Streaming category incremental learning method and system for open environment
By performing incremental learning of streaming categories in an open environment, using dynamic feature weight adjustment and confidence threshold detection, combined with internal and external optimization models, the instability and memory consumption problems of category identification in streaming data are solved, and efficient new category detection and model updates are achieved.
Patent Information
- Application Number
- CN202510526861.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-25
- Publication Date
- 2025-08-08
AI Technical Summary
The existing technology has problems such as class variance deviation, distribution offset sensitivity and catastrophic forgetting in streaming category incremental learning in open environments, and the existing methods are sensitive to noise, have high error detection rates, are large in memory consumption, and cannot meet real-time requirements.
By receiving streaming data, adaptive adjustment of dynamic feature weights is performed, feature extraction is performed, and unknown category detection is performed based on confidence threshold and pseudo-label generation. The internal and external optimization model is used to integrate new and old category information, and combine adaptive weighting schemes and distillation loss function optimization model.
Effectively respond to distribution changes, adapt to unlabeled data and new categories, realize model updates and data regularization in a short time, improve the identification ability of new data and new categories, reduce the error detection rate and optimize memory consumption.
Smart Images
Figure CN120448868A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of artificial intelligence and machine learning technologies, and specifically to a streaming category incremental learning method and system for an open environment. Background Art
[0002] In recent years, incremental learning has made significant progress in fields such as image classification and object detection. However, its primary assumption is a "closed environment," where all categories are predetermined and fixed before training. However, in real-world dynamic environments such as live video streaming and online user behavior analysis, where data continuously arrives in a streaming format and may contain previously unseen categories, numerous problems arise, such as class variance bias, sensitivity to distribution shift, and catastrophic forgetting.
[0003] Currently, existing technologies have great limitations. Pseudo-label-based methods rely on thresholds to select pseudo-labels, are sensitive to noise, and have a false detection rate of up to 20%-30%. Replay-driven methods require a large amount of old data and consume a huge amount of memory. Dynamic architecture methods cannot meet real-time requirements. Summary of the Invention
[0004] In order to address the deficiencies mentioned in the above background technology, the purpose of the present invention is to provide a streaming category incremental learning method and system for open environments.
[0005] In a first aspect, the purpose of the present invention can be achieved by the following technical solution: a streaming category incremental learning method for an open environment, the method comprising the following steps:
[0006] Receive streaming data, adaptively adjust dynamic feature weights based on the streaming data, and perform feature extraction to obtain features to be detected for new and old categories;
[0007] The new and old category features to be detected are used to detect unknown categories based on the confidence threshold and pseudo-label generation to obtain new category data and old category data;
[0008] The new category data and the old category data are input into the pre-established inner and outer layer optimization models, and the updated and integrated new category information and the old category information are output and saved.
[0009] In conjunction with the first aspect, in certain implementations of the first aspect, the method further includes: the process of adaptively adjusting the dynamic feature weights based on streaming data includes:
[0010] Scale and align the convolution weights in the neural network layer and adjust the norm of the fully connected layer weights corresponding to the new and old categories. The formula is as follows:
[0011] K q =(W q ⊙Ω q)K q-1
[0012] where Ω q is the scaling weight matrix, ⊙ represents element-by-element multiplication, K q represents the output of the qth layer, K q-1 Represents the input of the qth layer; then normalizes the norm of the new and old category weights of the fully connected layer:
[0013]
[0014] where N old and N new are the norm vectors of the old and new category weights respectively.
[0015] In combination with the first aspect, in some implementations of the first aspect, the method further includes: the dynamic feature aggregation for adaptively adjusting the dynamic feature weights based on streaming data includes: the output features of the stable layer and the plastic layer are aggregated by weights and Combined, the formula is as follows:
[0016]
[0017] in And dynamically optimize the weights through back-propagation.
[0018] In combination with the first aspect, in some implementations of the first aspect, the method further includes: the process of performing unknown category detection based on the new and old category features to be detected based on the confidence threshold and pseudo label generation:
[0019] Computing confidence scores for unlabeled data The formula is as follows:
[0020]
[0021] like Where τ is the threshold, it is marked as an unknown category;
[0022] Constructing a supervised loss function With pairwise loss function Optimize the model to reduce the learning bias of known categories.
[0023] In combination with the first aspect, in some implementations of the first aspect, the method further includes: the supervised loss function The expression is as follows:
[0024]
[0025] where λ controls the interval strength, is the dynamically adjusted confidence mean, and s is an additional scaling factor.
[0026] In conjunction with the first aspect, in certain implementations of the first aspect, the method further includes: the pre-established inner and outer layer optimization model includes:
[0027] Inner layer optimization: fixed aggregation weight α i , optimize network parameters X i and ψ i , the formula is as follows:
[0028]
[0029] Outer layer optimization: fix network parameters and optimize aggregation weight α i , the formula is as follows:
[0030]
[0031] Historical information is retained through distillation loss, the formula is as follows:
[0032]
[0033] where q j is the probability distribution of the historical model output.
[0034] In conjunction with the first aspect, in certain implementations of the first aspect, the method further includes: a calculation process of the total loss function is as follows:
[0035]
[0036] in is the pairwise loss function, and the formula is as follows:
[0037]
[0038] R is the maximum entropy regularization term, and the formula is as follows:
[0039]
[0040] In a second aspect, in order to achieve the above-mentioned objectives, the present invention discloses a streaming category incremental learning system for an open environment, comprising:
[0041] The weight adjustment module is used to receive streaming data, perform adaptive adjustment of dynamic feature weights based on the streaming data, and perform feature extraction to obtain features to be detected for new and old categories;
[0042] The category detection module is used to perform unknown category detection based on the new and old category features to be detected based on the confidence threshold and pseudo-label generation to obtain new category data and old category data;
[0043] The information processing module is used to input new category data and old category data into the pre-established inner and outer layer optimization models, output the updated and integrated new category information and old category information and save them.
[0044] In another aspect of the present invention, in order to achieve the above-mentioned purpose, a terminal device is disclosed, including a memory, a processor, and a computer program stored in the memory and capable of running on the processor. The memory stores a computer program capable of running on the processor, and when the processor loads and executes the computer program, it adopts the above-mentioned streaming category incremental learning method for an open environment.
[0045] In another aspect of the present invention, in order to achieve the above-mentioned purpose, a computer-readable storage medium is disclosed, in which a computer program is stored. When the computer program is loaded and executed by a processor, the above-mentioned streaming category incremental learning method for an open environment is adopted.
[0046] Beneficial effects of the present invention:
[0047] This method effectively addresses the challenges of changing distributions and continuously adapts to emerging unlabeled data and new categories. It optimizes streaming data samples internally and externally to unify their scale, allowing for rapid model updates and fine-tuning of data regularization strategies. Its adaptive weighting scheme and new category detection exhibit unique performance in each environment, enabling better identification of new data and categories. Furthermore, this framework can be easily ported to other online or incremental semi-supervised methods. BRIEF DESCRIPTION OF THE DRAWINGS
[0048] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, those skilled in the art can derive other drawings based on these drawings without inventive effort.
[0049] Figure 1 It is a schematic flow chart of the method of the present invention;
[0050] Figure 2 Schematic diagram of the framework of the streaming category incremental learning model for open environments of the present invention;
[0051] Figure 3 is the performance graph of each method on the dataset;
[0052] Figure 4 Figure 5 is the performance graph of the method under different class mismatch rates in various datasets;
[0053] Figure 6 Schematic diagram of the system structure of the present invention; DETAILED DESCRIPTION
[0054] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts shall fall within the scope of protection of the present invention.
[0055] Example 1:
[0056] like Figure 1 As shown, a streaming category incremental learning method for open environments includes the following steps:
[0057] S101: Receive streaming data, adaptively adjust dynamic feature weights based on the streaming data, and perform feature extraction to obtain features to be detected for new and old categories;
[0058] The specific operations are as follows:
[0059] Input streaming dataset D, which contains labeled data with known classes C l and unlabeled data, i.e., unknown class C u .
[0060] Initialize the input data set and set the basic model parameters to θ base , the adaptive weight parameter is set to And initialized to 0.5, the scaling weight X is initialized to a full 1 matrix, and the memory M is stored 0 Initially empty.
[0061] The process of adaptively adjusting dynamic feature weights based on streaming data includes:
[0062] For the input data X, it is extracted through different layers of the model, and the formula is as follows:
[0063]
[0064] The formula for the aggregated features is as follows:
[0065]
[0066] Scale and align the convolution weights in the neural network layer and adjust the norm of the fully connected layer weights corresponding to the new and old categories. The formula is as follows:
[0067] K q =(W q ⊙Ω q )K q-1
[0068] where Ω q is the scaling weight matrix, ⊙ represents element-by-element multiplication, K q represents the output of the qth layer, K q-1 Represents the input of the qth layer; then normalizes the norm of the new and old category weights of the fully connected layer:
[0069]
[0070] where N old and N new are the norm vectors of the weights of the old and new categories respectively;
[0071] Finally, correct the output logic, the formula is as follows:
[0072]
[0073] S102: Perform unknown category detection on the features to be detected of the new and old categories based on the confidence threshold and pseudo label generation to obtain new category data and old category data;
[0074] The process of detecting unknown categories by using the new and old category features to be detected based on the confidence threshold and pseudo label generation:
[0075] Computing confidence scores for unlabeled data The formula is as follows:
[0076]
[0077] like Where τ is the threshold, it is marked as an unknown category;
[0078] Constructing a supervised loss function With pairwise loss function Optimize the model to reduce the learning bias of known categories.
[0079] The supervised loss function The expression is as follows:
[0080]
[0081] where λ controls the interval strength, is the dynamically adjusted confidence mean, and s is an additional scaling factor.
[0082] Generate pseudo labels by using cosine distance to generate pseudo labels for unlabeled data, and only retain samples with high confidence. The formula is as follows:
[0083]
[0084] S103: Input the new category data and the old category data into the pre-established inner and outer layer optimization model, output the updated and integrated new category information and old category information, and save them.
[0085] Pre-built inner and outer layer optimization models include:
[0086] Outer layer problem of inner and outer layer optimization: optimizing the aggregation weight α i , the formula is as follows:
[0087]
[0088] The inner layer problem of inner and outer layer optimization: optimizing network parameters and ψ i , the formula is as follows:
[0089]
[0090] After the internal and external layer problems are processed, the new class information and the old class information are integrated and the storage memory M is updated. t , retain the key samples of the old class, and the formula is as follows:
[0091]
[0092] The calculation process of the total loss function is as follows:
[0093]
[0094] in is the pairwise loss function, and the formula is as follows:
[0095]
[0096] R is the maximum entropy regularization term, and the formula is as follows:
[0097]
[0098] Dataset comparison experiment: Figure 3 The performance of the proposed method was compared with baseline methods on eight datasets, where performance was compared by adding or subtracting the standard deviation from the mean, with a fixed distribution mismatch rate of 10%. The proposed method achieved superior accuracy to other baseline methods on all streaming data, particularly on the CIFAR-10 dataset, achieving an accuracy of 85.11%.
[0099] Parameter comparison experiment: The method of the present invention is compared with the baseline method under different labeled / unlabeled class ratios, such as Figure 4 Figure 5As shown, each curve represents the average performance of 10 independent random runs of the method. As the ratio increases, performance becomes increasingly dependent on the number of known class instances. However, our method consistently outperforms the baseline method, effectively narrowing the gap between unknown and known classes. At a ratio of 60%, our method achieved an accuracy of 70.23%, 8.9% higher than other methods. Significant advantages can also be observed in other datasets.
[0100] Example 2: Figure 6 As shown, in order to achieve the above-mentioned purpose, the present invention discloses a streaming category incremental learning system for an open environment, comprising:
[0101] The weight adjustment module 11 is used to receive streaming data, perform adaptive adjustment of dynamic feature weights based on the streaming data, and perform feature extraction to obtain features to be detected for new and old categories;
[0102] Category detection module 12, used to perform unknown category detection based on the new and old category features to be detected based on the confidence threshold and pseudo label generation to obtain new category data and old category data;
[0103] The information processing module 13 is used to input the new category data and the old category data into the pre-established inner and outer layer optimization model, output the updated and integrated new category information and old category information, and save them.
[0104] Based on the same inventive concept, the present invention also provides a computer device, which includes: one or more processors and a memory for storing one or more computer programs; the program includes program instructions, and the processor is used to execute the program instructions stored in the memory. The processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing core and control core of the terminal, which is used to implement one or more instructions, specifically for loading and executing one or more instructions in a computer storage medium to implement the above method.
[0105] It should be further explained that, based on the same inventive concept, the present invention also provides a computer storage medium having a computer program stored thereon, which executes the above method when executed by a processor. The storage medium can be any combination of one or more computer-readable media. The computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. The computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electrical, magnetic, infrared, or semiconductor system, device or component, or any combination thereof. More specific examples (a non-exhaustive list) of computer-readable storage media include: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof. In the present invention, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, device or component.
[0106] Throughout this specification, references to terms such as "one embodiment," "example," or "specific example" indicate that a specific feature, structure, material, or characteristic described in conjunction with that embodiment or example is included in at least one embodiment or example of the present disclosure. In this specification, schematic representations of these terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.
[0107] The above shows and describes the basic principles, main features and advantages of the present disclosure. Those skilled in the art should understand that the present disclosure is not limited to the above embodiments. The above embodiments and descriptions are merely illustrative of the principles of the present disclosure. Various changes and improvements may be made to the present disclosure without departing from the spirit and scope of the present disclosure, and such changes and improvements shall fall within the scope of the present disclosure.
Claims
1. A streaming category incremental learning method for open environments, characterized by: The method comprises the following steps: Receive streaming data, adaptively adjust dynamic feature weights based on the streaming data, and perform feature extraction to obtain features to be detected for new and old categories; The new and old category features to be detected are used to detect unknown categories based on the confidence threshold and pseudo-label generation to obtain new category data and old category data; The new category data and the old category data are input into the pre-established inner and outer layer optimization models, and the updated and integrated new category information and the old category information are output and saved.
2. The method for incremental stream-based learning in an open environment according to claim 1, characterized in that: The process of adaptively adjusting dynamic feature weights based on streaming data includes: Scale and align the convolution weights in the neural network layer and adjust the norm of the fully connected layer weights corresponding to the new and old categories. The formula is as follows: K q =(W q ⊙Ω q )K q-1 where Ω q is the scaling weight matrix, ⊙ represents element-by-element multiplication, K q represents the output of the qth layer, K q-1 Represents the input of the qth layer; then normalizes the norm of the new and old category weights of the fully connected layer: where N old and N new are the norm vectors of the old and new category weights respectively.
3. The method for incremental stream-based learning in an open environment according to claim 2, wherein: The dynamic feature aggregation based on the adaptive adjustment of the dynamic feature weights of the streaming data includes: the output features of the stable layer and the plastic layer are aggregated by the weights and Combined, the formula is as follows: in And dynamically optimize the weights through back-propagation.
4. The method for incremental stream-based learning in an open environment according to claim 1, wherein: The process of detecting unknown categories by using the new and old category features to be detected based on the confidence threshold and pseudo label generation: Computing confidence scores for unlabeled data The formula is as follows: like Where τ is the threshold, it is marked as an unknown category; Constructing a supervised loss function With pairwise loss function Optimize the model to reduce the learning bias of known categories.
5. The method for incremental stream-based learning in an open environment according to claim 4, characterized in that: The supervised loss function The expression is as follows: where λ controls the interval strength, is the dynamically adjusted confidence mean, and s is an additional scaling factor.
6. The method for incremental stream-based learning in an open environment according to claim 1, wherein: The pre-established inner and outer layer optimization models include: Inner layer optimization: fixed aggregation weight α i , optimize network parameters X i and ψ i , the formula is as follows: Outer layer optimization: fix network parameters and optimize aggregation weight α i , the formula is as follows: Historical information is retained through distillation loss, the formula is as follows: where q j is the probability distribution of the historical model output.
7. The method for incremental stream-based learning in an open environment according to claim 1, wherein: The calculation process of the total loss function is as follows: in is the pairwise loss function, and the formula is as follows: R is the maximum entropy regularization term, and the formula is as follows:
8. A streaming category incremental learning system for open environments, characterized by: include: The weight adjustment module is used to receive streaming data, perform adaptive adjustment of dynamic feature weights based on the streaming data, and perform feature extraction to obtain features to be detected for new and old categories; The category detection module is used to perform unknown category detection based on the new and old category features to be detected based on the confidence threshold and pseudo-label generation to obtain new category data and old category data; The information processing module is used to input new category data and old category data into the pre-established inner and outer layer optimization models, output the updated and integrated new category information and old category information and save them.
9. A terminal device comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that: The memory stores a computer program that can be run on the processor. When the processor loads and executes the computer program, it adopts a streaming category incremental learning method for an open environment according to any one of claims 1 to 7.
10. A computer-readable storage medium storing a computer program, wherein: When the computer program is loaded and executed by the processor, the method for incremental streaming category learning for an open environment according to any one of claims 1 to 7 is adopted.