on historical experience fragments in the experience memory library, thereby improving the effectiveness and pertinence of the adjustment and realizing active risk avoidance based on experience.

[0108] This embodiment, when operating, achieves precise, intelligent allocation and dynamic adjustment of neuron resources based on the aforementioned resource management and performance protection mechanism evaluation results, in combination with the predictive analysis and recommendations of the experience memory library. This strategy ensures efficient, intelligent, and adaptive resource utilization during the continuous learning process through precise assessment of neuron importance, utilization of historical experience based on clear similarity calculations and prediction models, and clear decision fusion logic.

[0109] Similarly, in terms of concurrent task processing, a task priority-aware resource allocation mechanism can also be adopted. The dynamic scoring weight of task priority and the resource quota adjustment strategy can also be optimized by querying and analyzing historical data on mixed workload processing in the experience memory library. For new task learning scenarios, through progressive resource allocation strategies, the size of the initial resource pool, the hierarchical division ratio, and the subsequent resource growth rate can all be set and adjusted based on the historical successful strategies in similar new task startup scenarios retrieved from the experience memory library.

[0110] In specific implementation, it is necessary to implement an adaptive resource reallocation mechanism in terms of dynamic adjustment, update the importance by monitoring the changes in neuron performance contribution, and when the changes are significant, for example, exceeding the threshold and the interval is greater than the preset minimum time, trigger the above-mentioned complete decision-making process including situational awareness, prediction and fusion to reallocate resources.

[0111] Example 4:

[0112] See also Figure 1 This embodiment provides a large-model continuous learning system using a neuron allocation strategy. The large-model continuous learning method is implemented by applying a neuron allocation strategy as described in the above embodiment, including:

[0113] Cognitive resource management module 1, used for comprehensive real-time monitoring and output of decision-making basis;

[0114] Performance protection mechanism adjustment module 2 is used to monitor the performance of large models in real time and output optimization suggestions through an asymmetric response mechanism to make timely adjustments to large model performance changes;

[0115] Experience memory module 3, used to store historical experience fragments for retrieval when formulating resource allocation plans;

[0116] Neuron resource dynamic allocation module 4, used to generate and execute resource allocation plans based on decision-making basis and optimization suggestions;

[0117] Knowledge integration module 5 is used for continuous learning of large models and the effective integration of new and old knowledge under the premise of ensuring resource constraints;

[0118] The cognitive resource management module 1, the performance protection mechanism adjustment module 2, the experience memory library module 3, the neuron resource dynamic allocation module 4 and the knowledge integration module 5 are data connected, and the monitoring end of the cognitive resource management module 1 and the monitoring end of the performance protection mechanism adjustment module 2 are both connected to the knowledge integration module 5 to monitor their resource status data flow and performance data flow.

[0119] When working in this example, a systematic information transmission scheme including experience feedback is adopted. A structured experience data stream is introduced on the basis of the original performance data stream, resource status stream, scheduling instruction stream, and knowledge status stream to ensure the coordinated operation of all links in the system. While ensuring the effective implementation of the overall function and the improvement of intelligence, it can also perform experience-based intelligent prediction and adaptation. Through the collaborative work of various modules, the system's resource optimization allocation in a dynamic environment is highly intelligent, robust, and forward-looking, providing a more reliable guarantee for the long-term stable and efficient operation of the system.

[0120] During operation, the cognitive resource management module 1 provides real-time resource status information and constructs a context vector. The performance protection mechanism adjustment module 2 outputs optimization suggestions based on the performance monitoring results and the quantified historical experience similarity analysis. The experience memory module 3 interacts with other modules to perform structured experience storage and efficient retrieval. The neuronal resource dynamic allocation module 4 integrates various aspects of information, especially combines clearly defined similarity calculations, prediction models and fusion logic to generate and execute scheduling plans. The knowledge integration module 5 maintains the stability of the knowledge state under a stable resource environment. Preferably, the interaction frequency between modules is designed to be configurable. For example, routine decisions may rely only on lightweight queries, while major status changes trigger in-depth analysis and prediction.

[0121] The beneficial technical effects of this embodiment include: the present invention can significantly improve the system resource utilization efficiency and effectively reduce computing overhead through a dynamic resource management mechanism and a resource allocation strategy based on importance. Through real-time monitoring of large model performance and computing resources, resource allocation can be dynamically adjusted according to the actual training effect, thereby achieving optimal adjustment of overall efficiency. At the same time, the knowledge protection mechanism based on importance and the progressive knowledge integration method can significantly enhance the transfer learning effect of the model between tasks in different fields, achieve an asymmetric response to changes in system performance, shorten the adaptation time to new tasks, and have stronger generalization capabilities.

[0122] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Those skilled in the art will understand that the present invention includes, but is not limited to, the contents described in the drawings and the above specific embodiments. Any modifications that do not deviate from the functional and structural principles of the present invention are intended to be included within the scope of the claims.

Claims

1. A method for implementing continuous learning of large models using a neuron allocation strategy, characterized in that: The following steps are involved: The large model adopts a multi-level knowledge storage structure. When updating knowledge, it protects knowledge units through a selective protection mechanism based on importance. When a large model is continuously learning, the performance and computing resources of the large model are monitored in real time. Based on the real-time performance monitoring information and resource status information, the task resource allocation plan is dynamically optimized and executed; By evaluating the importance of neurons and introducing the degree of contribution of neurons to the performance of the current task, the computing resources and update frequency of each neuron are dynamically allocated in layers according to the importance of each neuron. The specific steps are as follows: B1: By monitoring the performance data stream of the large model during continuous learning, the activation frequency of neurons, the strength of connection weights, and the task relevance factor that reflects the contribution of neurons to the current task are obtained, and the importance of neurons is calculated; B2: Divide the range according to the preset importance and set the neurons to match the resource allocation priority; B3: Based on the resource allocation priority, the preset adjustment mechanism is called to hierarchically allocate the computing resources and update frequency of each neuron. The adjustment mechanism includes a performance elasticity adjustment mechanism based on elastic coefficient allocation, a task characteristic adjustment mechanism that sets allocation templates for different task types, a periodic allocation adjustment mechanism that reserves a preset proportion of exploration time for periodic resource experiments, and an intra-level allocation adjustment mechanism for implementing refined allocation within each importance division range.

2. The method for implementing large-scale continuous learning using a neuron allocation strategy according to claim 1, characterized in that: The multi-level knowledge storage structure includes a basic knowledge layer for storing the general extraction capabilities of large models, a task-specific layer for storing professional knowledge related to specific tasks, and a cross-task relationship layer for recording the knowledge transfer relationship between different tasks.

3. The method for implementing large-scale continuous learning using a neuron allocation strategy according to claim 1, characterized in that: When protecting knowledge units through a selective protection mechanism based on importance, the following steps are adopted to obtain the verification performance, usage frequency and timeliness of the knowledge unit, and the importance of the knowledge unit is output through a weighted summation method. When the importance of a knowledge unit is greater than a preset threshold, the knowledge unit is protected by adjusting the learning rate and update threshold.

4. The method for implementing large-scale continuous learning using a neuron allocation strategy according to claim 1, characterized in that: When monitoring the performance and computing resources of a large model in real time, the following steps are used: by monitoring the resource status data stream during the continuous learning of the large model, a number of resource status parameters corresponding to each task are obtained, and real-time resource status information is obtained through calculation. The resource status parameters include at least one of memory usage, processor load, and energy consumption indicators; By monitoring the performance data flow of the large model during continuous learning, several performance evaluation indicators corresponding to the large model are obtained, and real-time performance monitoring information is obtained through calculation. The performance evaluation indicators include at least one of accuracy, recall rate, task completion time and resource utilization.

5. The method for implementing large-scale continuous learning using a neuron allocation strategy according to claim 1, characterized in that: To dynamically optimize and execute task resource allocation based on real-time performance monitoring information and resource status information, the following steps are taken: A1: Obtain real-time performance monitoring information, set dynamic adjustment coefficients based on the acquired resource status parameters and a preset coefficient adjustment mechanism, and calculate resource consumption information using a weighted summation method. A2: Calculate the efficiency of the current task resource allocation plan based on the performance monitoring information. If the efficiency of the current task resource allocation plan is lower than a preset threshold, proceed to step A3. A3: Adopting a performance-constrained optimization method, the task resource allocation scheme is dynamically optimized by solving the optimization problem so that the task performance is greater than the preset minimum performance threshold.

6. The method for implementing large-scale continuous learning using a neuron allocation strategy according to claim 1, characterized in that: The method also includes the following steps: constructing a current scenario vector, triggering and / or actively performing an associated query in a preset experience memory library, and outputting and executing matching response strategies under similar historical scenarios based on scenario similarity calculation.

7. The method for implementing large-scale continuous learning using a neuron allocation strategy according to claim 6, characterized in that: The experience memory bank is established by the following steps: obtaining historical scenarios, corresponding response strategies and subsequent actual effects, and storing the experience memory bank through a preset hybrid data storage solution.

8. The method for implementing large-scale continuous learning using a neuron allocation strategy according to claim 1, characterized in that: When a large model continues to learn, the following steps are also included: obtaining representative historical samples through the importance sampling method, establishing a dynamically updated experience pool, selecting samples from the experience pool for playback training through a preset memory playback cycle, and enhancing the retention of important knowledge.

9. A neuron allocation strategy for implementing a large model continuous learning system, applying a neuron allocation strategy as described in any one of claims 1 to 8 to implement a large model continuous learning method, characterized in that: include: Cognitive resource management module (1), used for comprehensive real-time monitoring and output of decision-making basis; Performance protection mechanism adjustment module (2), used to monitor the performance of the large model in real time and output optimization suggestions through an asymmetric response mechanism to make timely adjustments to the performance changes of the large model; Experience memory module (3), used to store historical experience fragments for retrieval when formulating resource allocation plans; Neuron resource dynamic allocation module (4), used to generate and execute resource allocation plans based on decision-making basis and optimization suggestions; Knowledge integration module (5), used for continuous learning of large models and effective integration of new and old knowledge under the premise of ensuring resource constraints; The cognitive resource management module (1), the performance protection mechanism adjustment module (2), the experience memory library module (3), the neuron resource dynamic allocation module (4) and the knowledge integration module (5) are data-connected. The monitoring end of the cognitive resource management module (1) and the monitoring end of the performance protection mechanism adjustment module (2) are both connected to the knowledge integration module (5) to monitor their resource status data flow and performance data flow.

Images