CN112734037A - Memory-guidance-based weakly supervised learning method, computer device and storage medium - Google Patents
Memory-guidance-based weakly supervised learning method, computer device and storage medium Download PDFInfo
- Publication number
- CN112734037A CN112734037A CN202110048718.8A CN202110048718A CN112734037A CN 112734037 A CN112734037 A CN 112734037A CN 202110048718 A CN202110048718 A CN 202110048718A CN 112734037 A CN112734037 A CN 112734037A
- Authority
- CN
- China
- Prior art keywords
- data
- memory
- label
- information
- new
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 60
- 230000015654 memory Effects 0.000 claims abstract description 144
- 239000013598 vector Substances 0.000 claims abstract description 68
- 238000012549 training Methods 0.000 claims abstract description 17
- 238000013145 classification model Methods 0.000 claims abstract description 14
- 230000008569 process Effects 0.000 claims abstract description 12
- 230000006870 function Effects 0.000 claims description 27
- 238000004590 computer program Methods 0.000 claims description 12
- 238000004364 calculation method Methods 0.000 claims description 7
- 238000011156 evaluation Methods 0.000 claims description 6
- 238000012545 processing Methods 0.000 claims description 5
- 238000013508 migration Methods 0.000 abstract description 4
- 230000005012 migration Effects 0.000 abstract description 4
- 230000007246 mechanism Effects 0.000 description 9
- 238000012937 correction Methods 0.000 description 6
- 238000010586 diagram Methods 0.000 description 4
- 230000003321 amplification Effects 0.000 description 3
- 238000000605 extraction Methods 0.000 description 3
- 238000002372 labelling Methods 0.000 description 3
- 238000003199 nucleic acid amplification method Methods 0.000 description 3
- 238000012216 screening Methods 0.000 description 3
- 238000006243 chemical reaction Methods 0.000 description 2
- 238000010276 construction Methods 0.000 description 2
- 238000013434 data augmentation Methods 0.000 description 2
- 239000000463 material Substances 0.000 description 2
- 230000003287 optical effect Effects 0.000 description 2
- 238000007781 pre-processing Methods 0.000 description 2
- 238000013526 transfer learning Methods 0.000 description 2
- 230000003044 adaptive effect Effects 0.000 description 1
- 230000004075 alteration Effects 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 238000013528 artificial neural network Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000002955 isolation Methods 0.000 description 1
- 230000010387 memory retrieval Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 239000013307 optical fiber Substances 0.000 description 1
- 230000000306 recurrent effect Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 230000002441 reversible effect Effects 0.000 description 1
- 230000006403 short-term memory Effects 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Computational Linguistics (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Image Analysis (AREA)
Abstract
The application provides a memory-guided weak supervised learning method, a computer device and a storage medium. The method comprises the following steps: and (4) encoding the old data set with the correct label, training the model and storing the model in a basic memory pool in a form of (feature vector, label). When new data is learned, part of memory information is extracted from the memory pool and is input into a prediction model together with the encoded new data for training, the prediction label is judged by taking a label similar to the memory in the memory pool as a reference, the judgment result and the new data are arranged into a new memory with the same form (characteristic vector, label), and the data which can not be labeled temporarily are screened out for later-stage study. And continuously repeating the process to update the model parameters until the new information of the batch is completely processed and then outputting the classification model. The classification model with strong learning ability is trained under the condition that the dependence of the model on large-scale labeled data is reduced as far as possible, and knowledge migration is realized.
Description
Technical Field
The application relates to the field of computer artificial intelligence, in particular to a memory-guidance-based weak supervised learning method, computer equipment and a storage medium.
Background
For small sample learning, when the input data includes a part of labeled data and a part of unlabeled data, the problem can be classified as incomplete supervision, and there are two general types of methods to solve. One is Active Learning (Active Learning). The goal of active learning is to minimize the number of queries, i.e., to select the most valuable unlabeled data for querying, requiring manual labeling by experts. The second is Semi-supervised Learning (Semi-supervised Learning). Semi-supervised learning is a learning paradigm in which the characteristics of the distribution characteristics of unlabeled data and the characteristics of the distribution of labeled data are analyzed without the participation of experts. In addition, researchers have also combined Transfer Learning (Transfer Learning), which has been extensively studied in recent years, with active Learning, and have achieved some good results by simultaneously migrating knowledge from the source domain and selecting the most valuable data for annotation. However, there are some problems at present, active learning depends heavily on the labeling of data or the clustering result thereof when establishing an initial model for selecting a query sample, and human experts still need to be queried continuously, which consumes a lot of manpower; and the migration learning has too high requirement on the data distribution consistency of the source domain and the target domain, and does not completely meet the real requirement.
From the current research, although the existing weak supervised learning methods can effectively reduce the dependence on large-scale labeled data, most of the methods consider tasks in isolation, lack the capabilities of continuous learning and adaptive learning, and have high requirements on data and less memory utilization.
Disclosure of Invention
The present application aims to solve at least one of the above mentioned technical problems to a certain extent.
Therefore, a first objective of the present application is to provide a memory-guided weak supervised learning method, so as to solve the problem of how to learn new images and data information through a memory knowledge guided model, and the problem of weak supervised and small sample learning through memory-assisted correction of wrong data categories, and train a classification model with strong learning ability under the condition of reducing the dependence of the model on large labeled data as much as possible, thereby implementing knowledge migration.
A second object of the present application is to propose a computer device.
A third object of the present application is to propose a non-transitory computer-readable storage medium.
A fourth object of the present application is to propose a computer program product.
In order to achieve the above object, an embodiment of a first aspect of the present application provides a method for weakly supervised learning based on memory guidance, including:
encoding each old data in the old data set with the correct label to obtain vector information of each old data, and training a prediction model according to the vector information of each old data and the correct label of each old data;
constructing a memory pool based on the vector information of each old data and the correct label of each old data; the memory pool comprises a plurality of memory information, and each memory information is used for representing the corresponding relation between data vector information and a label;
when new data is learned, encoding the new data to obtain vector information of the new data, extracting at least part of memory information from the memory pool, inputting the at least part of memory information and the vector information of the new data into the prediction model for training, and obtaining a prediction label aiming at the new data;
acquiring a reference label according to the at least part of memory information, and judging the prediction label according to the reference label;
forming a new memory information by the evaluation result and the vector information of the new data and storing the new memory information into the memory pool;
and continuously repeating the learning process of the new data to update the model parameters until the classification model is output after the new data of the batch are completely processed.
In a second aspect of the present application, a computer device is provided, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor, and when the processor executes the computer program, the memory-based learning method is implemented.
An embodiment of the third aspect of the present application provides a computer-readable storage medium, on which a computer program is stored, which, when being executed by a processor, implements the memory-based guiding weak supervised learning method described in the embodiment of the first aspect of the present application.
An embodiment of a fourth aspect of the present application provides a computer program product, where instructions of the computer program product, when executed by a processor, perform the memory-based guiding weak supervised learning method described in the embodiment of the first aspect of the present application.
According to the technical scheme of the embodiment of the application, a prediction model is trained through correctly labeled old data, a memory pool module is formed, part of memory guide models are extracted to learn new information and update self parameters, meanwhile, a memory-assisted self-error correction mechanism is established to judge whether the label is correct or not, the robustness of the system is enhanced, and finally a classification model with weak dependence on labeled data is output. Specifically, the system encodes the old data set with the correct label, trains the model and stores it in the basic memory pool in the form of (feature vector, label). When new data is learned, part of memory information is extracted from the memory pool and is input into a prediction model together with the encoded new data for training, the prediction label is judged by taking a label similar to the memory in the memory pool as a reference, the judgment result and the new data are arranged into a new memory with the same form (characteristic vector, label), and the data which can not be labeled temporarily are screened out for later-stage study. And continuously repeating the process to update the model parameters until the new information of the batch is completely processed and then outputting the classification model. In addition, the application realizes the calling of the stored knowledge when new information is learned by extracting partial memorized labels as reference and generating the prediction labels by data amplification, minimum entropy and other operations when new information is learned. .
Additional aspects and advantages of the present application will be set forth in part in the description which follows and, in part, will be obvious from the description, or may be learned by practice of the present application.
Drawings
The foregoing and/or additional aspects and advantages of the present application will become apparent and readily appreciated from the following description of the embodiments, taken in conjunction with the accompanying drawings of which:
FIG. 1 is a flowchart of a weak supervised learning method based on memory guidance according to an embodiment of the present application;
FIG. 2 is a schematic diagram of a tag evaluation flow according to an embodiment of the present application;
FIG. 3 is a schematic diagram of a predictive tag generation flow according to an embodiment of the application;
FIG. 4 is a flowchart illustrating a method for memory-guided learning based weakly supervised learning in accordance with an embodiment of the present application; and
FIG. 5 is a schematic block diagram of a computer device according to one embodiment of the present application.
Detailed Description
Reference will now be made in detail to embodiments of the present application, examples of which are illustrated in the accompanying drawings, wherein like or similar reference numerals refer to the same or similar elements or elements having the same or similar function throughout. The embodiments described below with reference to the drawings are exemplary and intended to be used for explaining the present application and should not be construed as limiting the present application.
A memory-oriented based weakly supervised learning method, a computer device, and a storage medium of embodiments of the present application are described below with reference to the accompanying drawings.
Fig. 1 is a flowchart of a weak supervised learning method based on memory guidance according to an embodiment of the present application. As shown in fig. 1, the memory-guided weak supervised learning based method may include the following steps.
In step 101, each old data in the old data set with the correct label is encoded to obtain vector information of each old data, and a prediction model is trained according to the vector information of each old data and the correct label of each old data.
In step 102, constructing a memory pool based on the vector information of each old data and the correct label of each old data; the memory pool comprises a plurality of memory information, and each memory information is used for representing the corresponding relation between the data vector information and the label.
In step 103, when new data is learned, encoding the new data to obtain vector information of the new data, extracting at least part of memory information from the memory pool, inputting at least part of the memory information and the vector information of the new data into a prediction model training, and obtaining a prediction label for the new data.
In some embodiments, the specific implementation process of extracting at least part of the memory information from the memory pool may be as follows: and performing similarity calculation on the vector information of the new data and the data vector information in each memory information in the memory pool, and taking the memory information of which the similarity calculation result meets the similarity memory matching index as at least part of memory information. In the embodiment of the present application, the similar calculation is performed by the following formula:
dist(xi,mj)＝||xi-mj||2
wherein x isiVector information representing new data, mjRepresents the data vector information in the jth memory information in the memory pool, dist (x)i,mj) Indicating the degree of vector similarity between the new data calculated with the norm L2 and the jth memory information.
In some embodiments of the present application, the specific implementation process of inputting at least part of the memory information and the vector information of the new data into the prediction model training to obtain the prediction label for the new data may be as follows: performing data enhancement for R times on unlabeled data input by a prediction model to obtain R enhanced data, and respectively obtaining R prediction classification probabilities from the R enhanced data through the prediction model; acquiring a reference label according to at least part of the memory information; and processing the average value of the R prediction classification probabilities and the reference label based on a sharpening function to obtain a prediction label aiming at new data.
In step 104, a reference label is obtained according to at least part of the memory information, and the predicted label is judged according to the reference label.
It should be noted that, in the embodiment of the present application, when new data is learned, if the new data is not labeled, the input of the prediction model is at least part of the memory information and the unlabeled data to be learned; if the new data is marked, temporarily storing the label carried by the new data as label-free data, and using at least part of memory information as the input of the prediction model.
In some embodiments, the distance between the vector information of the unlabeled data input by the prediction model and each data vector information among the at least partially memorized information may be calculated, and the distance between each data vector information may be weighted as the label of each data vector information; and acquiring a reference label according to the label corresponding to each data vector information in at least part of the memory information and the label weight of each data vector information.
In the embodiment of the present application, the implementation manner of evaluating the prediction tag according to the reference tag is as follows: new data input with a tag and new data input without a tag. When the model prediction tag of the new data with the tag is judged, as shown in fig. 2, the most similar memory can be inquired and extracted in the memory pool according to the feature vector of the tagged information, the extracted and memorized tag is used as a reference, the more accurate of the tag of the information and the prediction tag output by the model is judged and selected to be output as a result, and finally, the new information feature vector is combined to form a memory in the form of (feature vector, tag).
In this embodiment, the model prediction labels of the new data without labels are evaluated: in a real scene, the non-label data is generally not manually screened, so for the input of the type, a judging mechanism is arranged behind a model prediction module for screening, the data with the same prediction probability of each category but without obvious category direction is filtered, and only the non-label data with the probability exceeding a certain threshold value is reserved. The data to be filtered out can be understood as new information outside the category covered by the memory pool or new domain knowledge which is never contacted, so that the data is temporarily not classified and stored for later relearning.
In step 105, the evaluation result and the vector information of the new data are combined into a new memory information and stored in the memory pool.
In step 106, the learning process of the new data is repeated to update the model parameters until the classification model is output after all the new data of the batch is processed.
In some embodiments of the present application, a loss function may also be constructed, wherein the constructed loss function integrates cross-entropy loss of tagged data and loss of norm L2 after non-tagged data enhancement;
calculating loss values of the reference label and the prediction label based on the loss function, and updating model parameters of the prediction model according to the loss values; wherein the loss function is expressed as follows:
LOSS°＝Lx+λuLu
wherein | X | and | U | are the number of data samples captured in one training, and the numerical values correspond to the size of a standard data set and a standard-free data set; l is the number of classification categories, namely the number of all categories in the memory pool; x, l are labeled data and labels thereof, u, ypIs the unlabeled data and its predictive label; h (l, ° P (y | x; θ)) is a cross entropy function; lambda [ alpha ]uIs a balance factor for balancing supervision loss and unsupervised loss.
In order to facilitate a more clear understanding of the present application for those skilled in the art, the present application will be described in detail with reference to fig. 2 to 4.
As shown in fig. 2 to 4, the weak supervised learning method based on memory guidance in the embodiment of the present application mainly includes: the method comprises the steps of data preprocessing, prediction model construction, target function construction, prediction label generation, label judgment mechanism (including label input and label-free input), memory extraction and classification model generation. The specific implementation mode is as follows:
(1) data pre-processing
In order to better describe the content contained in the information, the input information needs to be encoded in the form of a vector, i.e., feature extraction is performed. For different types of information input such as audio or images, the application uniformly provides a general vector representation form:
wherein x iskjRepresents the jth component in the kth feature vector, and
the last component in the k-th feature vector is represented. For information xiTo the category ofiAnd (4) showing. After the encoding task is finished, a feature vector and category label pair (x) is formedi,li). For information to be stored in the memory pool, m is usediIn place of xiTo (x)i,li) Is stored.
(2) Building a prediction model
The method and the device guide the learning of new information by using memory, so that after the coding of a correctly labeled data set is completed, the data set is divided into a training set and a testing set to train a prediction model, so that the related memory information of old knowledge is reserved in the prediction model, and the related memory information is used for predicting the label of the new information later. The generated classification prediction model is marked as P (y | x, theta), wherein y is a classification class label of input x, and theta is a model training parameter. At present, various networks are used for data classification, and an existing Recurrent Neural Network (RNN) and a Long Short-Term Memory Network (LSTM) can be selected as basic models, and according to actual problem requirements and a form of learning information, training is firstly carried out on a corresponding large number of corpora, such as ImageNet image sets, and then fine tuning is carried out on the original information sets.
(3) Constructing an objective function
The method is based on a weak supervised learning method, namely inputting a small amount of labeled data and unlabeled data and outputting the classification label of the unlabeled data. For guiding the learning of new information by means of memory, for the prediction model constructed in (2), when new data is not labeled, the input is partial correct label data (x, l) randomly extracted from a memory pool and unlabeled data u to be learned; when new data is marked, the label y carried by the data (u, y) is temporarily stored, is also regarded as label-free data, and is input into the model together with part of label data (x, l) extracted from the memory pool.
On the basis of a weak supervised learning framework, under the condition of ensuring that the semantics are not changed, the method carries out conversion processing on the non-standard Data through series operations such as Data Augmentation (Data Augmentation) and Entropy Minimization (Entropy Minimization), and then restrains the invariance of the model to the Data before and after conversion through methods such as Consistency Regulation. The method integrates cross entropy loss of labeled data and L2 loss after non-labeled data enhancement, and forms a loss function as follows:
LOSS°＝Lx+λuLu
wherein | X | and | U | are the number of data samples captured in one training, and the numerical values correspond to the size of a standard data set and a standard-free data set; l is the number of classification categories, namely the number of all categories in the memory pool; x, l are labeled data and labels thereof, u, ypIs the unlabeled data and its predictive label; h (l, ° P (y | x; θ)) is a cross entropy function; lambda [ alpha ]uThe balance factors are balance supervision loss and unsupervised loss and can be specifically set according to experimental conditions.
The loss function considers both marked data and non-marked data, namely memory and new knowledge, and can effectively utilize knowledge to assist new information classification and enhance the robustness of the system. The mean square error MSE is selected as a consistency loss function, and KL divergence can be selected to replace MSE according to the actual data set condition.
(4) Generating predictive labels
As shown in FIG. 3, the prediction tag y using unlabeled data is used in calculating the loss function L (θ)pThis is a "guess" tag that is obtained through operations such as data enhancement, memory-assisted adjustment, sharpening, etc. First, for input non-standard dataCarrying out R times data enhancement to obtain uiAnd (i ═ 1.. times, R), obtaining R predicted classification probabilities through the model respectively. At the same time, k memorized labels with the most similar characteristics to the unlabeled data are searched in the memory pool, and a reference label y is generated by the following formular：
Where a is the softmax function, 0<a(u,xi)<1, c are the calculation of unlabeled data u and labeled data xiCosine distance between feature vectors. Selecting cosine distance between marked data and unmarked data as label weight of marked data, and referring to label yrIt is weighted by k similarly memorized tags.
In order to enhance the confidence degree of the prediction result, the method for minimizing the entropy is adopted. Processing reference tag y using sharpening function Sharpen (p, T)rAnd the average value of the R prediction classification probabilities reduces the entropy of label distribution so as to ensure that the decision boundary does not pass through a high-density area of edge data distribution. Finally, the category with the highest prediction probability is output, namely the prediction label yp. The sharpening function is formulated as follows:
wherein p is the probability that the model prediction data belongs to a certain class, and T is a temperature parameter.
(5) Label evaluation mechanism-labeled input
The application provides a memory-assisted error correction mechanism for judging whether the model predicts the marked new information correctly. As shown in FIG. 2, the most similar memory is searched and extracted in the memory pool according to the feature vector of the labeled information, and the label y of the memory is extractedbAs a reference, the label y of the information and the label y of the model output are judged and selectedbThe new information is finally combined with the feature vector of the new information to form a memory in the form of (feature vector, label).
(6) Label evaluation mechanism-label-free input
In a real scene, the non-label data is generally not manually screened, so for the input of the type, a judging mechanism is arranged behind a model prediction module for screening, the data with the same prediction probability of each category but without obvious category direction is filtered, and only the non-label data with the probability exceeding a certain threshold value is reserved. The data to be filtered out can be understood as new information outside the category covered by the memory pool or new domain knowledge which is never contacted, so that the data is temporarily not classified and stored for later relearning.
(7) Memory retrieval
Similar memory extraction is involved in many places in the above steps, and the degree of similarity of feature vectors is considered to represent the degree of possible data information belonging to the same category in the present application, because the distance between feature vectors is uniformly used as an index of similar memory matching, taking the L2 norm as an example:
dist(xi,mj)＝||xi-mj||2
wherein x isiVector information representing new data, mjRepresents the data vector information in the jth memory information in the memory pool, dist (x)i,mj) Indicating the degree of vector similarity between the new data calculated with the norm L2 and the jth memory information.
(8) Generating classification models
As described above, the learning of new data of only one batch is completed in steps (1) to (6), and the above process is repeated until all new data are learned, and the parameters of the model are updated, so that a classification model P (y | x, θ) with a higher learning capability than that of the previous model, that is, a new data-oriented model that depends on a small amount of labeled data, can be obtained. The model can be used for learning, classifying and screening subsequent new data.
In summary, as shown in fig. 4, in the embodiment of the present application, a prediction model is trained by correctly labeled old data, a memory pool module is formed, a part of memory-guided models is extracted to learn new information and update parameters of the prediction model, a memory-assisted self-error correction mechanism is established to judge whether a label is correct or incorrect, the robustness of the system is enhanced, and finally a classification model with weak dependency on labeled data is output. Specifically, the system encodes the old data set with the correct label, trains the model and stores it in the basic memory pool in the form of (feature vector, label). When new data is learned, part of memory information is extracted from the memory pool and is input into a prediction model together with the encoded new data for training, the prediction label is judged by taking a label similar to the memory in the memory pool as a reference, the judgment result and the new data are arranged into a new memory with the same form (characteristic vector, label), and the data which can not be labeled temporarily are screened out for later-stage study. And continuously repeating the process to update the model parameters until the new information of the batch is completely processed and then outputting the classification model. In addition, the application realizes the calling of the stored knowledge when new information is learned by extracting partial memorized labels as reference and generating the prediction labels by data amplification, minimum entropy and other operations when new information is learned.
From the above description, it can be seen that the method and the device can learn new data information by constructing a memory-guided weak supervised learning framework and using the model retaining old knowledge information, integrate a memory-assisted self-error correction mechanism, enhance the identification learning capability of the model on new information in different scenes and tasks, finally generate a classification model capable of effectively learning new data by using a small amount of labeled data, and save the labeling cost. Compared with the existing weak supervised learning method, the learning efficiency and accuracy are obviously improved on the basis of greatly reducing the labeled data quantity, and the method has a strong practical application prospect.
Therefore, the memory is introduced to guide the model to learn the unlabeled new information on the basis of the weak supervised learning framework, the identification learning capacity of the new information in different scenes and tasks is enhanced, the amount of labeled data required in the learning of the new task is effectively reduced, and knowledge migration and autonomous learning are realized. In addition, the memory and the new knowledge are comprehensively considered, cross entropy loss of the marked data and consistency loss of the unmarked data after amplification are linearly combined, the generalization capability of the strange data is enhanced, the old knowledge is used for assisting in correct classification of new information, and the robustness of the system is enhanced.
In the process of generating the prediction label, the prediction level of the label can be improved through memory-assisted adjustment on the basis of extracting the effective signal of the non-standard data, and the accuracy of the model is improved. In addition, the method and the device can judge the correctness of the prediction label output by the model more accurately through a memory auxiliary error correction mechanism, and simultaneously screen out information in a new field, thereby being beneficial to improving the robustness of the model.
In order to implement the above embodiments, the present application also provides a computer device.
FIG. 5 is a schematic block diagram of a computer device according to one embodiment of the present application. As shown in fig. 5, the computer device 500 may include: a memory 501, a processor 502 and a computer program 503 stored on the memory 501 and capable of running on the processor 502, wherein the processor 502 executes the program 503 to implement the memory-oriented learning method according to any of the above embodiments of the present application.
In order to achieve the above embodiments, the present application further proposes a non-transitory computer-readable storage medium, on which a computer program is stored, which when executed by a processor implements the memory-oriented learning method based on weak supervision according to any of the above embodiments of the present application.
In order to implement the foregoing embodiments, the present application also proposes a computer program product, wherein when the instructions in the computer program product are executed by a processor, the memory-oriented weak supervised learning based method described in any of the foregoing embodiments of the present application is executed.
In the description herein, reference to the description of the term "one embodiment," "some embodiments," "an example," "a specific example," or "some examples," etc., means that a particular feature, structure, material, or characteristic described in connection with the embodiment or example is included in at least one embodiment or example of the application. In this specification, the schematic representations of the terms used above are not necessarily intended to refer to the same embodiment or example. Furthermore, the particular features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples. Furthermore, various embodiments or examples and features of different embodiments or examples described in this specification can be combined and combined by one skilled in the art without contradiction.
Furthermore, the terms "first", "second" and "first" are used for descriptive purposes only and are not to be construed as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one such feature. In the description of the present application, "plurality" means at least two, e.g., two, three, etc., unless specifically limited otherwise.
Any process or method descriptions in flow charts or otherwise described herein may be understood as representing modules, segments, or portions of code which include one or more executable instructions for implementing steps of a custom logic function or process, and alternate implementations are included within the scope of the preferred embodiment of the present application in which functions may be executed out of order from that shown or discussed, including substantially concurrently or in reverse order, depending on the functionality involved, as would be understood by those reasonably skilled in the art of the present application.
The logic and/or steps represented in the flowcharts or otherwise described herein, e.g., an ordered listing of executable instructions that can be considered to implement logical functions, can be embodied in any computer-readable medium for use by or in connection with an instruction execution system, apparatus, or device, such as a computer-based system, processor-containing system, or other system that can fetch the instructions from the instruction execution system, apparatus, or device and execute the instructions. For the purposes of this description, a "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transport the program for use by or in connection with the instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of the computer-readable medium would include the following: an electrical connection (electronic device) having one or more wires, a portable computer diskette (magnetic device), a Random Access Memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disc read-only memory (CDROM). Additionally, the computer-readable medium could even be paper or another suitable medium upon which the program is printed, as the program can be electronically captured, via for instance optical scanning of the paper or other medium, then compiled, interpreted or otherwise processed in a suitable manner if necessary, and then stored in a computer memory.
It should be understood that portions of the present application may be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, the various steps or methods may be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. If implemented in hardware, as in another embodiment, any one or combination of the following techniques, which are known in the art, may be used: a discrete logic circuit having a logic gate circuit for implementing a logic function on a data signal, an application specific integrated circuit having an appropriate combinational logic gate circuit, a Programmable Gate Array (PGA), a Field Programmable Gate Array (FPGA), or the like.
It will be understood by those skilled in the art that all or part of the steps carried by the method for implementing the above embodiments may be implemented by hardware related to instructions of a program, which may be stored in a computer readable storage medium, and when the program is executed, the program includes one or a combination of the steps of the method embodiments.
In addition, functional units in the embodiments of the present application may be integrated into one processing module, or each unit may exist alone physically, or two or more units are integrated into one module. The integrated module can be realized in a hardware mode, and can also be realized in a software functional module mode. The integrated module, if implemented in the form of a software functional module and sold or used as a stand-alone product, may also be stored in a computer readable storage medium.
The storage medium mentioned above may be a read-only memory, a magnetic or optical disk, etc. Although embodiments of the present application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present application, and that variations, modifications, substitutions and alterations may be made to the above embodiments by those of ordinary skill in the art within the scope of the present application.
Claims (9)
1. A weak supervised learning method based on memory guidance is characterized by comprising the following steps:
encoding each old data in the old data set with the correct label to obtain vector information of each old data, and training a prediction model according to the vector information of each old data and the correct label of each old data;
constructing a memory pool based on the vector information of each old data and the correct label of each old data; the memory pool comprises a plurality of memory information, and each memory information is used for representing the corresponding relation between data vector information and a label;
when new data is learned, encoding the new data to obtain vector information of the new data, extracting at least part of memory information from the memory pool, inputting the at least part of memory information and the vector information of the new data into the prediction model for training, and obtaining a prediction label aiming at the new data;
acquiring a reference label according to the at least part of memory information, and judging the prediction label according to the reference label;
forming a new memory information by the evaluation result and the vector information of the new data and storing the new memory information into the memory pool;
and continuously repeating the learning process of the new data to update the model parameters until the classification model is output after the new data of the batch are completely processed.
2. The method of claim 1, wherein the extracting at least part of the memory information from the memory pool comprises:
performing similarity calculation on vector information of the new data and data vector information in each memory information in the memory pool;
and taking the memory information of which the similar calculation result meets the similar memory matching index as the at least part of memory information.
3. The method of claim 2, wherein the similarity calculation is performed by the following formula:
dist(xi,mj)＝||xi-mj||2
wherein x isiVector information, m, representing the new datajRepresents data vector information, dist (x) in jth memory information in the memory pooli,mj) Indicating the similarity degree of the vector between the new data and the j-th memory information calculated by the L2 norm.
4. The method according to claim 1, characterized in that when learning new data, if the new data is not labeled, the inputs of the prediction model are the at least partial memory information and the unlabeled data to be learned; if the new data is marked, temporarily storing the label carried by the new data as label-free data, and using the label and the at least part of memory information as the input of the prediction model.
5. The method of claim 4, wherein said retrieving a reference tag from said at least partially remembered information comprises:
calculating the distance between the vector information of the unlabeled data input by the prediction model and each data vector information in the at least partial memory information;
taking the distance between the data vector information as the label weight of the data vector information;
and acquiring the reference label according to the label corresponding to each data vector information in at least part of memory information and the label weighting of each data vector information.
6. The method of claim 4, wherein the inputting the at least partial memory information and the vector information of the new data into the prediction model training, resulting in a prediction label for the new data, comprises:
performing data enhancement for R times on the unlabeled data input by the prediction model to obtain R enhanced data, and respectively obtaining R prediction classification probabilities from the R enhanced data through the prediction model;
acquiring a reference label according to the at least part of memory information;
processing the average of the R prediction classification probabilities and the reference label based on a sharpening function to obtain a prediction label for the new data.
7. The method of any one of claims 1 to 6, further comprising:
constructing a loss function, wherein the constructed loss function integrates cross entropy loss of labeled data and norm L2 loss after unlabeled data enhancement;
calculating loss values of the reference label and the prediction label based on the loss function, and updating model parameters of the prediction model according to the loss values; wherein the loss function is represented as follows:
LOSS°＝Lx+λuLu
wherein | X | and | U | are the number of data samples captured in one training, and the numerical values correspond to the size of a standard data set and a standard-free data set; l is the number of classification categories, namely the number of all categories in the memory pool; x, l are labeled data and labels thereof, u, ypIs the unlabeled data and its predictive label; h (l, ° P (y | x; θ)) is a cross entropy function; lambda [ alpha ]uIs a balance factor for balancing supervision loss and unsupervised loss.
8. A computer device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, characterized in that the processor implements the memory-based weakly supervised learning method of any one of claims 1 to 7 when executing the program.
9. A computer-readable storage medium, on which a computer program is stored, which, when being executed by a processor, carries out a memory-guidance-based weakly supervised learning method as claimed in any one of claims 1 to 7.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110048718.8A CN112734037A (en) | 2021-01-14 | 2021-01-14 | Memory-guidance-based weakly supervised learning method, computer device and storage medium |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110048718.8A CN112734037A (en) | 2021-01-14 | 2021-01-14 | Memory-guidance-based weakly supervised learning method, computer device and storage medium |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN112734037A true CN112734037A (en) | 2021-04-30 |
Family
ID=75593067
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110048718.8A Pending CN112734037A (en) | 2021-01-14 | 2021-01-14 | Memory-guidance-based weakly supervised learning method, computer device and storage medium |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112734037A (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113255807A (en) * | 2021-06-03 | 2021-08-13 | 北京的卢深视科技有限公司 | Face analysis model training method, electronic device and storage medium |
| CN115082955A (en) * | 2022-05-12 | 2022-09-20 | 华南理工大学 | Deep learning global optimization method, recognition method, device and medium |
| CN116524297A (en) * | 2023-04-28 | 2023-08-01 | 迈杰转化医学研究(苏州)有限公司 | Weak supervision learning training method based on expert feedback |
-
2021
- 2021-01-14 CN CN202110048718.8A patent/CN112734037A/en active Pending
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113255807A (en) * | 2021-06-03 | 2021-08-13 | 北京的卢深视科技有限公司 | Face analysis model training method, electronic device and storage medium |
| CN113255807B (en) * | 2021-06-03 | 2022-03-25 | 北京的卢深视科技有限公司 | Face analysis model training method, electronic device and storage medium |
| CN115082955A (en) * | 2022-05-12 | 2022-09-20 | 华南理工大学 | Deep learning global optimization method, recognition method, device and medium |
| CN115082955B (en) * | 2022-05-12 | 2024-04-16 | 华南理工大学 | Deep learning global optimization method, recognition method, device and medium |
| CN116524297A (en) * | 2023-04-28 | 2023-08-01 | 迈杰转化医学研究(苏州)有限公司 | Weak supervision learning training method based on expert feedback |
| CN116524297B (en) * | 2023-04-28 | 2024-02-13 | 迈杰转化医学研究(苏州)有限公司 | Weak supervision learning training method based on expert feedback |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112232416B (en) | Semi-supervised learning method based on pseudo label weighting | |
| CN108960073B (en) | Cross-modal image mode identification method for biomedical literature | |
| CN112734037A (en) | Memory-guidance-based weakly supervised learning method, computer device and storage medium | |
| CN110866536A (en) | Cross-regional enterprise tax evasion identification method based on PU learning | |
| CN109522961B (en) | Semi-supervised image classification method based on dictionary deep learning | |
| CN113807420A (en) | Domain self-adaptive target detection method and system considering category semantic matching | |
| CN113688851B (en) | Data labeling method and device and fine granularity identification method and device | |
| CN114357221B (en) | Self-supervision active learning method based on image classification | |
| CN111813618A (en) | Data anomaly detection method, device, equipment and storage medium | |
| CN111241987B (en) | Multi-target model visual tracking method based on cost-sensitive three-branch decision | |
| CN115471739A (en) | Cross-domain remote sensing scene classification and retrieval method based on self-supervision contrast learning | |
| CN115439685A (en) | Small sample image data set dividing method and computer readable storage medium | |
| CN111191033A (en) | Open set classification method based on classification utility | |
| CN113705215A (en) | Meta-learning-based large-scale multi-label text classification method | |
| CN113128410A (en) | Weak supervision pedestrian re-identification method based on track association learning | |
| Myers-Dean et al. | Generalized few-shot semantic segmentation: All you need is fine-tuning | |
| CN114913546A (en) | Method and system for detecting character interaction relationship | |
| CN117195031A (en) | Electromagnetic radiation source individual identification method based on neural network and knowledge-graph dual-channel system | |
| CN114462526A (en) | Classification model training method and device, computer equipment and storage medium | |
| CN114418111A (en) | Label prediction model training and sample screening method, device and storage medium | |
| CN114678060A (en) | Protein modification method based on amino acid knowledge map and active learning | |
| CN116580272A (en) | Radar target classification method and system based on model fusion reasoning | |
| CN114648094A (en) | Machine learning apparatus and method | |
| Dyballa et al. | A separability-based approach to quantifying generalization: which layer is best? | |
| CN116975695B (en) | Limb movement recognition system based on multi-agent reinforcement learning |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20210430 |