User-guided neural network model dynamic training control method and system

Abstract

The invention relates to the technical field of neural network training, and discloses a user-guided neural network model dynamic training control method and system, and the method comprises the steps: receiving a training intention defined by a user, and enabling the training intention to encode a training priority and a safety condition in a structured constraint form; in the training iteration process of the neural network model, structured training telemetry information is generated and summarized in real time; inputting the training intention and the structured training telemetry information into a constrained training controller based on a language model; reasoning by the training controller based on the training intention and the structured training telemetry information, and determining a control action during training; according to the method, static optimization of the training process of the neural network is changed into intelligent control guided by the safety intention of the user in real time, and the safety, the stability and the decision transparency of the neural network are remarkably improved while the performance of the model is guaranteed.

Description

Technical Field

[0001] This invention relates to the field of neural network training technology, and in particular to a user-guided dynamic training control method and system for neural network models. Background Technology

[0002] In the field of artificial intelligence, the performance of deep neural networks is highly dependent on their training process. Currently, the mainstream training paradigm is essentially static optimization based on gradient descent, which iteratively updates network parameters by minimizing a pre-defined, fixed loss function. While effective, this paradigm's limitations are becoming increasingly apparent when facing applications requiring high reliability and security.

[0003] First, the training process is severely disconnected from high-level security intent. Traditional loss functions (such as cross-entropy), as simplified mathematical proxies, fail to adequately express the complex, non-differentiable constraints and priorities of users in real-world scenarios. These constraints include stringent suppression of specific error types (such as false positives), rigid requirements for model prediction and calibration stability, or the mandatory maintenance of validated core capabilities during continuous learning. This makes training a "blind" process that merely optimizes mathematical objectives, failing to directly address human security concerns and domain rules.

[0004] Secondly, the training process lacks a real-time, targeted intelligent control mechanism. Existing processes are open-loop or simple feedback mechanisms, with strategies (such as learning rate decay) typically predefined. When training telemetry data reveals specific risk patterns (such as an abnormal confidence distribution in a certain type of sample), the system cannot proactively intervene to perform diagnostic or remedial actions. Therefore, the evolution of the model representation space is implicit and uncontrollable, potentially introducing untraceable security risks.

[0005] To address these challenges, existing research directions have significant shortcomings. Automated machine learning primarily focuses on automating architecture or hyperparameter search, and its optimization objectives remain limited to traditional performance metrics, failing to incorporate safety constraints as core optimization variables. Methods using reinforcement learning to control training suffer from instability in policy network training, opaque decision logic, and difficulty in seamlessly integrating prior human knowledge as hard constraints. Furthermore, research on modular or dynamic networks largely concentrates on structural adaptation during the inference phase, without directly addressing the dynamic and safety-oriented regulation issues inherent in the training phase itself.

[0006] In summary, current technologies lack a framework that can directly translate explicit human intentions and safety constraints into real-time, targeted adjustments to the model's internal representation capacity in an interpretable and auditable manner during the training phase. This results in a rigid, opaque training process for neural networks facing safety-critical tasks, with potential risks difficult to manage. Therefore, the industry urgently needs a fundamental paradigm shift to reshape model training from a static numerical optimization problem into a dynamic, controllable, and safe intelligent system-guided process under explicit human guidance. Summary of the Invention

[0007] The purpose of this invention is to address the shortcomings of the existing technology by proposing a user-guided neural network dynamic training control method and system. Its core is to use a constrained large language model as a controller, and based on the user-defined security intent and training telemetry data, dynamically and safely regulate the model training process through atomic actions such as filter synthesis, creation of temporary branches and parameter freezing, without affecting the final inference model.

[0008] On the one hand, a user-guided dynamic training control method for neural network models is provided, including the following steps: Receive a user-defined training intent, which encodes one or more training priorities and security conditions in the form of structured constraints; During the training and iteration process of the neural network model, structured training telemetry information reflecting the model's performance and safety-critical failure modes is generated and summarized in real time. The training intent and the structured training telemetry information are input into a constrained, language model-based training controller. The training controller infers based on the training intent and the structured training telemetry information, and determines the control actions during training from a predefined finite set of actions; During the training iteration, certain training control actions are executed, and the training process is adjusted based on the execution results, wherein the execution of all control actions does not change the logic of the neural network model in the inference phase.

[0009] Furthermore, the constrained, language model-based training controller is configured not to directly access the original training data, gradient data, label data, and inference phase sample-level output data of the neural network model, and its decisions are based solely on the training telemetry information summary after the training process has been summarized, abstracted, and structured, as well as the user-defined training intent.

[0010] Furthermore, the control actions in the predefined finite set of actions include filter synthesis, temporary branch creation, and parameter freezing.

[0011] Specifically, the filter synthesis includes: A constrained linear fusion operation is performed on at least two existing convolutional filters in the neural network model to synthesize a new convolutional filter, wherein the linear fusion operation is implemented by weighted summation, expressed as: ; in, This represents the newly synthesized filter. Indicates the first An existing filter, Indicates corresponding to The fusion coefficient, and the fusion coefficient The value of is constrained to ensure that Located in existing filters Within the represented space.

[0012] Preferably, during the filter synthesis, the system creates and stores spectral information, which at least records information used for synthesizing new filters. Each of the pre-drive filters The identifier and its corresponding fusion coefficient This allows for full traceability of filter evolution during the training process.

[0013] Furthermore, the creation of the temporary branch specifically includes: After determining to perform the filter synthesis action, the system creates a temporary training branch independent of the main training trajectory and applies the newly synthesized convolutional filter... Introduce this into the model corresponding to this branch; The temporary training branch uses the same training data and optimizer as the main training branch, and is trained and evaluated independently within a preset bounded evaluation period.

[0014] Preferably, it also includes an evaluation and merging decision for the temporary training branches: After the bounded evaluation period ends, a set of performance metrics is obtained by weighting according to the user-defined training intent, and the model performance of the temporary training branch is evaluated based on the performance metrics. The model parameter changes in the temporary training branch are merged into the main training trajectory only when the evaluation results meet the preset acceptance criteria; otherwise, the temporary training branch is terminated and discarded, so that the main training trajectory remains unchanged.

[0015] Furthermore, the parameter freezing includes: When the structured training telemetry information indicates that a specific parameter or subset of parameters in the neural network model has a stable positive contribution to achieving the target specified in the training intent, the training controller issues an instruction to mark the specific parameter or subset of parameters as protected parameters. In subsequent training iterations, the values ​​of the protected parameters will be frozen and will no longer participate in gradient updates.

[0016] Furthermore, during the training control actions, adjustments to the model structure or training process are performed in a reversible and rollback manner, and can be restored to the state before execution if the preset acceptance conditions are not met.

[0017] In addition, the controlled actions during training also include: Within a predefined safety range, one or more hyperparameters of the optimizer used to train the neural network model are subject to constrained adjustments. The hyperparameters include at least one of the learning rate, weight decay, or momentum parameters, and the adjustments do not change the optimizer's basic update rules.

[0018] Preferably, the training controller is located in a control layer independent of the numerical optimization calculation of model parameters, and is used to make decisions on the set of behaviors allowed to be executed during training without participating in gradient calculation or weight update, thereby covering the control implementation of non-language models but having equivalent decision-making capabilities.

[0019] Furthermore, the training intent is encoded in the form of structured constraints to encode one or more training priorities and security conditions, enabling the training controller to weight or prioritize multiple performance or security metrics based on user intent before making decisions, so as to prevent others from circumventing intent guidance through fixed rules or static thresholds.

[0020] Furthermore, the structured training telemetry information includes at least one of the following types of metrics: Statistics on false positive and false negative rates for misclassification, distribution calibration stability measures for model output confidence, entropy measures for model prediction confidence, and variance measures for performance metrics across multiple training epochs.

[0021] Furthermore, the method also includes recording audit logs: Training control decisions are recorded in a structured manner to an audit log, which includes: training intent, timestamp, triggering telemetry context, controller decision, details of the action performed, evaluation results, and execution results, forming an auditable training evolution trajectory.

[0022] On the other hand, a user-guided dynamic training control system for neural network models is provided, including: The training intent receiving and parsing module is used to receive user-defined training intents, which are encoded in the form of structured constraints to include one or more training priorities and security conditions. The training telemetry generation and aggregation module is used to generate and aggregate structured training telemetry information that reflects model performance and safety-critical failure modes in real time during the training iteration of the neural network model. The controller input interface module is used to input the training intent and the structured training telemetry information to the constrained, language model-based training controller. The intention reasoning and control decision module is used by the training controller to reason based on the training intention and the structured training telemetry information, and to determine the control action during training from a predefined finite set of actions; The training action execution and scheduling module is used to execute determined training control actions during the training iteration process and adjust the training process based on the execution results, wherein the execution of all control actions does not change the logic of the neural network model in the inference phase.

[0023] In addition, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the user-guided dynamic training control method for neural network models described in any of the preceding claims.

[0024] Meanwhile, an electronic device is provided, comprising: one or more processors; and a storage device for storing one or more programs, which, when executed by the one or more processors, cause the one or more processors to implement the user-guided neural network model dynamic training control method described in any of the preceding claims.

[0025] Compared with the prior art, the beneficial effects of the present invention are: This invention receives and parses user-defined training intentions, transforming traditionally difficult-to-quantify security concerns, compliance requirements, and business priorities into structured constraints that the controller can directly follow. Decisions are made under these intention constraints, eliminating the risk of training behavior deviating from human values ​​and safety red lines. This solves the core problem of the difficulty in embedding and guaranteeing security objectives in traditional methods. This invention introduces a constrained language model-based training controller as the decision core, which utilizes its powerful semantic understanding and reasoning capabilities to perform in-depth analysis of structured training telemetry information. This achieves a leap from "conditional reflex" adjustment that relies on predefined rules or simple feedback to "intelligent scheduling" based on comprehensive reasoning of multi-source information, significantly enhancing the training system's ability to cope with complex and dynamic risks. All training control actions executed under the command of the controller in this invention are designed with safety buffers and rollback mechanisms. At the same time, the parameter freezing action can actively protect the verified core representations and prevent performance regression, so that the model capacity can evolve in a directional and non-destructive manner within the safety boundary, which greatly improves the reliability of the output model during the training process. Every decision (basis, action, result) in the entire control process of this invention is fully recorded, forming an auditable trajectory, realizing the traceability and interpretability of the training evolution history, meeting the requirements of high-standard compliance review. In addition, it clearly limits all actions to not changing the logic of the model in the inference stage, eliminating the need for a customized inference engine or complex runtime support, greatly reducing the deployment threshold and integration cost. Attached Figure Description

[0026] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a user-guided dynamic training control method for a neural network model according to the present invention. Figure 2 This is a block diagram of a user-guided neural network model dynamic training control system according to the present invention. Figure 3 This is a schematic diagram of an embodiment of an electronic device according to the present invention. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0028] The specific embodiments of the present invention will be described below with reference to the accompanying drawings and examples.

[0029] Example 1 This embodiment sets up a medical image analysis scenario, specifically the automatic classification of benign and malignant nodules in lung CT images. In this scenario, the cost of model errors is highly asymmetric, including critical risks and relatively acceptable risks. Critical risks, such as misclassifying malignant nodules as benign (false negatives), may lead to delayed treatment for patients and serious clinical consequences. Relatively acceptable risks, such as misclassifying benign nodules as malignant (false positives), usually lead to further examinations, which, although increasing costs, can be corrected.

[0030] Therefore, the user's core training intention is to strictly suppress the false negative rate (FNR) while maintaining a high overall accuracy, and to ensure that the model's prediction confidence is reliable (well calibrated).

[0031] In this scenario, the method of this invention drives the LLM controller's decision-making by real-time monitoring of safety telemetry data such as false negative rate and calibration error. When an increase in false negative risk is detected, filter synthesis is automatically triggered to enhance the extraction of malignant features, and the results are validated in a temporary branch. Simultaneously, the system freezes parameters that are crucial for the stable identification of benign nodules to prevent performance regression. Ultimately, guided by the user's "safety first" intention, an image-assisted diagnostic model with a significantly lower false negative rate and more reliable predictions is trained.

[0032] For details, please refer to [link / reference]. Figure 1 The technical solution of the user-guided dynamic training control method for neural network models provided in this embodiment includes the following steps: S1: Receive a user-defined training intent, which encodes one or more training priorities and security conditions in the form of structured constraints; S2: During the training and iteration process of the neural network model, generate and summarize structured training telemetry information that reflects the model performance and safety-critical failure modes in real time; S3: Input the training intent and the structured training telemetry information into a constrained, language model-based training controller; S4: The training controller performs reasoning based on the training intention and the structured training telemetry information, and determines the control action during training from a predefined finite set of actions; S5: Execute determined training control actions during the training iteration process and adjust the training process based on the execution results, wherein the execution of all control actions does not change the logic of the neural network model in the inference phase.

[0033] The first step, S1, involves system initialization, where the user defines the training intent and configures the initial model and training data. Specifically, the user intent is expressed as structured, computable constraints. This training intent encodes one or more training priorities and safety conditions in the form of structured constraints, allowing the training controller to weight or prioritize multiple performance or safety metrics based on the user intent before making decisions. This prevents others from circumventing intent guidance through fixed rules or static thresholds. For example, in the medical image classification task (distinguishing between benign and malignant nodules) of this embodiment, the intent can be defined as: json { "primary_objective": "Improves overall accuracy" “safety_constraints”: [ { "type": "asymmetric_error", “description”: “The cost of misclassifying a malignant nodule as benign (missed detection) is far greater than the cost of misclassifying a benign nodule as malignant (false positive).” "tolerance_ratio": 0.1 / / The maximum allowed false negative rate is 10% of the false positive rate. }, { "type": "calibration_stability", “description”: “The confidence level of the model prediction should closely match the actual accuracy.” "expected_calibration_error_max": 0.05 } ], “stability_goal”: { "type": "performance_regression_protection", “description”: “The learned ability to recognize typical benign features should not degenerate.” } } The intent file is parsed by the system and transformed into an internal logical representation that can be queried by the controller.

[0034] Next, in step S2, the structured training telemetry information includes at least one of the following metrics: The calculations include: statistics on false positive and false negative rates for misclassification; a stability measure for the distribution calibration of model output confidence; an entropy measure for model prediction confidence; and a variance measure for performance metrics across multiple training epochs. In this embodiment, the core calculations include: False positive rate (FPR) and false negative rate (FNR) calculated from the validation set by batch or period. The Expected Calibration Error (ECE), which measures the accuracy of confidence levels, is calculated using the following formula: ; in, This represents the number of confidence intervals. To predict the confidence level falling into the first A sample set of intervals, The total number of samples, This represents the actual accuracy of the set. The average prediction confidence level for this set; The confidence entropy, which reflects the certainty of the model's predictions, is calculated using the following formula: ; in, This is the softmax probability vector output by the model. The number of categories is used to count the proportion of high-entropy samples in the training batch, which can help identify potential problem areas with ambiguous classifications.

[0035] These indicators, along with regular indicators, are compiled into periodic telemetry summary reports.

[0036] Next, as described in step S3, the training intent and the structured training telemetry information are input to a constrained, language model-based training controller. This constrained, language model-based training controller is configured not to directly access the original training data, gradient data, label data, and inference-stage sample-level output data of the neural network model. Its decisions are based solely on a summary of the training telemetry information after the training process has been aggregated, abstracted, and structured, as well as the user-defined training intent. The training telemetry information summary may include statistical indicators, distribution characteristics, stability measures, or security-related indicators, but does not contain information that can reconstruct the prediction or gradient of a single sample.

[0037] In this embodiment, the training controller is implemented by a large language model (LLM) that has been fine-tuned by instructions. For example, the prompt word template is as follows: You are a neural network training safety controller. Your decisions must strictly adhere to the following user intent: [Insert the parsed user intent from step 1 here] The telemetry summary for the current training cycle is as follows: - Cycle: Round E - Main loss: L_train - Verify accuracy: Acc_val - **Key Security Indicators**: - False Negative Rate (FNR): fnr_current - False Positive Rate (FPR): fpr_current - FNR / FPR ratio: ratio_current (calculated value) - Expected calibration error (ECE): ece_current - Proportion of samples with high confidence entropy: high_entropy_ratio - Model status: Currently, M parameters are frozen.

[0038] Please analyze the situation and strictly select one action from the following action space to execute: 1. `NO_OP`: Do not perform any operation and continue training.

[0039] 2. `FILTER_FUSE`: Performs filter synthesis to enhance the feature extraction capability for high-entropy samples.

[0040] 3. `CREATE_BRANCH`: Creates a temporary evaluation branch for a potential risk pattern.

[0041] 4. `FREEZE_PARAMETERS`: Freezes parameters that contribute to current stability.

[0042] 5. `ROLLBACK`: Terminates a poorly performing temporary branch.

[0043] Your output must be in JSON format: `{"decision": "action name", "reason": "brief reasoning, referencing intent terms and telemetry data", "parameters": {action-related parameters}}`.

[0044] Finally, in this embodiment, the JSON instructions output by the controller will be passed to the action execution engine.

[0045] Based on the training controller described above, the training control actions in step S4 are determined. The control actions in the predefined finite set of actions include filter synthesis, temporary branch creation, and parameter freezing.

[0046] Specifically, the filter synthesis includes: S401: Perform a constrained linear fusion operation on at least two existing convolutional filters in the neural network model to synthesize a new convolutional filter, wherein the linear fusion operation is implemented by weighted summation, expressed as: ; in, This represents the newly synthesized filter. Indicates the first An existing filter, Indicates corresponding to The fusion coefficient, and the fusion coefficient The value of is constrained to ensure that Located in existing filters Within the represented space.

[0047] S402: During the filter synthesis, the system creates and stores spectrum information, which at least records information used for synthesizing the new filter. Each of the pre-drive filters The identifier and its corresponding fusion coefficient This allows for full traceability of filter evolution during the training process.

[0048] Specifically, in this embodiment, when the controller decision is FILTER_FUSE, the action execution engine performs the following operations: From a certain convolutional layer of the current network, select the two filters with the greatest difference in activation on the samples that lead to high confidence entropy. and As a precursor, a new filter is synthesized through constrained linear fusion. : in, Fusion coefficient ( (constraining it to be somewhere in between) For a very small random noise vector (such as from Sampling is used to break perfect symmetry. This operation ensures... lie in and Interpretability is maintained within the spanned representation subspace.

[0049] Among them, recorded in metadata The genealogy is .

[0050] Furthermore, the creation of the temporary branch in the control action during the training process specifically includes: S403: After determining to execute the filter synthesis action, the system creates a temporary training branch independent of the main training trajectory and applies the newly synthesized convolutional filter... Introduce this into the model corresponding to this branch; S404: The temporary training branch uses the same training data and optimizer as the main training branch, and is independently trained and evaluated within a preset bounded evaluation period.

[0051] It also includes the evaluation and merging decision of the temporary training branches: S405: After the bounded evaluation period ends, a set of performance metrics is obtained by weighting according to the user-defined training intent, and the model performance of the temporary training branch is evaluated based on the performance metrics. S406: The model parameter changes in the temporary training branch are merged into the main training trajectory only when the evaluation result meets the preset acceptance criteria; otherwise, the temporary training branch is terminated and discarded, so that the main training trajectory remains unchanged.

[0052] Specifically, the system automatically creates a file containing... This is a temporary training branch. This branch copies the parameters of the current main model, replaces a minor filter in the original layer with F_new, and trains independently within a bounded evaluation window (e.g., 2 training epochs) using the same optimizer settings as the main branch.

[0053] After the temporary branch evaluation window closes, the system decides whether to merge branches based on evaluation metrics weighted by the user's intent. An example of the weighted evaluation formula is as follows: in, This represents the weighted evaluation score. This indicates the amount of improvement relative to the main branch. , and These represent the weights of the three metrics automatically assigned based on the user's intent (for example, if the intent is to suppress false negatives, then...). (Set to maximum), while also satisfying all hard constraints in the intent (such as...) The ratio is less than 0.1).

[0054] Determine whether the merging requirements are met based on the calculated score. If they are met, the branch will be modified accordingly. The branch and its optimized parameters are merged into the main model. Lineage information is also saved; if the conditions are not met, the branch is terminated and discarded, and the main model state is rolled back to before the branch was created, ensuring that the evolution process is non-destructive.

[0055] In addition to the control actions described above, the parameter freezing action includes: When the structured training telemetry information indicates that a specific parameter or subset of parameters in the neural network model has a stable positive contribution to achieving the target specified in the training intent, the training controller issues an instruction to mark the specific parameter or subset of parameters as protected parameters. In subsequent training iterations, the values ​​of the protected parameters will be frozen and will no longer participate in gradient updates.

[0056] In this embodiment, when telemetry data shows that certain filters consistently make a high contribution to the stable identification of benign samples (corresponding to stability_goal) (judged by calculating the correlation between filter activation and task accuracy), the controller may issue a FREEZE_PARAMETERS instruction. The action engine marks the parameters of these filters as "protected," and their gradients are set to zero in subsequent gradient updates during training.

[0057] During the above training, all adjustments to the model structure or training process are executed in a reversible and rollback manner, and can be restored to the state before execution if the preset acceptance conditions are not met.

[0058] In addition, the controlled movements during training also include: Within a predefined safety range, one or more hyperparameters of the optimizer used to train the neural network model are subject to constrained adjustments. The hyperparameters include at least one of the learning rate, weight decay, or momentum parameters, and the adjustments do not change the optimizer's basic update rules.

[0059] The training controller is located in a control layer independent of the numerical optimization calculation of model parameters. It is used to make decisions on the set of behaviors allowed to be executed during training without participating in gradient calculation or weight update, thereby covering the control implementation of non-language models but having equivalent decision-making capabilities.

[0060] Finally, key information from all the above steps is structured and recorded in the audit log. Each record includes: training intent, timestamp, triggering telemetry context, controller decision, details of the action performed, evaluation result, execution result, and final disposal (merging / discarding). This provides complete traceability for the training process.

[0061] The core of the system, achieved by combining all the above methods, is a closed-loop control architecture centered on a constrained LLM training controller. The entire system is strictly divided into an intent and decision layer and a training execution layer, with the following structure: Figure 2 As shown, the system is clearly divided into an intent and decision layer and a training execution layer. The controller is the intelligent hub connecting the two layers; it does not access the raw data but directs all training actions. User intent and training data are the two main input sources for the system. The structured training telemetry feedback generated by the training execution layer is fed back to the controller. Outputs include the optimized final model and a complete audit log. Simultaneously, after confirming the control actions during training, the action execution engine selects specific action modules (filter synthesis, parameter freezing, temporary training branches) for execution; the dashed lines indicate the selection and scheduling.

[0062] Based on this, the present invention provides a user-guided dynamic training control system for neural network models, comprising: The training intent receiving and parsing module is used to receive user-defined training intents, which are encoded in the form of structured constraints to include one or more training priorities and security conditions. The training telemetry generation and aggregation module is used to generate and aggregate structured training telemetry information that reflects model performance and safety-critical failure modes in real time during the training iteration of the neural network model. The controller input interface module is used to input the training intent and the structured training telemetry information to the constrained, language model-based training controller. The intention reasoning and control decision module is used by the training controller to reason based on the training intention and the structured training telemetry information, and to determine the control action during training from a predefined finite set of actions; The training action execution and scheduling module is used to execute determined training control actions during the training iteration process and adjust the training process based on the execution results, wherein the execution of all control actions does not change the logic of the neural network model in the inference phase.

[0063] It should be noted that the steps in the user-guided neural network model dynamic training control method provided in this embodiment can be implemented based on the corresponding modules in the user-guided neural network model dynamic training control system. Those skilled in the art can refer to the technical solution of the system to implement the steps of the method. That is, the embodiments in the system can be understood as preferred examples of implementing the method, and will not be elaborated here.

[0064] Furthermore, in the embodiments disclosed in this invention, the model representation unit for dynamic training control is not limited to the convolutional filter in a convolutional neural network. Depending on the model architecture, the representation unit may also include, but is not limited to, attention weight substructures in attention mechanisms, feature projection matrices, intermediate representation modules in encoders or decoders, feature channel combination units, or other composable parameter units capable of characterizing the internal feature representation of the model.

[0065] In these implementations, the training controller still performs constrained synthesis, freezing, branch evaluation, or rollback operations on the representation unit based solely on structured training telemetry information and user-defined training intentions, without directly accessing the original sample data or participating in the numerical optimization calculation of model parameters. This allows the method of the present invention to adapt to different types of neural network architectures, including but not limited to attention-based model structures, where the representation unit can be a functional submodule or a black-box module.

[0066] This embodiment also provides an electronic device, such as... Figure 3As shown, the electronic device includes a processor 14 and a memory 13. The memory 13 stores machine-executable instructions that can be executed by the processor 14. The processor 14 executes the machine-executable instructions to implement the aforementioned user-guided neural network model dynamic training control method.

[0067] Furthermore, Figure 3 The electronic device shown also includes a bus 12 and a communication interface 11, with the processor 14, the communication interface 11 and the memory 13 connected via the bus 12.

[0068] The memory 13 may include high-speed random access memory (RAM) and may also include non-volatile memory, such as at least one disk storage device. Communication between this system network element and at least one other network element is achieved through at least one communication interface 11 (which can be wired or wireless), such as the Internet, wide area network, local area network, metropolitan area network, etc. The bus 12 may be an ISA bus, PCI bus, or EISA bus, etc. The bus can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 3 The symbol is represented by a single double-headed arrow, but this does not mean that there is only one bus or one type of bus.

[0069] Processor 14 may be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method can be completed by the integrated logic circuitry in the hardware of processor 14 or by instructions in software form. Processor 14 can be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it can also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in this embodiment. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this embodiment can be directly manifested as execution by a hardware decoding processor, or execution by a combination of hardware and software modules in the decoding processor. The software modules can reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. The storage medium is located in memory 13. The processor 14 reads the information in memory 13 and, in conjunction with its hardware, completes the steps of the user-guided neural network model dynamic training control method.

[0070] This disclosure also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, storing a computer program that, when run on a computer, causes the computer to execute the steps of a user-guided dynamic training control method for a neural network model.

[0071] Finally, it should be noted that the above description is only a preferred embodiment of the present invention, and the scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be pointed out that for those skilled in the art, any improvements and modifications made without departing from the principle of the present invention should also be considered within the scope of protection of the present invention.

[0072] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A user-guided dynamic training control method for a neural network model, characterized in that, Includes the following steps: Receive a user-defined training intent, which encodes one or more training priorities and security conditions in the form of structured constraints; During the training and iteration process of the neural network model, structured training telemetry information reflecting the model's performance and safety-critical failure modes is generated and summarized in real time. The training intent and the structured training telemetry information are input into a constrained, language model-based training controller. The training controller infers based on the training intent and the structured training telemetry information, and determines the control actions during training from a predefined finite set of actions; During the training iteration, certain training control actions are executed, and the training process is adjusted based on the execution results, wherein the execution of all control actions does not change the logic of the neural network model in the inference phase.

2. The user-guided dynamic training control method for neural network models according to claim 1, characterized in that, The constrained, language model-based training controller is configured not to directly access the original training data, gradient data, label data, and inference phase sample-level output data of the neural network model. Its decisions are based solely on the training telemetry information summary after the training process has been aggregated, abstracted, and structured, as well as the user-defined training intent.

3. The user-guided dynamic training control method for neural network models according to claim 1, characterized in that, The control actions in the predefined finite set of actions include filter synthesis, temporary branch creation, and parameter freezing.

4. The user-guided dynamic training control method for neural network models according to claim 3, wherein the filter synthesis specifically includes: A constrained linear fusion operation is performed on at least two existing convolutional filters in the neural network model to synthesize a new convolutional filter, wherein the linear fusion operation is implemented by weighted summation, expressed as: ； in, This represents the newly synthesized filter. Indicates the first An existing filter, Indicates corresponding to The fusion coefficient, and the fusion coefficient The value of is constrained to ensure that Located in existing filters Within the represented space.

5. The user-guided dynamic training control method for neural network models according to claim 4, characterized in that, During the synthesis of the filter, the system creates and stores spectral information, which records at least the information used to synthesize the new filter. Each of the pre-drive filters The identifier and its corresponding fusion coefficient This allows for full traceability of filter evolution during the training process.

6. The user-guided dynamic training control method for neural network models according to claim 5, characterized in that, The creation of the temporary branch specifically includes: After determining to perform the filter synthesis action, the system creates a temporary training branch independent of the main training trajectory and applies the newly synthesized convolutional filter... Introduce this into the model corresponding to this branch; The temporary training branch uses the same training data and optimizer as the main training branch, and is trained and evaluated independently within a preset bounded evaluation period.

7. The user-guided dynamic training control method for neural network models according to claim 6, characterized in that, This also includes the evaluation and merging decisions for the aforementioned temporary training branches: After the bounded evaluation period ends, a set of performance metrics is obtained by weighting according to the user-defined training intent, and the model performance of the temporary training branch is evaluated based on the performance metrics. The model parameter changes in the temporary training branch will only be merged into the main training trajectory if the evaluation results meet the preset acceptance criteria. Otherwise, terminate and discard the temporary training branch, leaving the main training trajectory unchanged.

8. The user-guided dynamic training control method for neural network models according to claim 3, characterized in that, The parameter freezing specifically includes: When the structured training telemetry information indicates that a specific parameter or subset of parameters in the neural network model has a stable positive contribution to achieving the target specified in the training intent, the training controller issues an instruction to mark the specific parameter or subset of parameters as protected parameters. In subsequent training iterations, the values ​​of the protected parameters will be frozen and will no longer participate in gradient updates.

9. The user-guided dynamic training control method for neural network models according to claim 1, characterized in that, During the training process, any adjustments to the model structure or training process are performed in a reversible and rollback manner, and can be restored to the state before execution if the preset acceptance conditions are not met.

10. The user-guided dynamic training control method for neural network models according to claim 1, characterized in that, The controlled actions during training also include: Within a predefined safety range, one or more hyperparameters of the optimizer used to train the neural network model are subject to constrained adjustments. The hyperparameters include at least one of the learning rate, weight decay, or momentum parameters, and the adjustments do not change the optimizer's basic update rules.

11. The user-guided dynamic training control method for neural network models according to claim 1, characterized in that, The training controller is located in a control layer independent of the numerical optimization calculation of model parameters. It is used to make decisions on the set of behaviors allowed to be executed during training without participating in gradient calculation or weight update, thereby covering the control implementation of non-language models but having equivalent decision-making capabilities.

12. The user-guided dynamic training control method for neural network models according to claim 1, characterized in that, The training intent is encoded in the form of structured constraints, which encode one or more training priorities and security conditions. This allows the training controller to weight or prioritize multiple performance or security metrics based on the user's intent before making a decision, thereby preventing others from circumventing intent guidance through fixed rules or static thresholds.

13. The user-guided dynamic training control method for neural network models according to claim 1, characterized in that, The structured training telemetry information includes at least one of the following metrics: Statistics on false positive and false negative rates for misclassification, distribution calibration stability measures for model output confidence, entropy measures for model prediction confidence, and variance measures for performance metrics across multiple training epochs.

14. The user-guided dynamic training control method for neural network models according to claim 1, characterized in that, The method also includes recording audit logs: Training control decisions are recorded in a structured manner to an audit log, which includes: training intent, timestamp, triggering telemetry context, controller decision, details of the action performed, evaluation results, and execution results, forming an auditable training evolution trajectory.

15. A user-guided dynamic training control system for a neural network model, characterized in that, include: The training intent receiving and parsing module is used to receive user-defined training intents, which are encoded in the form of structured constraints to include one or more training priorities and security conditions. The training telemetry generation and aggregation module is used to generate and aggregate structured training telemetry information that reflects model performance and safety-critical failure modes in real time during the training iteration of the neural network model. The controller input interface module is used to input the training intent and the structured training telemetry information to the constrained, language model-based training controller. The intention reasoning and control decision module is used by the training controller to reason based on the training intention and the structured training telemetry information, and to determine the control action during training from a predefined finite set of actions; The training action execution and scheduling module is used to execute determined training control actions during the training iteration process and adjust the training process based on the execution results, wherein the execution of all control actions does not change the logic of the neural network model in the inference phase.

16. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the user-guided dynamic training control method for neural network models as described in any one of claims 1-14.

17. An electronic device, characterized in that, include: One or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to implement the user-guided dynamic training control method for neural network models as described in any one of claims 1-14.

Images