A multi-task large model training system and method of adaptive incremental learning

Abstract

The application relates to the technical field of artificial intelligence, and discloses a multi-task large model training system and method for self-adaptive incremental learning. The system comprises a knowledge cell library, a dynamic knowledge covariant graph, a path planning module, a forward calculation and parameter updating module, and a structure evolution module. The method comprises the following steps: receiving a task, planning an optimal calculation path and a counterfactual path on the dynamic knowledge covariant graph; driving input to be calculated along the two paths, and updating parameters of knowledge cells involved in the paths based on a composite loss comprising a task loss and a counterfactual verification loss; and when performance stagnation is determined, deriving and specifically training new knowledge cells to realize self-adaptive updating of a model knowledge system. Through modular large model capability units, dynamic planning of an optimal calculation path, self-adaptive expansion of a model structure, and introduction of a counterfactual mechanism, incremental learning, reduction of calculation overhead, and system optimization are realized.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to an adaptive incremental learning multi-task large model training system and method. Background Technology

[0002] In recent years, large-scale artificial intelligence models, represented by pre-trained language models, have demonstrated outstanding performance in various complex tasks such as natural language processing and computer vision due to their powerful representation learning capabilities, and have become the core driving force for the development of artificial intelligence technology.

[0003] The training process of these large models typically relies on a static, large-scale dataset for one-time pre-training to build a general knowledge base. However, real-world applications are dynamic, and models need to continuously adapt to new tasks or learn incremental knowledge. In this process, existing technical approaches have revealed inherent limitations. When using traditional full-scale fine-tuning methods, models not only face high computational resource consumption but also generally suffer from the risk of catastrophic forgetting, meaning that while learning new knowledge, the model significantly impairs its performance already mastered on older tasks.

[0004] To alleviate this problem, although techniques such as Parameter-Efficient Fine-Tuning (PEFT) have been proposed, the core idea of ​​these methods is still to adjust existing parameters or add a small number of new parameters. They do not fundamentally solve the problem of insufficient inherent knowledge capacity and structural complexity of the model when facing new tasks that are significantly different from the existing knowledge system.

[0005] At a deeper level, existing models, whether monolithic or modular (MoE) structures, typically have a fixed network topology before training begins. This means that when a model encounters performance bottlenecks due to its own structural limitations during the learning process, there is a lack of an intrinsic mechanism that can autonomously diagnose the source of the bottleneck and dynamically expand its structure and generate new functional modules as needed.

[0006] Therefore, this invention proposes an adaptive incremental learning multi-task large model training system and method to address the shortcomings of existing technologies. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides an adaptive incremental learning multi-task large model training system and method, which solves the problem that large models, when performing incremental learning, are difficult to balance between avoiding catastrophic forgetting and efficiently learning new knowledge, and that their fixed model structure limits their ability to adaptively expand.

[0008] To achieve the above objectives, the present invention provides the following technical solution: an adaptive incremental learning multi-task large model training system, the system comprising:

[0009] The knowledge cell library is used to store multiple knowledge cells that serve as basic capability units.

[0010] Dynamic knowledge covariance graph is used to represent the functional synergistic relationships between knowledge cells in the knowledge cell library in a graph structure.

[0011] The task receiving and reconstruction module is used to reconstruct the task input based on the core knowledge cells in the dynamic knowledge covariance graph after receiving the task input, and generate an abstract input representation.

[0012] The path planning module is used to receive the abstract input representation and plan the optimal computation path and counterfactual path on the dynamic knowledge covariance graph.

[0013] The forward computation and parameter update module is used to drive the abstract input representation to perform computation along the optimal computation path and the counterfactual path respectively, to obtain the optimal path prediction output and the counterfactual path prediction output and calculate the composite loss, and then update the parameters of the knowledge cells on the optimal path prediction output and the counterfactual path prediction output.

[0014] The structural evolution module is used to calculate the task residual based on the optimal path prediction output if the task performance stagnates after the forward calculation and parameter update module updates the parameters, and to directionally differentiate new knowledge cells based on the task residual, add the new knowledge cells to the knowledge cell library, and update the dynamic knowledge covariance graph.

[0015] Preferably, the task receiving and reconstructing module is further configured to:

[0016] Receive task input containing raw task data and task metadata, and parse the task metadata to determine task type information;

[0017] Based on the parsed task type information, a set of core knowledge cells is queried and determined in the dynamic knowledge covariance graph. The query operation includes matching the task type information with the functional tags of each knowledge cell pre-stored in the knowledge cell library.

[0018] The core knowledge cell is invoked to perform collaborative processing on the original task data to generate initial features, and the initial features are output as abstract input representations; wherein, the abstract input representations are high-dimensional feature vectors or tensors.

[0019] Preferably, the path planning module is further used for:

[0020] Based on the abstract input representation generated by the task receiving and reconstruction module of the meta-controller, and with the abstract input representation as the initial state, a search process is performed on the dynamic knowledge covariance graph with the optimization objective of maximizing the expected performance of the task, so as to determine the optimal computation path;

[0021] A counterfactual path with expected performance lower than the optimal computation path is generated using a preset strategy, wherein the preset strategy includes:

[0022] During the search process, one path with the second-best score is retained as the counterfactual path;

[0023] Alternatively, one or more knowledge cells in the optimal computation path can be replaced to generate the counterfactual path.

[0024] Preferably, the forward calculation and parameter update module is further used for:

[0025] The abstract input representation is driven to perform calculations along the optimal calculation path and the counterfactual path, respectively, to obtain the optimal path prediction output and the counterfactual path prediction output;

[0026] Based on the difference between the optimal path prediction output and the task objective, and using the task loss formula, the task loss is calculated. The task loss formula is as follows:

[0027] ;

[0028] In the formula, It is the value of the task loss. It is the actual target value or label of the current task. It is a loss function pre-defined according to the task type. It is the final optimal path prediction output of the cell output;

[0029] Based on the difference between the counterfactual path prediction output and the task objective, and using the counterfactual task loss formula, the counterfactual task loss is calculated. The counterfactual task loss formula is as follows:

[0030] ;

[0031] In the formula, It is the value of the counterfactual task loss. It is the output of counterfactual path prediction;

[0032] Based on the topological centrality and output entropy of knowledge cells in the dynamic knowledge covariance graph, and the entropy constraint loss calculated using the graph-guided entropy constraint loss formula, the entropy constraint loss formula is as follows:

[0033] ;

[0034] In the formula, This is the value of the graph-guided entropy constraint loss. Refers to the set of knowledge cells along the optimal computation path. It is a knowledge cell on the path. It is a cell Topological centrality metric in dynamic knowledge covariance graphs It is a cell The output characteristics, It is the Shannon entropy calculated from the output features;

[0035] Using the task loss and counterfactual task loss as input, the counterfactual verification loss is calculated using the counterfactual verification loss formula. This counterfactual verification loss ensures that the performance of the planned optimal computation path is significantly better than the counterfactual path. The counterfactual verification loss formula is as follows:

[0036] ;

[0037] In the formula, It is the value of counterfactual verification loss. It is a pre-defined nonnegative constant;

[0038] Based on the composite loss formula, the task loss, the graph-guided entropy constraint loss, and the counterfactual verification loss are combined to obtain the composite loss, which is defined as follows:

[0039] ;

[0040] In the formula, It is the value of the composite loss. and It is a preset non-negative hyperparameter used to adjust the weights of each loss term;

[0041] The composite loss employs a gradient backpropagation algorithm to construct the gradients of the trainable parameters within all knowledge cells of the optimal computation path and counterfactual path, and uses a gradient descent optimizer to update the parameters of the knowledge cells of the optimal computation path and counterfactual path.

[0042] Preferably, the calculation of the counterfactual verification loss is used to compare the counterfactual task loss with the task loss, and when the counterfactual task loss exceeds the task loss by no more than a boundary value, the counterfactual verification loss is set to a non-zero penalty value, and when the counterfactual task loss exceeds the task loss by more than a boundary value, the counterfactual verification loss is set to zero.

[0043] Preferably, the structural evolution module is further used for:

[0044] After the forward calculation and parameter update module updates the parameters, if it is determined that the task performance has stalled, the task residual is calculated based on the difference between the optimal path prediction output and the task objective.

[0045] Based on the task residual, among the knowledge cells that constitute the optimal computation path, the knowledge cell most relevant to the task residual is selected as the parent cell.

[0046] The structural evolution module derives new knowledge cells with the same network structure as the parent cell based on the parent cell, and performs targeted training on the new knowledge cells with the optimization objective of minimizing the fitting error of the task residual.

[0047] The new knowledge cells that have completed the targeted training are added to the knowledge cell library, and the node set and edge set of the dynamic knowledge covariance graph are updated.

[0048] Preferably, when the structural evolution module derives a new knowledge cell, the parameters of the new knowledge cell are initialized based on the parameters of the parent cell; and the specialized training is an isolated training phase, in which only the parameters of the new knowledge cell are updated to fit the task residual, while the parameters of the knowledge cells constituting the dynamic knowledge covariance graph remain unchanged.

[0049] This invention also provides an adaptive incremental learning method for training large multi-task models, the method comprising the following steps:

[0050] S1. Store multiple knowledge cells as basic capability units in a knowledge cell library, and use a dynamic knowledge covariance graph to represent the functional synergy relationship between the knowledge cells.

[0051] S2. After receiving the task input, the task input is reconstructed based on the core knowledge cells in the dynamic knowledge covariance graph to generate an abstract input representation.

[0052] S3. Receive the abstract input representation and plan the optimal computation path and counterfactual path on the dynamic knowledge covariance graph;

[0053] S4. Drive the abstract input representation to perform calculations along the optimal calculation path and the counterfactual path respectively to obtain the optimal path prediction output and the counterfactual path prediction output, and calculate the composite loss, and then update the parameters of the knowledge cells that constitute the optimal calculation path and the counterfactual path based on the composite loss.

[0054] S5. After the parameters of the knowledge cell are updated, if the task performance is determined to be stagnant, the task residual is calculated based on the optimal path prediction output, and new knowledge cells are directionally differentiated based on the task residual. The new knowledge cells are added to the knowledge cell library, and the dynamic knowledge covariance graph is updated.

[0055] This invention provides an adaptive incremental learning multi-task large model training system and method. It has the following beneficial effects:

[0056] 1. This invention modularizes the capability units of a large model into independent knowledge cells, and updates parameters only for relevant cells along the optimal computation path when learning a new task, effectively mitigating the catastrophic forgetting problem. When encountering performance bottlenecks that existing knowledge cannot solve, the structural evolution module can directionally derive new knowledge cells to learn new knowledge, thereby achieving incremental learning of the model and continuous expansion of the knowledge base without compromising existing capabilities.

[0057] 2. The path planning module established in this invention can dynamically plan an optimal computational path on a dynamic knowledge covariance graph based on the specific task. This mechanism avoids calling all parameters of the entire large model when processing each task, and instead forms a targeted temporary sub-network composed of a few knowledge cells, which greatly reduces the computational overhead of a single task and improves the system's operating efficiency and resource utilization.

[0058] 3. The structural evolution mechanism proposed in this invention enables adaptive and on-demand growth of model complexity. By analyzing task residuals, this mechanism can accurately identify the weaknesses in the current knowledge base and generate new knowledge cells specifically targeting those weaknesses for compensation. This goal-oriented structural expansion method ensures that model growth is efficient and beneficial, avoiding the resource waste caused by blindly expanding the model size.

[0059] 4. This invention introduces counterfactual paths and counterfactual verification loss, adding an internal comparison and verification step to the model training process. This mechanism ensures that the optimal computational path selected by the path planning module is not only effective but also significantly superior to other alternative paths, thus providing stronger supervision signals for the rationality of path selection. This enhances the stability of the entire training process, ensuring that the path selection strategy learned by the system has higher reliability and generalization ability. Attached Figure Description

[0060] Figure 1 This is a schematic diagram of the structure of the adaptive incremental learning multi-task large model training system of the present invention;

[0061] Figure 2 This is a flowchart of the adaptive incremental learning multi-task large model training method of the present invention;

[0062] Figure 3 This is a schematic diagram illustrating the structural evolution process of the present invention;

[0063] Figure 4This is a schematic diagram of the computer device structure of the present invention.

[0064] Among them, 100, Multi-task large model training system; 110, Knowledge cell library; 120, Dynamic knowledge covariance graph; 130, Task receiving and reconstruction module; 140, Path planning module; 150, Forward computation and parameter update module; 160, Structure evolution module; 170, Computer equipment; 171, Processor; 172, Memory; 173, Storage medium. Detailed Implementation

[0065] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0066] See attached document Figure 1 , Figure 1 This is a schematic diagram of the structure of a multi-task large model training system for adaptive incremental learning according to an embodiment of the present invention. The present invention provides a multi-task large model training system 100 for adaptive incremental learning, which may include: a knowledge cell library 110, a dynamic knowledge covariance graph 120, a task receiving and reconstruction module 130, a path planning module 140, a forward computation and parameter update module 150, and a structure evolution module 160.

[0067] Knowledge Cell Library 110 is a data storage entity whose function is to centrally store multiple knowledge cells. Each knowledge cell is a computational unit containing independent functions and a set of trainable parameters, and can be regarded as a basic capability component that constitutes a large model.

[0068] The dynamic knowledge covariant graph 120 is a graph data structure whose node set consists of knowledge cells from the knowledge cell library 110, and whose edge set represents the functional collaboration and calling relationships between knowledge cells. The structure of this graph is dynamically changing; when the structure evolution module 160 generates new knowledge cells, the node set and edge set of the graph will be updated accordingly.

[0069] The task receiving and reconstruction module 130 has its input end connected to an external task source and its output end connected to the path planning module 140. This module is used to receive external task input and, based on the prior knowledge stored in the dynamic knowledge covariance graph 120, process the original task input to generate a structured abstract input representation.

[0070] The path planning module 140 has its input connected to the output of the task receiving and reconstruction module 130, and its output connected to the forward computation and parameter update module 150. This module is used to receive abstract input representations and plan an optimal computation path and a counterfactual path on the dynamic knowledge covariance graph 120 to perform the current task.

[0071] The forward computation and parameter update module 150 has its input connected to the output of the path planning module 140. This module performs the main training process of the model. It drives the abstract input representation to compute along two paths, obtains the predicted output, calculates a composite loss based on the predicted output and the task objective, and finally updates the parameters of the knowledge cells involved in the path based on this composite loss.

[0072] The structural evolution module 160 has its input connected to the forward computation and parameter update module 150, and its output connected to the knowledge cell library 110 and the dynamic knowledge covariance graph 120. This module is used to expand the model's capabilities by directionally differentiating new knowledge cells when the system determines that the task performance has stalled, and adds the newly generated cells to the knowledge cell library 110, while updating the structure of the dynamic knowledge covariance graph 120.

[0073] The specific steps and structure of the multi-task large model training system 100 and method according to embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0074] See attached document Figure 1 The technical solution of this invention is built upon two fundamental data structures: a knowledge cell library 110 and a dynamic knowledge covariance graph 120. The knowledge cell library 110 is a collection for storing and managing multiple knowledge cells. A knowledge cell is a functionally independent, modular computational unit that encapsulates a specific network structure and a corresponding set of trainable parameters. For example, a knowledge cell can be a complete Transformer encoder layer, a convolutional neural network (CNN) block, or a multilayer perceptron (MLP) network, designed to perform a basic, atomic data transformation or feature extraction function.

[0075] Dynamic knowledge covariant graph 120 is a directed graph structure, denoted as... This is used to mathematically represent the functional collaboration relationships between individual knowledge cells in the knowledge cell library 110. In this graph structure, the node set... Each node in Uniquely corresponds to a knowledge cell in knowledge cell library 110 E represents the edge set symbolizing the functional collaboration and invocation relationships between knowledge cells, and G is a directed graph structure consisting of node set V and edge set E. The edge set... Each directed edge in From node Pointing to node , representing knowledge cells The output can be used as a knowledge cell The inputs of the two have a potential relationship of sequential invocation or functional collaboration.

[0076] In some embodiments, the edges in the dynamic knowledge covariance graph 120 A weight value can be assigned. This weight value is used to quantify knowledge cells. and The effectiveness or probability of collaborative work. This weight value can be updated based on the frequency with which the two cells are jointly activated and generate positive task benefits in historical tasks.

[0077] The dynamic characteristics of this dynamic knowledge covariance graph 120 are reflected in its node set. and edge set It is not fixed, but updated based on the output of the structural evolution module 160. When the structural evolution module 160 successfully derives and trains a new knowledge cell to learn new knowledge, the new cell is added to the knowledge cell library 110. Simultaneously, the node set of the dynamic knowledge covariance graph 120... A new node corresponding to the new cell will be added, and edges will be established from the new node to other related nodes in the graph, or from other related nodes to the new node, thus structurally reflecting the growth of the overall system capability and the evolution of the knowledge system.

[0078] See attached document Figure 1 and Figure 2 In the overall workflow of the multi-task large model training system 100, the task receiving and reconstruction module 130 performs the initial task processing steps. The function of the task receiving and reconstruction module 130 is to receive an externally submitted task input and convert it into an internally unified, structured, abstract input representation for subsequent modules to process.

[0079] Specifically, the task input consists of two parts: raw task data and task metadata. Raw task data is the unstructured or semi-structured data to be processed, such as a piece of text, an image, or an audio clip. Task metadata is descriptive information about the task type, such as a label identifying it as text classification or image object detection.

[0080] Upon receiving task input, the task receiving and reconstructing module 130 first parses the task metadata to identify the specific type of the current task. Then, using the identified task type as an index, the module 130 queries the dynamic knowledge covariance graph 120 to determine a set of core knowledge cells directly related to the function of that task type. In one specific embodiment, each knowledge cell in the knowledge cell library 110 is accompanied by one or more functional tags. The query process involves matching the current task type with the functional tags of the knowledge cells to filter out the corresponding set of core knowledge cells.

[0081] After determining the core knowledge cell set, the task receiving and reconstruction module 130 calls one or more knowledge cells from this set to perform a collaborative preprocessing of the original task data. This collaborative processing process inputs the original task data into the selected core knowledge cells for computation, generating a set of initial features. For example, for a natural language processing task, this process might call a text embedding cell and a syntactic analysis cell to jointly process the input text.

[0082] The output of this collaborative processing is the abstract input representation. This representation is no longer the original task data, but rather a structured data set or a group of data formed after preliminary processing using existing system knowledge, such as a high-dimensional feature vector or tensor. This abstract input representation encapsulates the basic understanding of the original data and is passed to the path planning module 140 as the input basis for its planned computation path.

[0083] See attached document Figure 1 and Figure 2 After the task receiving and reconstruction module 130 generates an abstract input representation, this representation is passed to the path planning module 140. The core of the path planning module 140 is a meta-controller. This meta-controller is a trainable decision unit whose function is to plan two computational paths for executing the current task on the dynamic knowledge covariance graph 120 based on the received abstract input representation.

[0084] The first path is the optimal computational path. The meta-controller performs a search process to determine this path with the optimization objective of maximizing the expected task performance. In one embodiment, this search process is constructed as a sequential decision problem. The meta-controller (e.g., a recurrent neural network or a graph neural network) takes an abstract input representation as its initial state and progressively selects the next knowledge cell to be added to the path on the dynamic knowledge covariance graph 120. In each step, the meta-controller selects one from the neighboring nodes of the current path's end node, with the goal of ensuring that the final complete path (i.e., an ordered combination of knowledge cells) achieves optimal expected performance when performing the task. This search process can be driven by a reinforcement learning algorithm, where the meta-controller, as a policy network, is trained by maximizing the reward signal associated with the final task performance.

[0085] The second path is the counterfactual path. This path is generated to provide a benchmark for comparison in subsequent training phases. A counterfactual path is an alternative computational path that is equally effective but is expected to have lower performance than the optimal computational path. The meta-controller employs a specific strategy to generate this path. In one embodiment, during the process of finding the optimal path using methods such as reinforcement learning or Monte Carlo tree search, in addition to determining the highest-scoring path as the optimal computational path, a second-best-scoring path is also retained as a counterfactual path. In another embodiment, the optimal computational path can be determined first, and then perturbed to generate a counterfactual path, for example, by replacing one or more knowledge cells in it with other functionally similar but lower-performing cells in the dynamic knowledge covariance graph 120 according to the meta-controller's evaluation.

[0086] Finally, the path planning module 140 outputs the definition information of the determined optimal computation path and counterfactual path (e.g., two ordered lists composed of knowledge cell identifiers) to the forward computation and parameter update module 150.

[0087] See attached document Figure 1 and Figure 2 The forward computation and parameter update module 150 receives the optimal computation path and counterfactual path output by the path planning module 140, as well as the abstract input representation generated by the task receiving and reconstruction module 130. This forward computation and parameter update module 150 is used to perform forward computation, loss calculation, and parameter updates of the model, and is the core execution unit for the system's learning process.

[0088] First, the forward computation and parameter update module 150 executes the forward computation process. The abstract input representation is fed into the starting knowledge cell of the optimal computation path, and the output of that cell is used as the input to the next cell in the path. This process is repeated sequentially until the last cell outputs the final optimal path prediction, denoted as... In parallel or serial mode, the forward computation and parameter update module 150 drives the abstract input representation to compute along the counterfactual path in the same manner, obtaining the final counterfactual path prediction output, denoted as... .

[0089] After obtaining the two predicted outputs, the forward computation and parameter update module 150 calculates a composite loss. This composite loss is a weighted combination of multiple loss components. The specific calculation steps are as follows:

[0090] Calculate task loss This loss is used to measure the difference between the optimal path prediction output and the actual task objective. Its calculation formula is:

[0091] ;

[0092] In the formula, It is the actual target value or label of the current task. It is a loss function predefined according to the task type, such as the cross-entropy loss function for classification tasks or the mean squared error loss function for regression tasks.

[0093] Calculate the counterfactual task loss This loss is used to measure the difference between the counterfactual path prediction output and the actual task objective. Its calculation formula is as follows:

[0094] ;

[0095] Entropy-constrained loss guided by computational graph This loss term is used to constrain the uncertainty of the knowledge cell output along the computation path, especially for cells located at the core of the graph. Its calculation formula is:

[0096] ;

[0097] In the formula, Refers to the set of knowledge cells along the optimal computation path. It is a knowledge cell on this path. It is a cell Topological centrality metric in dynamic knowledge covariance graph 120 It is a cell The output characteristics, It is the Shannon entropy calculated for this output feature.

[0098] Calculate the counterfactual verification loss This loss term is used to ensure that the performance of the planned optimal computational path is significantly better than the counterfactual path. Its calculation formula is:

[0099] ;

[0100] In the formula, It is a pre-defined non-negative constant, called the boundary value. The form of this formula ensures that the loss of the counterfactual path only becomes negative if... Loss close to or even better than the optimal path Only when this happens will a non-zero penalty value be generated.

[0101] Finally, the weighted sum of the above loss components yields the final composite loss. The calculation formula is as follows:

[0102] ;

[0103] In the formula, and It is a preset non-negative hyperparameter used to adjust the weights of each loss term.

[0104] After calculating the composite loss Then, the forward calculation and parameter update module 150 uses the gradient backpropagation algorithm to calculate... The gradients are calculated relative to the trainable parameters within all knowledge cells that constitute the optimal computation path and the counterfactual path. These parameters are then updated using a gradient descent optimizer (e.g., the Adam optimizer). It is noteworthy that in this step, only the parameters of the knowledge cells on the two paths selected by the current task are updated, while the parameters of other unused knowledge cells in the knowledge cell library 110 remain unchanged.

[0105] See attached document Figure 1 , Figure 2 and Figure 3 The function of the structure evolution module 160 is to perform targeted expansion of the model's structure when the existing knowledge of the system cannot meet the performance requirements of the task. This module is triggered after the forward computation and parameter update module 150 completes each round or several rounds of parameter updates to determine whether structure evolution is needed.

[0106] The structural evolution module 160 first determines whether the task performance has stagnated. In one specific embodiment, the criterion is: the system's task loss on an independent validation dataset over a predetermined number of consecutive training cycles. The evolutionary process will cease, or its rate of decline will fall below a preset minimum threshold. Subsequent evolutionary steps will only be executed if this condition is met.

[0107] After determining that the task performance has stalled, the structural evolution module 160 first calculates the task residuals. The task residual is defined as the true objective of the task. The predicted output generated by the current optimal computation path The difference between them, namely The task residual This represents the part that the current model has failed to fit correctly.

[0108] Subsequently, the structural evolution module 160 needs to select a knowledge cell that corresponds to the task residual from the knowledge cells that constitute the optimal computation path. The cell most relevant to its generation is chosen as the parent cell. (Refer to...) Figure 3 This selection process is achieved through a correlation analysis. For example, the output features of each knowledge cell on the optimal computation path are calculated in relation to the final task residual. The gradient correlation or mutual information between them. Knowledge cells with the highest correlation measure (such as...) Figure 3 The parent cell in the model was chosen because it is considered to be the main source of the current performance bottleneck.

[0109] After selecting a parent cell, the structural evolution module 160 derives a new knowledge cell based on that parent cell, namely... Figure 3 The new knowledge cell is shown. The derivation process includes:

[0110] First, the network structure of the parent cell is copied as the network structure of the new knowledge cell. Second, the parameters of the new knowledge cell are initialized based on the parameters of the parent cell. For example, the parameter values ​​of the parent cell are given a small random perturbation and then assigned to the new knowledge cell.

[0111] Next, the new knowledge cell will enter an isolated, specialized training phase. During this phase, the parameters of all knowledge cells in the knowledge cell library 110, except for the new knowledge cell, are frozen and remain unchanged. The training data consists of task samples that cause performance stagnation, and the training objective is to minimize the task residual with respect to the output of the new knowledge cell. The fitting error. This process ensures that the new cells are specifically trained to compensate for the deficiencies in existing knowledge.

[0112] Once the specialized training of the new knowledge cell is complete (e.g., its fitting error to the task residual converges), the structural evolution module 160 performs a final update operation. The trained new knowledge cell is formally added to the knowledge cell library 110. Simultaneously, it updates the structure of the dynamic knowledge covariance graph 120 by adding a new node representing the new knowledge cell and establishing connections between this new node and other nodes in the graph (such as the node corresponding to the parent cell) to reflect the evolution of the system's knowledge structure.

[0113] See attached document Figure 2 , Figure 2This is a flowchart of a multi-task large model training method based on adaptive incremental learning according to an embodiment of the present invention. The overall workflow of the multi-task large model training system 100 of the present invention will be described in detail below, which organically connects the functions of the aforementioned modules.

[0114] In an initial state, the multi-task large model training system 100 already possesses a knowledge cell library 110 containing several basic knowledge cells, and an initial dynamic knowledge covariance graph 120 representing the cooperative relationships between these cells. The specific steps of this method are as follows:

[0115] Step S1: Task Reception and Reconstruction. When a new task arrives, the task reception and reconstruction module 130 receives the task input containing the original task data and task metadata. The task reception and reconstruction module 130 parses the task metadata to determine the task type and, based on this, identifies a set of core knowledge cells in the dynamic knowledge covariance graph 120. Subsequently, the original task data is processed by calling this set of core knowledge cells to generate a structured abstract input representation.

[0116] Step S2: Path Planning. After receiving the abstract input representation, the path planning module 140's internal meta-controller executes a search process on the dynamic knowledge covariance graph 120 with the goal of maximizing the expected performance of the task. The output of this process is the definition of two paths: one is the optimal computational path, and the other is the counterfactual path.

[0117] Step S3: Forward computation and parameter update. The forward computation and parameter update module 150 drives the abstract input representation to perform forward computation along the optimal computation path and the counterfactual path, respectively, to obtain the optimal path prediction output. and counterfactual path prediction output Subsequently, the forward computation and parameter update module 150 calculates the results based on these two predicted outputs and the actual task objective. Calculate composite loss The gradient backpropagation algorithm is used to update the trainable parameters inside the knowledge cells that constitute these two paths.

[0118] Step S4: Performance Determination and Structure Evolution. After the parameters are updated, the structure evolution module 160 determines the system's task performance. If the performance on the validation set has not stagnated (e.g., the loss is still decreasing effectively), the current task processing cycle ends, and the system prepares to process the next task.

[0119] If the task performance is determined to be stagnant, the process proceeds to step S5. The structure evolution module 160 first calculates the current optimal path prediction output. With the true objective of the mission Task residuals between Then, among the cells that constitute the optimal computation path, the cell with the highest correlation to the residual is selected as the parent cell.

[0120] Step S6: Derivation and Update. The structural evolution module 160 derives a new knowledge cell based on the selected parent cell and performs specialized training on it in an isolated state to fit the task residuals. After the specialized training is completed, the new knowledge cell is added to the knowledge cell library 110, and the node and edge set of the dynamic knowledge covariance graph 120 is updated. After this step, the system returns to step S1 to process subsequent tasks with an enhanced knowledge system.

[0121] Please see the appendix Figure 4 The system and method of this invention can be based on a computer device 170, including a processor 171 and a memory 172. The memory 172 stores a computer program that can be executed by the processor. When the computer program is executed by the processor, it performs the steps of the method described above.

[0122] The present invention also provides a storage medium 173 on which a computer program is stored, the computer program being executed by a processor 171 to perform the steps of the method described above.

[0123] The storage medium 173 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable red-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk or optical disk.

[0124] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A multi-task large model training system with adaptive incremental learning, characterized in that, The system includes: A knowledge cell library is used to store multiple knowledge cells as basic capability units. Each knowledge cell is a computational unit containing independent functions and a set of trainable parameters, including Transformer encoder layers, convolutional neural network blocks, or multilayer perceptron networks. A dynamic knowledge covariance graph is used to represent the functional collaboration relationships between knowledge cells in the knowledge cell library in a graph structure. The task receiving and reconstruction module is used to receive task input containing natural language text, and based on the core knowledge cell in the dynamic knowledge covariance graph, call the text embedding cell and the syntactic analysis cell to jointly process the natural language text to generate an abstract input representation. The path planning module is used to receive the abstract input representation and plan the optimal computation path and counterfactual path on the dynamic knowledge covariance graph. The forward computation and parameter update module is used to drive the abstract input representation to perform computation along the optimal computation path and the counterfactual path respectively, to obtain the optimal path prediction output and the counterfactual path prediction output and calculate the composite loss, and then update the parameters of the knowledge cells on the optimal path prediction output and the counterfactual path prediction output. The structural evolution module is used to calculate the task residual based on the optimal path prediction output if the task performance stagnates after the forward calculation and parameter update module updates the parameters, and to directionally differentiate new knowledge cells based on the task residual, add the new knowledge cells to the knowledge cell library, and update the dynamic knowledge covariance graph.

2. The adaptive incremental learning multi-task large model training system according to claim 1, characterized in that, The task receiving and reconstructing module is also used for: Receive task input containing raw task data and task metadata, and parse the task metadata to determine task type information; Based on the parsed task type information, a set of core knowledge cells is queried and determined in the dynamic knowledge covariance graph. The query operation includes matching the task type information with the functional tags of each knowledge cell pre-stored in the knowledge cell library. The core knowledge cell is invoked to perform collaborative processing on the original task data, i.e., natural language text, to generate initial features, and the initial features are output as abstract input representations; wherein, the abstract input representations are high-dimensional feature vectors or tensors.

3. The adaptive incremental learning multi-task large model training system according to claim 1, characterized in that, The path planning module is also used for: Based on the abstract input representation generated by the task receiving and reconstruction module of the meta-controller, and with the abstract input representation as the initial state, a search process is performed on the dynamic knowledge covariance graph with the optimization objective of maximizing the expected performance of the task, so as to determine the optimal computation path; And a counterfactual path with expected performance lower than the optimal computation path is generated using a preset strategy, wherein the preset strategy includes: During the search process, one path with the second-best score is retained as the counterfactual path; Alternatively, one or more knowledge cells in the optimal computation path can be replaced to generate the counterfactual path.

4. The adaptive incremental learning multi-task large model training system according to claim 1, characterized in that, The forward calculation and parameter update module is also used for: The abstract input representation is driven to perform calculations along the optimal calculation path and the counterfactual path, respectively, to obtain the optimal path prediction output and the counterfactual path prediction output; Based on the difference between the optimal path prediction output and the task objective, and using the task loss formula, the task loss is calculated. The task loss formula is as follows: ； In the formula, It is the value of the task loss. It is the actual target value or label of the current task. It is a loss function pre-defined according to the task type. It is the final optimal path prediction output of the cell output; Based on the difference between the counterfactual path prediction output and the task objective, and using the counterfactual task loss formula, the counterfactual task loss is calculated. The counterfactual task loss formula is as follows: ； In the formula, It is the value of the counterfactual task loss. It is the output of counterfactual path prediction; Based on the topological centrality and output entropy of knowledge cells in the dynamic knowledge covariance graph, and the entropy constraint loss calculated using the graph-guided entropy constraint loss formula, the entropy constraint loss formula is as follows: ； In the formula, This is the value of the graph-guided entropy constraint loss. Refers to the set of knowledge cells along the optimal computation path. It is a knowledge cell on the path. It is a cell Topological centrality metric in dynamic knowledge covariance graphs It is a cell The output characteristics, It is the Shannon entropy calculated from the output features; Using the task loss and counterfactual task loss as input, the counterfactual verification loss is calculated using the counterfactual verification loss formula. This counterfactual verification loss ensures that the performance of the planned optimal computation path is significantly better than the counterfactual path. The counterfactual verification loss formula is as follows: ； In the formula, It is the value of counterfactual verification loss. It is a pre-defined nonnegative constant; The task loss, the graph-guided entropy constraint loss, and the counterfactual verification loss are combined to obtain a composite loss, the formula of which is: ； In the formula, It is a compound loss. and It is a preset non-negative hyperparameter used to adjust the weights of each loss term; The composite loss employs a gradient backpropagation algorithm to construct the gradients of the trainable parameters within all knowledge cells of the optimal computation path and counterfactual path, and uses a gradient descent optimizer to update the parameters of the knowledge cells of the optimal computation path and counterfactual path.

5. The adaptive incremental learning multi-task large model training system according to claim 4, characterized in that, The calculation of the counterfactual verification loss is used to compare the counterfactual task loss with the task loss. When the counterfactual task loss exceeds the task loss by a margin but does not exceed the boundary value, the counterfactual verification loss is set to a non-zero penalty value. When the counterfactual task loss exceeds the task loss by a margin exceeding the boundary value, the counterfactual verification loss is set to zero.

6. The adaptive incremental learning multi-task large model training system according to claim 1, characterized in that, The structural evolution module is also used for: After the forward calculation and parameter update module updates the parameters, if it is determined that the task performance has stalled, the task residual is calculated based on the difference between the optimal path prediction output and the task objective. Based on the task residual, among the knowledge cells that constitute the optimal computation path, the knowledge cell most relevant to the task residual is selected as the parent cell. The structural evolution module derives new knowledge cells with the same network structure as the parent cell based on the parent cell, and performs targeted training on the new knowledge cells with the optimization objective of minimizing the fitting error of the task residual. The new knowledge cells that have completed the targeted training are added to the knowledge cell library, and the node set and edge set of the dynamic knowledge covariance graph are updated.

7. The adaptive incremental learning multi-task large model training system according to claim 6, characterized in that, When the structural evolution module derives a new knowledge cell, it initializes the parameters of the new knowledge cell based on the parameters of the parent cell. Furthermore, the specialized training is an isolated training phase, in which only the parameters of the new knowledge cell are updated to fit the task residual, while the parameters of the knowledge cells constituting the dynamic knowledge covariance graph remain unchanged.

8. An adaptive incremental learning method for training a large multi-task model, applied to the system described in any one of claims 1-7, characterized in that, The method includes the following steps: S1. Store multiple knowledge cells as basic capability units in a knowledge cell library. The knowledge cells include Transformer encoder layers, convolutional neural network blocks or multilayer perceptron networks, and represent the functional synergy between the knowledge cells with a dynamic knowledge covariance graph. S2. Upon receiving a task input containing natural language text, based on the core knowledge cell in the dynamic knowledge covariance graph, the text embedding cell and the syntactic analysis cell are invoked to jointly process the natural language text and generate an abstract input representation. S3. Receive the abstract input representation and plan the optimal computation path and counterfactual path on the dynamic knowledge covariance graph; S4. Drive the abstract input representation to perform calculations along the optimal calculation path and the counterfactual path respectively to obtain the optimal path prediction output and the counterfactual path prediction output, and calculate the composite loss, and then update the parameters of the knowledge cells that constitute the optimal calculation path and the counterfactual path based on the composite loss. S5. After the parameters of the knowledge cell are updated, if the task performance is determined to be stagnant, the task residual is calculated based on the optimal path prediction output, and new knowledge cells are directionally differentiated based on the task residual. The new knowledge cells are added to the knowledge cell library, and the dynamic knowledge covariance graph is updated.

Images