A small model driven large model collaborative continuous learning method and system

Abstract

The application discloses a kind of based on small model driving large model collaborative continuous learning method and system, belong to artificial intelligence and continuous learning technical field, method includes: S1.text data of current continuous learning task is obtained;S2.the text data is extracted using small model feature, and the extracted feature is input to parameter generator dynamic generation adapter parameter specific to continuous learning task;S3.based on the adapter parameter, the basic large model for executing continuous learning task is fine-tuned;S4.utilize the basic large model after fine-tuning to predict.Wherein, small model relieves the forgetting problem through task incremental fine-tuning and experience replay, at the same time, the selection of samples in buffer zone is optimized by introducing information model, the burden of continuous learning is transferred to small model by the application, the high computing cost and catastrophic forgetting problem brought by directly fine-tuning large model are avoided, and stable, efficient learning of model to continuous task flow in resource-constrained environment is effectively realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence and continuous learning, and particularly relates to a large model collaborative continuous learning method and system based on small model driving. BACKGROUND

[0002] In the field of artificial intelligence and financial technology, continuous learning has become an important direction for coping with multi-task and long-term adaptability in financial risk control, intelligent driving, industrial Internet of Things and other scenarios. With the strong generalization and representation ability of large models (such as BERT, GPT, ViT, etc.) in natural language processing, computer vision and financial risk control, for example, in the financial continuous learning scenario, it is expected to improve the model's ability to identify and model dynamic risk patterns. However, in practical applications, continuous learning based on large models faces technical bottlenecks such as catastrophic forgetting and computational efficiency in various fields. For example, financial institutions face multiple constraints such as computing power, data and privacy, making it difficult to implement continuous training and updating of large models.

[0003] The existing methods mainly have the following challenges: first, although large models have strong feature extraction and generalization capabilities, their large parameter size results in huge computational and storage overhead, making it difficult to efficiently deploy and continuously update in resource-constrained edge environments (such as local or edge devices of financial institutions, vehicle-mounted devices, medical terminals, industrial sensors, etc.); second, small models are more suitable for low-cost training locally, but they are prone to catastrophic forgetting in long task sequences, for example, it is difficult to ensure the stability and continuity of risk modeling in financial scenarios; third, the dynamic changes and differences of data make it difficult to unify modeling and knowledge transfer, for example, in financial business scenarios, the heterogeneity of tasks and data is strong, which is manifested in the differences in customer groups, transaction behavior, time distribution and risk categories, making it difficult to transfer knowledge. In addition, in some scenarios with privacy requirements for data, such as financial scenarios, as data compliance requirements continue to improve, financial risk control systems not only need to protect user privacy and avoid centralized processing of sensitive data, but also need to consider computational efficiency and communication costs during model collaboration.

[0004] In comparison, lightweight small models have developed a variety of mature anti-forgetting techniques in the field of continuous learning, and are more easily updated and deployed in resource-constrained local or edge environments. However, existing work has not fully addressed the bottlenecks in knowledge transfer and collaborative adaptation between small and large models, resulting in the realistic dilemma of "large models being strong but difficult to continuously adapt; small models being flexible but difficult to efficiently and securely transfer their learning results to large models".

[0005] Therefore, there is an urgent need for an innovative method that can combine the global representation capability of large models with the flexible and continuous adaptation capability of small models, while ensuring that the computational and communication overhead is controllable, and achieving efficient and stable continuous adaptation of the model to the dynamic environment throughout its life cycle. For example, in the financial field, under the premise of ensuring compliance and efficiency, long-term learning and stable identification of dynamic risk patterns are achieved. SUMMARY

[0006] The present application aims to address the challenges of existing continuous learning methods when dealing with large models, particularly the high overhead of direct fine-tuning of large models, the tendency to forget historical knowledge, and the difficulty of small models to stably maintain long-term memory. The present application provides a small model-driven large model collaborative continuous learning method and system, which constructs a large-small model collaborative framework for continuous learning, and is used to achieve continuous knowledge update and forgetting inhibition of large-scale base models in dynamic task sequences and resource-constrained environments.

[0007] The purpose of the present application is achieved by the following technical solutions:

[0008] In a first aspect, a small model-driven large model collaborative continuous learning method is provided, comprising the following steps:

[0009] S1. Obtain text data of a current continuous learning task;

[0010] S2. Use a small model to extract features from the text data, and input the extracted features into a parameter generator to dynamically generate adapter parameters for the continuous learning task;

[0011] S3. Fine-tune a base large model that performs the continuous learning task based on the adapter parameters;

[0012] S4. Use the fine-tuned base large model for prediction.

[0013] In some embodiments, the training of the small model includes:

[0014] The small model is trained based on a consistency fine-tuning strategy of buffer replay and knowledge distillation.

[0015] In some embodiments, the following loss function is used when training the small model:

[0016]

[0017] wherein, represents the loss function, represents a set of samples in the buffer, represents a data sample in represents a data sample in represents data of the middle sample, denotes label of the middle sample, and is a hyperparameter, denotes the feature extracted by the current small model, denotes the feature extracted by the historical small model, denotes the adapter parameter representation generated by the current parameter generator, denotes the adapter parameter representation generated by the historical parameter generator, denotes the square of the L2 norm.

[0018] In some embodiments, the adapter parameter is in any of the following forms:

[0019] a soft prompt vector for splicing into the input embedding sequence of the base large model;

[0020] a low-rank matrix of LoRA for adding to the weight matrix inside the base large model that is frozen to form an adapted weight.

[0021] In some embodiments, further comprising:

[0022] introducing the selection of samples in the information model optimization buffer.

[0023] In some embodiments, the selection of samples in the information model optimization buffer specifically comprises:

[0024] an initialization sub-step: taking the small model and the parameter generator of the last task as the initial parameters of the information model;

[0025] an updating sub-step: using the data of the current task to update the information model by minimizing a loss function containing a regularization term, which is used to constrain the deviation of the parameters of the information model from the initial parameters;

[0026] an importance calculation sub-step: in the updating process of the information model, the gradient norm of each sample is accumulated as its importance score.

[0027] In some embodiments, the importance score is calculated as follows:

[0028]

[0029] wherein, denotes the i-th sample in the training dataset from the n-th current task, denotes the corresponding label, denotes an importance score of the sample, S denotes a preset training step number, and p denotes a training step number, which is valued from 1 to S, denotes a state of an information model at the p-th training step, denotes a gradient operator, denotes a predicted output of the information model, denotes a loss function of the information model.

[0030] In some embodiments, the loss function of the fine-tuned base large model is:

[0031]

[0032] wherein, denotes a total loss function on the n-th task, is an expectation operator, denotes the n-th task, denotes a cross-entropy loss function, denotes a classification head, denotes the fine-tuned base large model, denotes an input sample, denotes a corresponding real label.

[0033] In a second aspect, a small model driven large model collaborative continual learning system is provided, and the system comprises:

[0034] a data acquisition module configured to acquire text data of a current continual learning task;

[0035] an adapter parameter extraction module configured to perform feature extraction on the text data by using a small model, and input the extracted features into a parameter generator to dynamically generate adapter parameters of the continual learning task;

[0036] a large model fine-tuning module configured to fine-tune a base large model performing a continual learning task based on the adapter parameters;

[0037] a prediction module configured to perform prediction by using the fine-tuned base large model.

[0038] It should be further noted that the technical features corresponding to the above-mentioned various option embodiments can be combined or replaced with each other to form new technical solutions without conflict.

[0039] Compared with the prior art, the present application has the following advantages:

[0040] 1. Significantly improves computing and storage efficiency, reduces deployment cost: the invention designs a small-large model collaborative training framework, forming a collaborative continuous learning mechanism. In this framework, the small model not only completes the local task of continuous learning, but also dynamically generates adaptive parameters through the parameter generator and injects them into the frozen large model. This way avoids the huge computing overhead and storage demand generated by directly fine-tuning the large model, greatly reducing the computing and storage burden, which enables the invention to efficiently run on edge devices with limited computing and storage resources (such as local financial institutions, vehicle-mounted computing units, medical terminals, industrial sensors), greatly expanding the application boundary of large models in real-time scenarios. At the same time, the small model can drive the large model to enhance its representation ability during the learning of new tasks, thereby realizing bridge-like knowledge transfer.

[0041] 2. Effectively alleviates catastrophic forgetting, enhances the long-term stability and robustness of the model: the invention introduces a consistency fine-tuning strategy based on buffer playback and knowledge distillation to constrain the small model and its parameter generator to maintain consistency in feature output and generated parameters when facing historical samples, thereby suppressing the forgetting of old knowledge when learning new tasks. This ensures the prediction accuracy and decision reliability of risk control, diagnosis, and control systems during long-term evolution.

[0042] 3. Optimizes knowledge retention efficiency under limited resources through intelligent sample selection: the invention proposes an information model-driven sample selection mechanism to optimize the sample composition of the buffer, which can intelligently identify and retain representative samples that are most critical to coordinating historical knowledge and new task knowledge. This method maximizes knowledge retention efficiency under limited memory buffer capacity, especially suitable for real-world scenarios where sample categories are imbalanced and key samples are scarce (such as rare fault diagnosis), thereby effectively reducing the false positive rate and false negative rate of the model. BRIEF DESCRIPTION OF DRAWINGS

[0043] Figure 1 A flowchart of the small model-driven large model collaborative continuous learning method of the invention;

[0044] Figure 2 A schematic diagram of the small-large model collaborative training framework of the invention;

[0045] Figure 3 A specific training flowchart of the small-large model collaborative training framework of the invention; DETAILED DESCRIPTION

[0046] The technical solutions of the present application will be described clearly and completely below in conjunction with the drawings. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. The components of the embodiments of the present application described and shown in the drawings can be arranged and designed in various different configurations. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0047] It should be noted that the defects in the above prior art solutions are the result of the inventors' practice and careful research, therefore, the discovery process of the above problems and the solutions proposed by the embodiments of the present application to the above problems should be the contribution of the inventors to the present application in the process of invention and creation, and should not be understood as technical content known to those skilled in the art.

[0048] Based on the above related statements, since the present embodiment relates to the technology related to continual learning, in order to make the purpose, technical solutions and advantages of the embodiments of the present application more clear, the related professional terms that the present application may involve are explained as follows:

[0049] 1. Continual Learning (CL) refers to the paradigm of incrementally training a model on a stream of task data, aiming to preserve the performance on previously learned tasks while learning new tasks. Each task usually has independent training / validation / testing sets, and the model can only access the data of the current task during training. The main challenge is catastrophic forgetting: the model's performance significantly decreases after learning a new task. To alleviate forgetting, common methods include regularization methods (such as EWC), modular / structured methods, and sample replay (Replay) methods; in the replay method, how to efficiently select and manage the limited capacity of the memory buffer (Memory Buffer) is a key problem, and the present application uses information model-based sample selection to improve the efficiency of replay.

[0050] 2. Foundation Models (FM) refer to large models (such as large-scale Transformers / ViTs, GPT-type models) that are pre-trained on large-scale data and have strong general-purpose representation capabilities, also known as foundation large models. FM usually has large parameter size and strong generalization ability, but directly fine-tuning it for continual learning is costly and more prone to forgetting or parameter conflicts; therefore, in resource-constrained or frequently changing task scenarios, parameter-efficient methods are needed to indirectly adapt FM.

[0051] 3. Small Model Lightweight neural network (e.g. ResNet-18, MobileNet, etc.) with few parameters, low training and updating cost, and more easily combined with mature CL techniques (EWC, LwF, Replay, etc.). In the present invention, the small model plays a dual role: on the one hand, it performs its own continual learning to adapt to the task sequence, on the other hand, it serves as a driver for the "parameter generator / adapter" to dynamically generate sample / task-specific adapter parameters for the frozen FM, indirectly adjusting the FM behavior.

[0052] 4. LoRA (Low-Rank Adaptation) represents the increment of the large model weight in the form of low-rank decomposition (∆W = A·B), and only trains the two small matrices A and B to achieve efficient adjustment of the original weight. LoRA preserves the continuity of the weight structure, making it easy to fine-tune the main weight while maintaining the stability of the main weight.

[0053] 5. Parameter-Efficient Fine-Tuning (PEFT) refers to a set of methods that achieve adaptation by introducing a small number of trainable parameters without or with minimal modification of the core parameters of the original large model. Common PEFT techniques include Prompt Tuning (concatenating or learning a small number of trainable prompt vectors in input embedding) and LoRA (Low-Rank Adaptation) (approximating the weight increment with a low-rank matrix and only training the increment). PEFT is widely used in CL to reduce computational / storage burden and maintain model stability.

[0054] To solve the technical problems pointed out in the background art, the embodiments provided by the present invention are as follows:

[0055] In an exemplary embodiment, a small model-driven large model collaborative continual learning method is provided, as shown in Figure 1 The method comprises the following steps:

[0056] S1. Obtain text data of a current continual learning task;

[0057] S2. Use the small model to extract features from the text data, and input the extracted features into the parameter generator to dynamically generate adapter parameters for the continual learning task;

[0058] S3. Fine-tune the base large model performing the continual learning task based on the adapter parameters;

[0059] S4. Use the fine-tuned base large model for prediction.

[0060] Specifically, as shown in Figure 2As shown, taking the financial continuous learning scenario as an example, a small model-driven large model collaborative continuous learning framework for financial applications is proposed. Lightweight small models are used as driving engines to perform online learning of continuous task sequences and generate adapter parameters for frozen large pre-trained financial models to indirectly adjust their behavior in financial risk control and credit scenarios. This method eliminates the dependence of traditional continuous learning methods on direct fine-tuning of large models and achieves efficient and auditable knowledge transfer. This method is suitable for financial continuous task flows such as abnormal transaction detection and credit evaluation, significantly improving the practicality, stability, and generalization ability of models in complex environments.

[0061] In this small model collaborative training framework, the cooperative continuous learning mechanism is that the small model (such as ResNet, CNN) not only completes the continuous learning of the local task, but also dynamically generates adaptive parameters (such as LoRA matrix or prompt vector) through a parameter generator and injects them into the frozen large model (such as ViT). This approach avoids direct continuous fine-tuning of large models, significantly reducing computational and storage burdens, while enabling small models to drive large models to enhance representation capabilities during the learning of new tasks, thereby achieving bridging knowledge transfer.

[0062] Under this framework, the small model plays a dual role, as it is responsible for both cross-task continuous learning and generating sample-specific adapter parameters that adjust the behavior of the frozen base model through a parameter generator. These adapter parameters are then injected into the large model to enable it to adjust its representation. Through this process, the small model can serve as a dynamic intermediary that evolves over time, effectively supporting continuous learning while enhancing the performance of static large models. For example, Figure 3 As shown, the specific workflow is as follows: the lightweight small model receives financial samples (such as transaction sequences, account features, or text notes) for feature extraction, and the extracted intermediate representation is input into a parameter generator network to dynamically generate task / sample-specific adapter parameters (such as soft prompt vectors for financial text or time series modules, or LoRA low-rank matrices for weight adjustment). The generated adapter is injected into the frozen base financial model (such as a pre-trained Transformer architecture large model or a cross-modal financial representation model) to achieve indirect fine-tuning of the large model without modifying its core parameters. This mechanism supports both prompt and low-rank adaptation modes, ensuring flexibility in dealing with abnormal transaction patterns or credit distribution drift, while reducing the computational and compliance audit costs associated with direct updates to large models.

[0063] Further, to address the catastrophic forgetting problem that small models may experience during continuous learning, the invention designs a consistency fine-tuning strategy based on buffer replay and knowledge distillation. Key historical financial samples (such as those previously identified as high-risk or transaction anomalies) are saved in the local buffer along with the small model and its parameter generator at the previous time. During training, the current small model is constrained by feature distillation and consistency loss of the generated parameter matrix, so that it retains historical discrimination ability when adapting to new rules or customer groups. This component ensures that the adapter generated by the small model remains stable over time, thereby indirectly improving the continuous performance of the frozen large model in credit scoring and anomaly detection tasks and the explainability of risk discrimination.

[0064] Further, an information model (IM) driven sample selection strategy is proposed to optimize the composition of the buffer samples. After each task update, the IM is updated regularly using the previous small model as the initial state. The importance of samples is measured by the gradient norm accumulated in the early stages of IM training using current financial data (including normal and abnormal samples). Based on this importance measure, the top-ranked representative samples are selected by class or risk level to enter the limited-capacity memory buffer, thereby preserving key samples for both historical and current tasks in the context of scarce abnormal samples and class imbalance, improving the effectiveness of replay training and reducing false positives and false negatives. For continuous fine-tuning of small models, knowledge distillation is used between the previous and current small models using buffered samples. The goal is to generate consistent parameters for fine-tuning the large model.

[0065] In summary, in this embodiment, the online learning ability of the small model is converted into sample-level adapter injection for the frozen large model by the parameter generator; the stability of the small model's generation ability is ensured through buffer review and knowledge distillation; and the selection of samples in the limited buffer is optimized through the information model. This solution can achieve rapid adaptation to new risk patterns, reduce historical knowledge forgetting, and improve the accuracy and robustness of risk control and credit assessment systems in resource-constrained, compliance audit, and real-time response demanding financial scenarios. Typical financial application scenarios include:

[0066] Anomaly transaction detection: Continuously learn new transaction patterns to improve the real-time identification of abnormal transactions by the financial system.

[0067] Credit risk assessment: Achieve long-term adaptation and robustness improvement of credit scoring models as customer group characteristics change dynamically.

[0068] Anomaly pattern monitoring: Continuously update the model's identification ability in complex cross-border transactions and new fund flow patterns to effectively detect abnormal risks.

[0069] Compliance review and regulatory technology: support financial institutions to continuously learn and update regulatory rule mapping in compliance review under the premise of privacy protection, and improve the level of automation and intelligence.

[0070] Therefore, the present application can be widely applied to the scenes of financial risk control, credit management and compliance detection which require continuous learning and privacy protection, and has important industrial promotion and application value.

[0071] In an example, a pre-trained ViT is used as the base large model, which remains frozen and untrainable throughout the entire continuous learning process. A small convolutional neural network, denoted as , is introduced, whose architecture can be varied to include models such as ResNet-18, VGG-16, etc. For a given input sample and its label , it first obtains an intermediate representation through the feature extractor of the local small model .

[0072] (1)

[0073] The representation is then input into the parameter generator , which generates a set of sample-specific parameters, denoted as .

[0074] In prompt-based collaborative continuous learning, can be flexibly organized as part of the input to the base large model. Let and their corresponding embedding features be . Then the input embedding can be adjusted as follows:

[0075] (2)

[0076] where is the adjusted input embedding sequence, is the concatenation operator, which concatenates along the token length dimension, and here is a single prompt with token length and the same embedding size D as .

[0077] In LoRA-based collaborative continuous learning, a pre-trained weight matrix is updated through low-rank decomposition, so that , represents the increment of the large model weight. During training, remains in a frozen state and does not accept gradient updates, while comprises trainable parameters. and are applied to the same input, their outputs are summed element-wise. For input , the standard projection is modified as follows:

[0078] (3)

[0079] To ensure stable optimization, the output of is scaled by during training, where denotes the rank of .

[0080] To jointly train the small model parameters , the adapter generator , and the classification head , the loss on task is defined as:

[0081] (4)

[0082] where denotes the overall loss function on the nth task, is the expectation operator, denotes the nth task, denotes the cross-entropy loss function, denotes the classification head, denotes the fine-tuned base large model, denotes the input sample, denotes the corresponding true label.

[0083] This collaborative optimization framework allows the small model to serve as a dynamic bridge, continuously improving the utility of the large model while adapting to new tasks.

[0084] Further, the small model can continuously adapt to evolving tasks and in turn drive the base large model through sample-aware parameter generation. This indirect learning path allows the large model to gradually adapt to new tasks without being explicitly and directly fine-tuned, thereby enhancing its continuous learning capabilities in dynamic environments.

[0085] Illustratively, the local client maintains the local CNN from the previous task and its corresponding parameter generator as and respectively. For each data sample in the buffer , which is processed by the previous local model and the current local model. Specifically, the previous CNN processes the buffer data samples to produce features and generates a matrix while the current CNN processes the buffer data samples to produce features and generates a matrix . The consistency loss can be summarized as:

[0086] (5)

[0087] where denotes the loss function, denotes the set of samples in the buffer, denotes the data sample in , denotes the data of the sample in , denotes the label of the sample in , and is a hyperparameter, denotes the features extracted by the current local model, denotes the features extracted by the historical local model, denotes the adapter parameter representation generated by the current parameter generator, denotes the adapter parameter representation generated by the historical parameter generator, denotes the square of the L2 norm. With this loss function, the goal is to ensure that the local model extracts consistent features for past samples and that the local parameter generator produces similar matrices. This mechanism enhances the consistency of the local model, enabling it to continuously fine-tune the large model.

[0088] Exemplarily, this embodiment introduces an information model (IM) for sample selection that combines current and past insights. A scaling factor is further proposed to adjust the information balance between these sources. During the update of the IM, the sample gradient norm is recorded to derive an importance score, integrating both local (current task) and global (overall task relevance) perspectives.

[0089] Specifically, after completing the training on task , the information model is updated to incorporate the knowledge of previous and current tasks. Let be the local model trained on task , be the current local model trained on task . The information model is initialized as and updated using samples from by minimizing the following regularized objective:

[0090] (6)

[0091] in, Represents the updated information model ， This represents the optimization operator. Represents the optimization variable. This represents the nth training dataset from the current task. i One sample, express The corresponding tags Represents the loss function. Indicates the sample Input parameters are v The predicted output obtained in the information model. The strength coefficient of the regularization term , ∈(0,1), and It controls the rate at which the update step size is increased. Hyperparameters The balance between previous and current information combined in the update has been adjusted.

[0092] To explain the update mechanism, an analogy can be drawn with momentum methods in optimization. Momentum-based methods utilize previous updates to guide the direction of the current update. Similarly, the second term, as a momentum component, integrates small models from the past. Information to influence Update. Hyperparameters. The weights controlling this momentum component lie within the interval (0,1). When When it is close to 0, IM is primarily intended to recover the previous model. This will ensure consistency with current data. Conversely, as... The increase in places greater emphasis on the current task. By utilizing Theorem 1, the convergence of IM is ensured, thus enabling the computation of the gradient norm during training.

[0093] Theorem 1 (Convergence of IM Updates): Assume the loss function ℓ(⋅) is smooth and bounded below, and the variance of the stochastic gradient is bounded. Let... The update sequence generated by SGD in S iterations is initialized to... Then, there exists a constant C < ∞ such that:

[0094] (7)

[0095] This result ensures that the information model converges to the regularized loss function. a stable point, effectively fusing the previous knowledge and the information of the current task through the proximal control update.

[0096] Since the information model integrates the past and current models, a sample with a larger gradient norm in the information model indicates that the sample can better promote the information model to combine previous and current knowledge to fit the task. This effect can be more pronounced in the early training stage of the information model, when there are fewer fluctuations around the optimal solution than in the later training stage. Therefore, the gradient norm is accumulated during the training process of the information model, and the sample importance is calculated in the early training stage, as follows:

[0097] (8)

[0098] wherein, represents the importance score of the sample, S represents the preset number of training steps, p represents the training step number, which takes a value from 1 to S , represents the information model state at the p th training step, represents the gradient operator, represents the predicted output of the information model, represents the loss function of the information model.

[0099] On the basis of constructing this importance measure, the large model is trained in cooperation with the small model to complete the current task . After the cooperative training of the large model and the small model, the top M samples with the highest importance in each class of the training data of the current task are selected. These selected samples are then stored to update the local buffer. This selection strategy ensures both class balance and priority consideration of stable and representative samples for future learning.

[0100] In another example embodiment, a small model driven large model cooperative continuous learning system is provided, comprising:

[0101] A data acquisition module is configured to acquire text data of a current continuous learning task.

[0102] An adapter parameter extraction module is configured to extract features from the text data using a small model, and input the extracted features into a parameter generator to dynamically generate adapter parameters for the continuous learning task.

[0103] A large model fine-tuning module is configured to fine-tune a base large model performing a continuous learning task based on the adapter parameters.

[0104] A prediction module is configured to use the fine-tuned base large model to make predictions.

[0105] The above detailed description is a detailed description of the application, and cannot be considered as limiting the specific embodiments of the application to these descriptions. For those skilled in the art, without departing from the concept of the application, a number of simple deductions and substitutions can be made, which should be considered as falling within the protection scope of the application.

Claims

1. A small-model-driven, large-model collaborative continuous learning method, applied to financial risk control, autonomous driving, and industrial IoT, characterized in that... Includes the following steps: S1. Obtain the text data for the current ongoing learning task; S2. Use a small model to extract features from the text data, and input the extracted features into a parameter generator to dynamically generate adapter parameters for the continuous learning task; The training of the small model includes: The small model is trained using a consistency fine-tuning strategy based on buffer replay and knowledge distillation; the small model is trained using the following loss function: ,in, Represents the loss function. This represents the set of samples in the buffer. express Data samples in express Data from the medium sample express Labels of the middle sample and It's a hyperparameter. This represents the features extracted by the current small model. This represents the features extracted by the historical small model. This represents the adapter parameter representation generated by the current parameter generator. This represents the adapter parameter representation generated by the history parameter generator. The square of the L2 norm is represented; the adapter parameters are in any of the following forms: Soft cue vectors are used to concatenate the input embedding sequence of the base model. The low-rank matrix of LoRA is used to add to the weight matrix frozen inside the underlying large model to form the adapted weights. S3. Fine-tune the underlying large model for performing the continuous learning task based on the adapter parameters; S4. Make predictions using the fine-tuned base model; the loss function of the fine-tuned base model is: ,in, This represents the overall loss function on the nth task. For expected value operators, This represents the nth task. Represents the cross-entropy loss function. Indicates the category header, This represents the base model after fine-tuning. Indicates the input sample. This represents the corresponding real label.

2. The method for collaborative continuous learning of large models driven by small models according to claim 1, characterized in that, Also includes: An information model is introduced to optimize the selection of samples within the buffer.

3. The method for collaborative continuous learning of large models driven by small models according to claim 2, characterized in that, The introduction of an information model to optimize the selection of samples within the buffer specifically includes: Initialization sub-step: Use the parameters of the small model and parameter generator from the previous task as the initial parameters of the information model; Update sub-step: Using the data from the current task, update the information model by minimizing a loss function containing a regularization term, which constrains the deviation of the information model's parameters from the initial parameters; Importance calculation sub-step: During the update process of the information model, the gradient norm of each sample is accumulated as its importance score.

4. The method for collaborative continuous learning of large models driven by small models according to claim 3, characterized in that, The importance score is calculated as follows: ,in, This represents the i-th sample from the training dataset for the n-th current task. express The corresponding tags This represents the importance score of the sample, S represents the preset number of training steps, and p represents the training step number, which ranges from 1 to S. This represents the information model state at the p-th training step. Represents the gradient operator. This represents the predicted output of the information model. This represents the loss function of the information model.

5. A collaborative continuous learning system for large models driven by small models, applied to financial risk control, autonomous driving, and industrial IoT, characterized in that... include: The data acquisition module is used to acquire text data for the current ongoing learning task; The adapter parameter extraction module is used to extract features from the text data using a small model, and input the extracted features into a parameter generator to dynamically generate adapter parameters for the continuous learning task; the small model uses the following loss function during training: ,in, Represents the loss function. This represents the set of samples in the buffer. express Data samples in express Data from the medium sample express Labels of the middle sample and It's a hyperparameter. This represents the features extracted by the current small model. This represents the features extracted by the historical small model. This represents the adapter parameter representation generated by the current parameter generator. This represents the adapter parameter representation generated by the history parameter generator. The square of the L2 norm is represented; the adapter parameters are in any of the following forms: Soft cue vectors are used to concatenate the input embedding sequence of the base model. The low-rank matrix of LoRA is used to add to the weight matrix frozen inside the underlying large model to form the adapted weights. The large model fine-tuning module is used to fine-tune the underlying large model for performing continuous learning tasks based on the adapter parameters; The prediction module is used to make predictions using a fine-tuned base model; the loss function of the fine-tuned base model is: ,in, This represents the overall loss function on the nth task. For expected value operators, This represents the nth task. Represents the cross-entropy loss function. Indicates the category header, This represents the base model after fine-tuning. Indicates the input sample. This represents the corresponding real label.

Images