Learning sample selection method in active learning process

Abstract

The invention discloses a learning sample selection method in an active learning process, and the method comprises the steps: carrying out the one-time feature embedding of original data through a pre-training model, and carrying out the selection of iterative sample data through a miniature neural network; and based on the iterative sample data, performing multiple rounds of iterative selection processing processes based on the micro neural network and the sample pooling process to realize selection of learning samples. According to the method, the information richness of the sample can be directly predicted by using the lightweight neural network, and the candidate sample of each round of sample information richness measurement can be reduced according to the historical information richness measurement result, so that the problem of long time consumption of data selection when the unlabeled data volume is huge is solved; the task model with high accuracy can be obtained by using less marking time budgeting.

Description

Technical Field

[0001] The present invention relates to the technical field of big data analysis and processing, and in particular to a method for selecting learning samples in an active learning process. Background Art

[0002] In the process of implementing large-scale data analysis services, efficient and regular model update conditions are required to adapt to the possible changes in data distribution. For the corresponding large-scale samples, the workload of manually labeling all samples that can be used for task model updates is too large, making it impossible to implement. To solve this problem, the prior art proposes to use active learning technology to iteratively select a subset of samples with high information richness for labeling to achieve the labeling of learning samples. However, through the experience of applying the corresponding active learning in the actual video analysis system, a previously overlooked time cost bottleneck problem was discovered, namely the problem of excessively high time cost of data selection.

[0003] That is, existing active learning methods select data by estimating the information richness of all unlabeled samples in each iteration (e.g., their output confidence, etc.). In the corresponding analysis system, this data selection process may account for a large amount of time cost of the entire model update. Experiments have found that the corresponding time cost accounts for 42.8% of the time cost of the entire model update. The total time cost of the corresponding entire model update includes the time for data selection, labeling, and model update.

[0004] Therefore, the existing data selection methods for active learning have the problem of taking up too much time, which in turn leads to the fact that the efficiency of the corresponding entire model update cannot be guaranteed. Summary of the invention

[0005] The purpose of the present invention is to provide a method for selecting learning samples in an active learning process, so as to improve the efficiency of sample selection in the active learning process, thereby solving the above-mentioned technical problems existing in the prior art.

[0006] The objective of the present invention is achieved through the following technical solutions:

[0007] A method for selecting learning samples in an active learning process, comprising:

[0008] The pre-trained model is used to perform one-time feature embedding on the original data. After the one-time feature embedding is completed, the micro neural network is used to iteratively select sample data.

[0009] Based on the iterative sample data, multiple rounds of iterative selection processing are performed based on the micro neural network and the sample pooling process to achieve the selection of learning samples; wherein, based on the micro neural network, information measurement of sample data can be performed, and the information richness of the sample data after each round of iterative processing is obtained, and the sample pooling process is a process of reducing the current sample data based on the information richness of the sample data of two adjacent rounds.

[0010] The one-time feature embedding process includes:

[0011] A pre-trained natural language processing model is used to obtain the embedding of the text input, and four fully connected layers are connected after the embedding to make informative predictions; and as a benchmark, a prediction model with the original file as input is also used, which also uses the said four fully connected layers.

[0012] The process of reducing the current sample data based on the information richness of two adjacent rounds of sample data includes:

[0013] Based on the information richness of two adjacent rounds of sample data and the determined sample discard rate, sample data reduction processing is performed on the current sample data; wherein, the sample discard rate can be used to determine the amount of sample data selected to be retained in the total current sample data.

[0014] The sample data reduction process includes:

[0015] The sample discard rate is determined based on a pipeline strategy, and sample data reduction processing is performed based on the sample discard rate based on the pipeline strategy.

[0016] The process of performing sample data reduction processing on the current sample data includes:

[0017] By comparing the comparative measurement results of the information richness of two adjacent rounds of sample data, if the information richness of the two adjacent rounds of sample data changes greatly, the comparative measurement result is larger, otherwise, it is smaller;

[0018] Based on the comparison and measurement results, the sample data after the reduction processing is obtained by selecting from the sample data based on the sample information discarding rate from the largest to the smallest.

[0019] In the process of determining the sample discard rate, it is necessary to achieve load balancing between the various stages of the pipeline.

[0020] The formula for determining the sample discard rate includes:

[0021]

[0022] Among them, K is the number of samples selected and labeled in each round, T a ,Tt , T e represents the unit time cost of annotation, training, and prediction for each sample, where 0 < q ≤ 1, N0 is the maximum number of rounds for pooling operations预定的进行池化操作的最大轮数, N is the number of unlabeled samples initially, and n represents the current nth round.

[0023] The said T a , T t , T e The determination formula of

[0024]

[0025] Compared with the prior art, the implementation solution provided by the present invention can directly predict the information richness of samples using a lightweight neural network through an efficient proxy information estimation method. At the same time, it can also narrow down the candidate samples for measuring the information richness of samples in each round according to the measurement results of historical information richness, thereby solving the problem of time-consuming data selection when the amount of unlabeled data is large. Furthermore, it enables obtaining a task model with high accuracy using less annotation time budget. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] To more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0027] Figure 1 is a schematic flowchart of the method for selecting active learning samples provided by the embodiment of the present invention;

[0028] Figure 2 is a schematic comparison diagram of the active learning sample annotation processes with and without using a three-parallel pipeline. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0029] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in combination with the specific content of the present invention; obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments, which does not constitute a limitation to the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the protection scope of the present invention.

[0030] First, the following explanations will be given for the terms that may be used in this article:

[0031] The term “and / or” means that either or both of them can be realized at the same time. For example, X and / or Y means both “X” or “Y” and “X and Y”.

[0032] The terms "include", "comprises", "contains", "has" or other descriptions with similar semantics should be interpreted as non-exclusive inclusion. For example, including certain technical feature elements (such as raw materials, components, ingredients, carriers, dosage forms, materials, dimensions, parts, components, mechanisms, devices, steps, procedures, methods, reaction conditions, processing conditions, parameters, algorithms, signals, data, products or products, etc.) should be interpreted as including not only certain technical feature elements explicitly listed, but also other technical feature elements known in the art that are not explicitly listed.

[0033] The term "consisting of..." means excluding any technical feature elements not explicitly listed. If this term is used in a claim, it will make the claim closed, so that it does not contain technical feature elements other than the technical feature elements explicitly listed, except for the conventional impurities related to them. If this term only appears in a clause of a claim, it only limits the elements explicitly listed in the clause, and the elements recorded in other clauses are not excluded from the overall claim.

[0034] Unless otherwise specified or limited, the terms "installed", "connected", "connected", "fixed" and the like should be understood in a broad sense, for example: it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be an indirect connection through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in this article can be understood according to specific circumstances.

[0035] When concentration, temperature, pressure, size or other parameters are expressed in the form of a numerical range, the numerical range should be understood to specifically disclose all ranges formed by the pairing of any upper limit, lower limit, and preferred value in the numerical range, regardless of whether the range is explicitly stated; for example, if a numerical range of "2 to 8" is stated, the numerical range should be interpreted as including ranges such as "2 to 7", "2 to 6", "5 to 7", "3 to 4 and 6 to 7", "3 to 5 and 7", "2 and 5 to 7", etc. Unless otherwise specified, the numerical ranges stated herein include both their end values ​​and all integers and fractions within the numerical range.

[0036] The orientation or position relationship indicated by terms such as "center", "longitudinal", "lateral", "length", "width", "thickness", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", "clockwise", "counterclockwise", etc. are based on the orientation or position relationship shown in the drawings and are only for the convenience and simplification of description, and do not explicitly or implicitly indicate that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as a limitation of this document.

[0037] In the implementation process of the present invention, in order to solve the time cost bottleneck problem caused by data selection in the prior art, a corresponding implementation scheme of proxy informativeness estimation is specifically proposed. Specifically, the time cost of data selection can be modeled first, from which three key factors are determined, namely: unit estimation cost, the number of samples used for estimation, and the number of iterations; wherein the influence of the first two factors gradually accumulates with the increase in the number of iterations; accordingly, the present invention designs a proxy estimator and a sample aggregation scheme. The proxy estimator can be a lightweight neural network, which can be trained to directly predict the informativeness of the sample, and it can replace the role of the high-cost task model in the original estimation, thereby reducing the unit cost. On the other hand, the sample aggregation scheme of the present invention uses historical estimation results to narrow the range of sample candidates.

[0038] Furthermore, in order to achieve the above technical effects, the present invention will face the following two key challenges in the specific implementation process, that is, two technical problems that need to be solved:

[0039] (1) Lightweight estimation (i.e., information measurement) problem

[0040] The information richness estimation process must be lightweight so that the unit estimation cost can be reduced; it is obvious that a lightweight model can perform faster predictions, so the present invention needs to solve the problem of how to make a lightweight estimator replace the high-cost task model in the original information richness estimation process;

[0041] (2) Problems with narrowing down the candidate pool

[0042] From the perspective of cost modeling, the data selection cost is linearly positively correlated with the number of candidate samples involved; the fewer candidate samples involved, the faster the prediction. Therefore, the present invention also needs to solve the problem of how to use existing information to narrow the range of sample candidates for information richness estimation.

[0043] Based on the two main technical problems that need to be solved above, the present invention provides a corresponding method for selecting learning samples in the active learning process, and specifically provides a corresponding efficient active learning sample selection method based on agent information estimation. During the implementation process, the present invention mainly involves fields such as data labeling methods, active learning, sample information richness estimation, and data selection methods.

[0044] The purpose of the present invention is to design an efficient proxy information estimation method so that the information richness of samples can be directly predicted using a lightweight neural network, and the candidate samples for each round of sample information richness measurement can be narrowed down according to the historical information richness measurement results to solve the problem of long data selection time when the amount of unlabeled data is huge. This allows a task model with higher accuracy to be budgeted using less labeling time.

[0045] To facilitate understanding of the present invention, a specific implementation scheme of a method for selecting learning samples in an active learning process provided by an embodiment of the present invention will be explained in detail below with reference to the accompanying drawings.

[0046] In the specific implementation process of the embodiment of the present invention, the selection process of the corresponding learning sample can be specifically but not limited to the corresponding concurrent video packet filtering process, etc., and its main processing process can be as follows: Figure 1 As shown, the following three processing steps may be included:

[0047] (i) Using one-shot feature embedding and lightweight information metrics for iterative data selection;

[0048] Specifically, data embedding can be performed by using a one-time feature embedding method, and after the one-time embedding is completed, iterative sample data selection is performed through a micro neural network; thereafter, subsequent multiple rounds of iterative data selection processing can be performed based on the iterative sample data;

[0049] The corresponding lightweight information measurement process can be implemented by a lightweight neural network (or micro neural network), which can also be called a proxy information estimator (referred to as proxy estimator), which can directly learn information measurement to reduce the unit information measurement cost in each iteration;

[0050] The specific implementation process of the processing (I) is as follows:

[0051] In each round of active learning, a portion of unlabeled data needs to be selected for labeling. Experiments have found that when active learning is applied in video analysis systems, the corresponding data selection process becomes a bottleneck of time cost. In order to understand this bottleneck, we can first model the time cost of data selection and obtain the following formula:

[0052] Labeling cost = unit cost × number of samples × number of iterations;

[0053] Among them, for existing active learning methods, the "unit cost" is the time cost of running the current task model on unlabeled samples and measuring informativeness; the corresponding "number of samples" and "number of iterations" usually represent the number of all unlabeled samples and the number of active learning iterations, respectively. It can be seen from the above cost formula that when the task model is large (for example, a deep neural network with billions of parameters), the unit cost may be very expensive because it involves inference of the task model. On the other hand, when the number of unlabeled samples is large (for example, millions of images), the overhead of data selection is unacceptable even for smaller task models. The impact of these two factors also gradually accumulates with the increase in the number of active learning rounds. The main purpose of the present invention in the implementation process is to focus on reducing the costs caused by these two key factors.

[0054] In the traditional active learning iteration process, the same sample needs to be repeatedly estimated for information richness in each round. Information richness requires the use of a task model to embed the original data into the feature space, and then measure the informativeness based on these embeddings. Therefore, the unit cost can be divided into: feature embedding cost Embed and information measurement cost Measure, and the corresponding data selection cost (i.e., annotation cost) is as follows:

[0055]

[0056] In the above learning iterations, the repeated feature embedding process takes up most of the time.

[0057] In order to reduce the cost of repeated feature embedding, the present invention explores the implementation method of one-time embedding using a pre-trained embedding model through experiments; specifically, for a text-based emotion recognition task, the processing process that can be used includes:

[0058] First, use a pre-trained natural language processing model, such as Bert, GPT2, etc. (each is a natural language processing model) to obtain the embedding of the text input, that is, perform a one-time feature embedding on the original data;

[0059] Then, four fully connected layers are connected after the embedding to predict the informativeness (i.e., perform an information measure); as a baseline, the present invention also implements a prediction model that takes the original text as input and uses the same four fully connected layers.

[0060] It is found in experiments that using pre-trained embeddings, lightweight neural networks can effectively learn information metrics; therefore, the present invention can use the pre-trained model to perform a one-time feature embedding of the original data, and use a lightweight neural network, called a proxy estimator, to directly predict informativeness based on the embedding; based on this, the formula of the corresponding data selection cost is rewritten as:

[0061] Data selection cost = Embed × number of samples × 1 + Measure × number of samples × number of iterations;

[0062] It can be seen that the above-mentioned one-time feature extraction design can reduce the time cost of repeated and redundant feature embedding from "Embed×number of samples×number of iterations" to "Embed×number of samples×1", achieving the purpose of the present invention, namely, reducing the unit estimation cost.

[0063] (ii) Sample pooling process based on the information richness of samples;

[0064] That is, the estimation results of the previous round are used to filter the low-information samples, so that the sample candidate range can be more effectively narrowed down in the next round of processing; that is, based on the iterative sample data obtained in the above processing process (I), the micro neural network and the sample pooling process are used to perform multiple rounds of iterative selection processing to achieve the selection of learning samples; wherein, based on the micro neural network being able to perform information measurement and obtain the information richness of the sample data after each round of iterative processing, the sample pooling process is a process of narrowing down the current sample data based on the information richness of the sample data of two adjacent rounds;

[0065] That is, for the sample quantity factor, the present invention designs a sample pooling module to narrow the sample candidate range for information richness estimation, and the processing performed by the sample pooling module may include: first, the comparative measurement results of the information richness of two adjacent rounds of samples can be obtained by comparison, if the information richness of the two adjacent rounds of samples changes greatly, the larger the comparative measurement result, and vice versa, the smaller; then, based on the comparative measurement results from large to small, the sample data after the narrowing process is selected in the samples based on the sample information discarding rate, wherein the sample discarding rate can be used to determine the amount of sample data selected to be retained in the total current sample data;

[0066] Specifically, in the process of implementing the present invention, in order to find opportunities to reduce the range of sample candidates, the informativeness changes between two rounds were studied; specifically, experiments were conducted on a text-based emotion recognition task, using a minimum confidence strategy and 6809 samples; by comparing the measurement results of the information richness of samples in two adjacent rounds, it was found that the high information richness samples had a great change in the information richness measurement results of the next round, while the low information richness samples remained stable; therefore, the sample pooling module of the present invention filtered out samples with low information richness predicted in the previous round, so as to narrow the range of sample candidates predicted by the proxy estimator, thereby further reducing the overall data selection cost;

[0067] In statistics, "sampling" is to obtain a subset with the same distribution as the original data set; in contrast, the present invention aims to obtain a subset within a specific distribution (a subset with high information richness); this is more like a "pooling" strategy, so it can be called "sample pooling"; specifically, in each round of active learning, the (1-q) proportion of samples with the lowest information richness will be eliminated, where 0 <q≤1。

[0068] (iii) Use pipeline strategy to select appropriate sample discard rate;

[0069] like Figure 2 As shown, in the processing process (three), the pipeline strategy can be used to select the sample discard rate, and in the process of determining the sample discard rate, load balancing can be achieved between the various stages of the pipeline, thereby determining the sample discard rate.

[0070] Specifically, the principle of determining the sample discard rate q comes from the advantages of the technical solution provided by the present invention, namely: it can support pipeline processing strategy;

[0071] For traditional active learning methods, the sample data selection for the next round must wait until the task model training is completed; this dependency prevents the overlapping execution of jobs, resulting in stagnation of the job process. The adoption of the proxy estimator (i.e., micro-neural network) of the present invention can eliminate the dependence of data selection on the task model, so the technical solution provided by the present invention can support pipeline processing; refer to Figure 2 As shown, it shows a comparison of active learning annotation processes with and without pipeline processing; in order to make full use of the pipeline, it is also possible to choose to maintain a balanced workload between the stages of the pipeline; therefore, the present invention can determine the value of q according to the following equation:

[0072] Among them, N represents the number of unlabeled samples at the beginning, n represents the current round n; N0 is the predefined maximum number of rounds for pooling operation.

[0073] Among them, K is the number of samples selected and labeled in each round, T a ,T t ,T e represents the unit time cost of annotation, training, and proxy estimator prediction for each sample;

[0074] The above equation is obtained by making the data selection time equal to the labeling and training time, that is:

[0075] For round n,

[0076] Using this equation and the condition q≤1, we can derive the expression for q and then determine the value of q.

[0077] The sample discard rate q value determined by the above processing process (three) can more effectively implement the corresponding sample pooling processing operation in the above processing process (two).

[0078] In order to further understand the specific application effects of the above-mentioned embodiments of the present invention, the specific application implementation process thereof will be compared with the implementation scheme in the prior art to illustrate the effects.

[0079] Specifically, a real-time video analysis system was developed for campus security at the university, in which 2,529 cameras were installed in public areas to generate millions of video frames every day. In order to adapt to the dynamic changes that may occur in campus data distribution and ensure timely detection of security incidents, the detection model needs to be regularly updated with newly generated videos every day. Obviously, it is impractical to annotate all video frames. Therefore, in the annotation process, an experimental comparison was carried out using the traditional annotation method and the annotation method provided by the present invention:

[0080] (1) We used the classic active learning method based on minimum output confidence and found that the data selection process took too long (accounting for 42% of the total time cost), becoming an efficiency bottleneck. The reason is very intuitive, because existing active learning methods require the task model to predict all unlabeled data and measure the information richness to select data based on these predictions of the task model. Therefore, for a large number of unlabeled datasets or large task models, the iterative data selection process can easily dominate the active learning process.

[0081] (2) Combining the technical solution provided by the present invention with the existing active learning method can save 66% of the labeling time while achieving the same model accuracy.

[0082] In summary, the efficient active learning sample selection scheme based on proxy information estimation provided by the present invention can directly predict the information richness of samples using a lightweight neural network through an efficient proxy information estimation method. At the same time, the candidate samples for each round of sample information richness measurement can be narrowed down according to the historical information richness measurement results to solve the problem of time-consuming data selection when the amount of unlabeled data is huge, thereby obtaining a high-accuracy task model using a smaller annotation time budget. And by combining the embodiment of the present invention as a plug-in with the most advanced active learning annotation algorithm, while achieving the same accuracy of the task model, after combining the embodiment of the present invention, 53.6-89.1% of the annotation time can be saved in the active learning annotation process.

[0083] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by any technician familiar with the technical field within the technical scope disclosed in the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims. The information disclosed in the background technology section of this article is only intended to deepen the understanding of the overall background technology of the present invention, and should not be regarded as an admission or in any form that the information constitutes prior art known to those skilled in the art.

Claims

1. A method for selecting learning samples in an active learning process, characterized in that: include: The pre-trained model is used to perform one-time feature embedding on the original data, and after the one-time feature embedding is completed, the micro neural network is used to iteratively select sample data; Based on the iterative sample data, multiple rounds of iterative selection processing are performed based on the micro neural network and the sample pooling process to achieve the selection of learning samples; wherein, based on the micro neural network, information measurement of sample data can be performed, and the information richness of the sample data after each round of iterative processing is obtained, and the sample pooling process is a process of reducing the current sample data based on the information richness of the sample data of two adjacent rounds.

2. The method according to claim 1, characterized in that: The one-time feature embedding process includes: A pre-trained natural language processing model is used to obtain the embedding of the text input, and four fully connected layers are connected after the embedding to make informative predictions; and as a benchmark, a prediction model with the original file as input is also used, which also uses the said four fully connected layers.

3. The method according to claim 1, characterized in that The process of reducing the current sample data based on the information richness of two adjacent rounds of sample data includes: Based on the information richness of two adjacent rounds of sample data and the determined sample discard rate, sample data reduction processing is performed on the current sample data; wherein, the sample discard rate can be used to determine the amount of sample data selected to be retained in the total current sample data.

4. The method according to claim 3, characterized in that The sample data reduction process includes: The sample discard rate is determined based on a pipeline strategy, and sample data reduction processing is performed based on the sample discard rate based on the pipeline strategy.

5. The method according to claim 3, characterized in that: The process of performing sample data reduction processing on the current sample data includes: By comparing the comparative measurement results of the information richness of two adjacent rounds of sample data, if the information richness of the two adjacent rounds of sample data changes greatly, the comparative measurement result is larger, otherwise, it is smaller; Based on the comparison and measurement results, the sample data after the reduction processing is obtained by selecting from the sample data based on the sample information discarding rate from the largest to the smallest.

6. The method according to any one of claims 1 to 5, characterized in that: In the process of determining the sample discard rate, it is necessary to achieve load balancing between the various stages of the pipeline.

7. The method according to claim 6, characterized in that The formula for determining the sample discard rate includes: Among them, K is the number of samples selected for annotation in each round, T a , T t , T e represents the unit time cost of annotation, training, and prediction for each sample, and 0 < q ≤ 1, N0 is the maximum number of rounds for the pooling operation, N is the number of unlabeled samples at the beginning, and n represents the nth round currently.

8. The method according to claim 7, characterized in that The T a ,T t ,T e The formulas for determining include:

Images