Data tag verification

Abstract

Aspects of the invention disclose a method for validating labels of records of a dataset. The records include sample data and a relevant label of a plurality of labels. The method includes one or more processors dividing the dataset into a training dataset and an inference dataset, the training dataset including records relevant to a selected label, the inference dataset including records having sample data relevant to the selected label and all other labels of the plurality of labels. The method further includes dividing the training dataset into a plurality of learner training datasets, the plurality of learner training datasets including at least one sample relevant to the selected label. The method further includes training a plurality of label-specific few-shot learners with one of the learner training datasets. The method further includes performing inference on the inference dataset by the plurality of trained label-specific few-shot learners to generate a set of a plurality of predicted label output values.

Description

BACKGROUND

[0001] The present invention relates generally to label quality of a dataset, and more specifically, to verifying labels of records of a labeled dataset.

[0002] Artificial intelligence (AI) and machine learning (ML) are taking center stage in enterprises of any size and in every vertical industry. While users are gaining more and more experience with ML technology, the training of algorithms remains a very time-consuming task. One area of burden is the provision of good training data, i.e. correctly labeled data for supervised learning. Thus, mislabeled data is a problem for a range of applications and methods in the field of machine learning. The performance of a machine learning classifier depends to a large extent on the quality of the data labels used for training and validation. Typically, human supervision is required to obtain labels for a dataset. Thus, the labeling process a) is time-consuming, b) can contain errors (mislabeling). However, it should be noted that mislabeling can not only occur due to human error, but mislabeling can be an inherent part of collecting data, as described in several publications.

[0003] Detecting and correcting mislabeled data is of interest to enterprise users, as mislabeled data directly impacts ML model performance. As datasets become larger, the process of manually verifying labels is not feasible in terms of time, leading to the need for automatic detection of mislabeled data.

[0004] The next wave of AI innovation can be facilitated by deep learning models trained using methods that lie somewhere between supervised and unsupervised learning. Computer scientists and engineers are exploring many such learning methods, some of which can bring a triple threat: less labeled data, less data volume, and less human intervention. Among them, “one-shot learning” is the closest to unsupervised learning. One-shot learning is based on the premise that most human learning occurs when receiving only one or two examples.

[0005] One-shot or few-shot learning is generally based on a “quick transfer” of previously learned concepts to new situations or objects not yet encountered. This can be achieved by hard-coding, a well-defined rule set in the respective model, or by leveraging a prior in a Bayesian approach. Thus, the idea is based on facilitating learning by shared generalizable concepts.

[0006] Once the concept of a characteristic for a broad class has been learned, the concept can need to identify a characteristic for none of the classes available or allow for example isolation is available for a class not seen before (identified by an independent label).

[0007] Some documents belonging to the general context of the concept presented here include document WO 2018 / 226492 Al. This document discloses a method and computer system for improving training to a base deep neural network by structurally altering the base deep neural network to create an updated deep neural network such that the updated deep neural network does not degrade in performance on training data relative to the base deep neural network.

[0008] In addition, document US 8,331,657 B2 discloses a computer-implemented pattern recognition method comprising electronically creating links between a plurality of models within a classifier module within a pattern recognition system such that any one of the plurality of models can be selected as an active model in a recognition process, electronically creating a null hypothesis between at least one model of the plurality of linked models and at least a second model of the plurality of linked models, electronically accumulating evidence to accept or reject the null hypothesis, and transmitting a portion of the electronically accumulated evidence to the pattern classifier module. However, so far, most known methods rely on well-labeled data.

[0009] Therefore, on the road from supervised learning to fully unsupervised learning via transfer learning, reinforcement, and one-shot learning, the requirement for correctly labeled data remains high. Therefore, embodiments of the present invention propose how to improve the quality of labeled data. SUMMARY

[0010] Aspects of the present invention disclose a method, computer program product, and system for verifying labels of records of a dataset. The records include sample data and a relevant label of a plurality of labels. The method includes one or more processors dividing the dataset into a training dataset and an inference dataset, the training dataset including records relevant to a selected label, the inference dataset including records having sample data relevant to the selected label and all other labels of the plurality of labels. The method further includes one or more processors dividing the training dataset into a plurality of learner training datasets. The learner training datasets include at least one sample relevant to the selected label. The method further includes one or more processors training a plurality of label-specific few-shot learners. The few-shot learners are trained with one of the plurality of learner training datasets. The method further includes one or more processors performing inference on the inference dataset by the plurality of trained label-specific few-shot learners to generate a set of a plurality of predicted label output values.

[0011] In further embodiments, the method further includes the one or more processors aggregating the predicted label output values within each of the sets of predicted label output values, thereby generating respective score values for the plurality of label-specific few-shot learners, the respective score values reflecting variability in the predicted selected label. In response to determining that at least a predefined number of the score values are below a predefined variability threshold, the method further includes the one or more processors determining that the dataset is verified with respect to the selected label.

[0012] In another aspect of the application, the computer program product includes one or more computer readable storage media and program instructions stored on the one or more computer readable storage media, the program instructions comprising program instructions to divide the dataset into a training dataset and an inference dataset, the training dataset including records related to the selected label, the inference dataset including records having sample data related to the selected label and all other labels of the plurality of labels. The computer program product further includes program instructions to divide the training dataset into a plurality of learner training datasets. A learner training dataset includes at least one sample related to the selected label. The computer program product further includes program instructions to train a plurality of label-specific few-shot learners. One of the plurality of learner training datasets is utilized to train the few-shot learner. The computer program product further includes program instructions to perform inference on the inference dataset by the plurality of trained label-specific few-shot learners to generate a plurality of sets of predicted label output values.

[0013] In further embodiments, the computer program product further includes program instructions to aggregate the predicted label output values within each of the sets of predicted label output values, thereby generating respective score values for each of the plurality of label-specific few-shot learners, the respective score values reflecting variability in the predicted selected label. In response to determining that at least a predefined number of the score values are below a predefined variability threshold, the computer program product further includes program instructions to determine that the dataset is verified with respect to the selected label.

[0014] In another aspect of the invention, a computer system comprises one or more computer processors, one or more computer-readable storage media, and program instructions stored on the computer-readable storage media for execution by at least one of the one or more processors, the program instructions comprising program instructions to partition a data set into a training data set and an inference data set, the training data set comprising records related to a selected label, the inference data set comprising records having sample data related to the selected label and all other labels of a plurality of labels. The computer system further comprises program instructions to partition the training data set into a plurality of learner training data sets. A learner training data set comprises at least one sample related to the selected label. The computer system further comprises program instructions to train a plurality of label-specific few-shot learners. A few-shot learner is trained with one of the plurality of learner training data sets. The computer system further comprises program instructions to perform inference on the inference data set by the plurality of trained label-specific few-shot learners to generate a set of predicted label output values.

[0015] In a further embodiment, the computer system further comprises program instructions to aggregate the predicted label output values within each of the set of predicted label output values, thereby generating respective score values for the plurality of label-specific few-shot learners, the respective score values reflecting variability in the predicted selected label. In response to determining that at least a predefined number of the score values are below a predefined variability threshold, the computer system further comprises program instructions to determine that the data set is verified with respect to the selected label.

[0016] The proposed computer-implemented method for verifying labels of records of a data set can provide a plurality of advantages, technical effects, contributions and / or improvements.

[0017] The various aspects of the invention can successfully overcome problems in existing proposals for detecting mislabeled samples that suffer from performance degradation, where the ratio of label noise is too high. In particular, ensemble classifiers trained on the same data set (of a subsample) whose label errors are to be checked can depend on the quality of the data that is trained on. Single-shot or few-shot learning modules trained on examples with known ground truth or known labels with high confidence do not tend to the aforementioned performance limitations; thus, providing an automated solution for successful detection of mislabeled data instances. This is generally important when working with data having labels (i.e., in order to ensure correctness of label data).

[0018] Thus, the proposed method can automatically significantly and reliably increase the label quality of labeled data for machine learning, even under difficult conditions (e.g., high mislabeling). BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1A block diagram illustrating an embodiment of the computer-implemented method of the present invention for verifying labels of records of a dataset according to an embodiment of the present invention.

[0020] Figure 2 A block diagram illustrating a more pictographic form of the method presented herein according to an embodiment of the present invention.

[0021] Figure 3 A series of matrices associated with the same example used in

[0022] Figure 4 A series of matrices associated with the same example used in Figure 3

[0023] Figure 5 A series of matrices associated with the same example used in Figure 3 and Figure 4

[0024] Figure 6 A block diagram illustrating an embodiment of the label verification system for verifying labels of records of a dataset according to an embodiment of the present invention.

[0025] Figure 7 A block diagram illustrating an embodiment of a computing system comprising a label verification system according to Figure 6 DETAILED DESCRIPTION

[0026] It should be noted that embodiments of the invention are described with reference to different subject matters. In particular, some embodiments are described with reference to method type claims whereas other embodiments are described with reference to device type claims. However, a person skilled in the art will gather from the above and the following description that, unless otherwise indicated, any aspect or feature described as pertaining to a particular subject matter can be implemented with a different subject matter, in particular a method type claim with a device type claim or vice versa. Thus, a person skilled in the art will appreciate that a method type claim can be implemented at the device level and vice versa.

[0027] The above-mentioned aspects and other aspects of the present invention will become apparent from examples of embodiments to be described hereinafter, and will be explained with reference to examples of embodiments, but the present invention is not limited thereto. Preferred embodiments of the present invention will be described hereinbelow with reference to the accompanying drawings.

[0028] In the context of the present specification, the following conventions, terms and / or expressions can be used:

[0029] ​​​The term“machine learning” (ML) can denote a number of algorithms and statistical models that a computer system can use in order to effectively perform a specific task without using explicit instructions, but rather relying on patterns and inferences. The behavior of the algorithms is not performed in a procedural manner. ML is seen as a subset of artificial intelligence (AI). ML algorithms build mathematical models based on sample data called“training data” in order to make predictions or decisions without being explicitly programmed to perform the task. Typical implementations of ML can be based on (deep) neural networks.

[0030] The term“inference” can denote the processing of new information by a trained machine learning system. For example, the result of an inference process can be represented as a prediction or a classification into a class by a classifier of unknown data.

[0031] The term“Bayesian inference” can denote a method of statistical inference wherein Bayes’ theorem is used to update the probability of a hypothesis as more evidence or information becomes available. Bayesian inference is an important technique in statistics, especially mathematical statistics. Bayesian updating is particularly important in the dynamic analysis of a sequence of data. A single-sample or small-sample learner can be implemented using a“Bayesian inference” model. Thereby, Bayesian inference derives a posterior probability from two predecessors: a prior probability and a“likelihood function” derived from a statistical model of the observed data. Bayesian inference computes the posterior probability according to Bayes’ theorem.

[0032] The term“data set” can denote a number of records comprising data that can be used to train a machine learning model. In addition to net training data, each record can comprise a label that shall describe the net data. For example, an image of an object can be seen as net training data, wherein a related label describes the content of the image. In an additional example, a scene on a street can show a car; therefore, a label can be“car”. However, other labels can also be used for the same scene (e.g.“traffic light”,“pedestrian”,“street”,“sidewalk”,“crossing”, etc.), as this object can also be shown in the image of the scene.

[0033] The term "one-shot or few-shot learning" can generally be based on "transferring" previously learned concepts to new cases of objects not yet encountered. One-shot or few-shot learning can be implemented by heart-encoding well-defined routes in a respective model or alternatively by exploiting a priori in a Bayesian approach. Thus, the concept is based on facilitating learning by shared generalized concepts. In other words, one-shot learners are object classification problem solvers often found in computer vision. While most machine learning based object classification algorithms require training on hundreds or thousands of samples / images and very large data sets, one-shot learning (or few-shot learning) aims to learn information about object classes from one or only a few training samples / images.

[0034] Once the feature concepts of a broad range of classes have been learned, they can have to recognize classes not seen before (e.g., distinguish between independent labels) or allow isolating examples that are not characteristic of any of the available classes.

[0035] In the following, a detailed description of the figures will be given. All illustrations in the figures are schematic. First, a block diagram of an embodiment of the computer-implemented method of the present invention for verifying labels of records of a data set will be given. Then, further embodiments as well as embodiments of a data label verification system for verifying labels of records of a data set will be described.

[0036] Figure 1 A block diagram of an example embodiment of a computer-implemented method 100 for verifying labels of records of a data set is shown, wherein the records comprise sample data and at least one relevant label of a plurality of labels (e.g., a record can have labels u, w, v, x, y, z). The method 100 comprises, in step 102, dividing the data set into a training data set comprising records relevant to one selected label (i.e., having the label y, such as "cat") and an inference data set comprising records having sample data relevant to the selected label (i.e., label "y") and all other labels of the plurality of labels (i.e., labels u, w, v, x, z in the above example).

[0037] The method 100 further comprises, in step 104, dividing the training data set into a plurality of learner training data sets (randomly, or according to another predefined method). The learner data sets can each comprise at least one sample relevant to the selected (relevant) label. The method 100 then comprises, in step 106, training a plurality of label-specific few-shot learners with one of the plurality of learner training data sets. In an example embodiment, the classifier implemented as a learner is a classifier with two output classes: good label y or bad label y, according to the above embodiment.

[0038] Further, the method 100 comprises performing inference on the inference dataset (i.e. typically the complete inference dataset) by the plurality of trained label-specific few-shot learners in step 110. In various embodiments, performing inference can result in a plurality of sets of predicted label output values (i.e. one set for each few-shot learner, which, due to the direct one-to-one relationship, is equivalent to one set for each learner training data set).

[0039] In addition, the method 100 comprises aggregating the predicted label output values within each of the sets of predicted label output values in step 112. Thereby, embodiments of the present invention can determine and provide a respective score value for each of the plurality of label-specific few-shot learners. The respective score value reflects the variability of the predicted selected label. This variability can also be understood as label noise.

[0040] Further, the method 100 comprises determining that the dataset is validated with respect to the selected label (i.e. label “y” (e.g. “cat”)) if at least a predefined number (or a predefined percentage) of the score values is below a predefined variability threshold in step 114. Further, a low number of score values can indicate that the label of the record and the related sub-data has a good quality.

[0041] Figure 2 A block diagram 200 is depicted according to embodiments of the present invention, i.e. a pictographic form of the proposed method.

[0042] The initial dataset 202 is split into two parts: an inference dataset 204 and training data, which is again randomly (or according to another algorithm) split, in particular in a learner training dataset 206. In further embodiments, the selected label (e.g. label “y”) can be selected from a potential plurality of labels for the records of the dataset. However, in other embodiments, the records do not correspond to a plurality of labels. For example, if simple animal images have to be classified / labelled, the images can correspond to one label (e.g. “cat”, “dog”,... “dolphin”, etc.).

[0043] Embodiments of the present invention can then train a label-specific (i.e. here for label “y”) learner 208 (i.e. a classifier) using the learner training dataset. The learner training dataset has been pre-trained with another independent dataset to classify according to the selected “y” label.

[0044] Each record in the learner training dataset 206 can produce (e.g. during occurrence) a prediction or a predicted label output value (a confidence level of “good label” or “bad label”). Thus, if the learner training dataset 206 comprises a capital number F of records, also a confidence level value of the query is generated.

[0045] Next, the embodiments of the present application aggregate the predicted label output values (e.g., by determining a mean or expected value according to another algorithm), resulting in a vector 210 of consolidated, predicted label output values. In an example, one dimension of the vector can represent one learner training dataset 206. If a threshold criterion is met, the associated initial dataset 202 can be deemed validated 212. For example, based on whether at least a predefined number of previously determined score values (dimensions of the resulting vector) for the learner training dataset are below a predefined "correct label" threshold.

[0046] Figure 3 A series of matrices 300 is depicted that illustrate a practical example of applying the proposed method 100 according to embodiments of the present application. Matrix 302 represents an initial dataset (i.e., initial dataset). In a first step, a label (e.g., "y") is selected that should be validated. The selected label can be any one of a possible plurality of labels in the initial dataset. The label can be represented by the exemplary black boxes around images 4 and 6 that have the label "cat".

[0047] In a second step, M individual drawings (without replacement) are made from a subset of matrix 302 that have the positive label "y" (i.e., "cat"), which can be named subset A 306. The remaining portion of matrix 304 (with and without label y) is considered the reference, i.e., validation of the inference dataset, also referred to as subset A1+B, 308.

[0048] Then, the above-described training is performed by using single-sample or few-sample learners, resulting in M trained models 310 (e.g., LI, L2, L3, L4, and potentially more models). In a fourth step, the previously trained M classifiers (i.e., single-sample or few-sample learners) are used on the samples of subset A1+B, 308, which results in a predicted label for each model and each sample in A1+B, 308.

[0049] In a fifth step, the classification result values (i.e., predicted label output values) for each sample in A1+B, 308 are aggregated (e.g., mean values). The main metric observed is the predicted label (large variation in effectiveness corresponding to low confidence), which can also be formulated as label entropy. Then, in a sixth step, it is determined whether a threshold criterion has been exceeded for subset A1+B, 308 (e.g., compare Figure 2 , 212). Matrix 312 depicts an example result of the above-described process. The two leftmost columns of matrix 312 correspond to matrix 308 (A1+B, 308). The "mean" column and "std" column (standard deviation) represent the aggregated results.

[0050] As a standard, the predicted value (P val, Ha > 50%) is depicted in the second column from the right. The classification can be done according to the following traffic light algorithm: "cat": not "cat"; not "cat": fail: The above part of the procedure can be seen as a first iteration (i = 1). Figure 4 The second iteration is explained in

[0051] Figure 4 A series of matrices 400 associated with the example used in Figure 3 is described during the second iteration (i = 2) according to embodiments of the present invention.

[0052] As y label, again "cat" is used for the optimization procedure, as can be seen in matrix 404. Again, a split of the training data set (A-A2, 406) and the interference data set (A2+B, 408) is done, resulting in further models 410 L1', L2', L3', L4' of the exemplary learner.

[0053] Matrix 412 depicts the result matrix of the above Figure 3 and Figure 4 procedures, with the same end criteria as in Figure 3 and the conditions for the traffic light scheme, the second row from the bottom of the matrix, P val = 0.876, marks "fail".

[0054] In an additional step of the proposed method, embodiments of the present invention can perform additional iterations (i = 3, 4, 5,...). If all samples are classified as "cat" or "not cat", the procedure can move to Figure 5 the next label shown.

[0055] Figure 5 A series of matrices 500 is described according to embodiments of the present invention with the same example used in Figure 3 and Figure 4 However, in Figure 5 the label y "dog" is chosen (matrix 504). Given the above explained example in Figure 3 and Figure 4 the matrix 502 with the label "dog" is chosen, the training data set 506 A-A1 is separated from the inference data set B+A1, 508, again resulting in different models 510 L1, L2, L3, L4. As described above, the result matrix of the aggregated predicted label output values is evaluated.

[0056] For completeness reasons, Figure 6A block diagram of a system 600 of a data label verification system for verifying labels of records of a data set according to embodiments of the application is depicted. In various embodiments, the system 600 comprises modules and components operating to perform the processes of the method 100 Figure 1 ) according to embodiments of the application.

[0057] Further, the records comprise sample data and a relevant label of the plurality of labels. The system 600 comprises a partitioning unit 602 adapted to partition the data set into a training data set comprising records relevant to one selected label and an inference data set comprising records having sample data relevant to the selected label and all other labels of the plurality of labels. Accordingly, the partitioning unit 602 is further adapted to randomly partition the training data set into a plurality of learner training data sets, each learner training data set comprising at least one sample relevant to the selected label.

[0058] The system 600 further comprises a training module 604 adapted to train a plurality of label-specific few-shot learners 606, each of the few-shot learners 606 being trained with one of the plurality of learner training data sets. The system 600 further comprises an inference unit 608 adapted to perform inference on the inference data set by the plurality of trained label-specific few-shot learners 606, resulting in a plurality of sets of predicted label output values.

[0059] Further, the system 600 comprises an aggregating module 610 adapted to aggregate the predicted label output values within each of the sets of predicted label output values, resulting in a respective score value for each of the plurality of label-specific few-shot learners corresponding to the relevant specific learner training data set. Accordingly, the respective score value reflects the variability of the predicted selected label, and a determining module 612 is adapted to determine the data set to be verified with respect to the selected label if at least a predefined number of score values is below a predefined variability threshold.

[0060] Embodiments of the application can in principle be implemented together with any kind of computer, regardless of the platform used to store and / or execute program code. Figure 7 As an example, a computing system 700 adapted to execute program code related to the proposed method is shown.

[0061] The computing system 700 is only one example of a suitable computer system and is not intended to suggest any limitation as to the scope of use or functionality of embodiments of the present application described herein, whether or not implementation is

[0062] As shown, the computer system / server 700 is in the form factor of a general- purpose computing device. The components of the computer system / server 700 can include, but are not limited to, one or more processors or processor(s) 702, a system memory 704, and a bus system 706 that couples various system components including the system memory 704 to the processor(s) 702. The bus system 706 represents one or more of several types of bus structures, including a memory bus or memory controller, a peripheral bus, an accelerated graphics port, and a processor or local bus using any of a variety of bus architectures. By way of example, and without limitation, such architectures include Industry Standard Architecture (ISA) bus, Micro Channel Architecture (MCA) bus, Enhanced ISA (EISA) bus, Video Electronics Standards Association (VESA) local bus, and Peripheral Component Interconnect (PCI) bus. The computer system / server 700 typically includes a variety of computer system readable media. Such media can be any available media that is accessible by the computer system / server 700 and it includes both volatile and non-volatile media, removable and non-removable media.

[0063] The system memory 704 can include computer system readable media in the form of volatile memory, such as random access memory (RAM) 708 and / or cache memory 710. Computer system / server 700 can further include other removable / non-removable, volatile / non-volatile computer system storage media. By way of example only, a storage system 712 can be provided for reading from and writing to non-removable, non-volatile magnetic media (not shown and typically called a "hard drive"). Although not specifically shown, a magnetic disk drive can also be used for reading and writing to a removable, non-volatile magnetic disk (e.g., a "floppy disk"), and an optical disk drive can be used for reading from or writing to a removable, non-volatile optical disk (such as a CD-ROM, DVD-ROM or other optical media). Each of these devices can be connected to bus system 706 by one or more data media interfaces. As will be further depicted and described below, memory 704 can include at least one program product having a set (e.g., at least one) of program modules that are configured to carry out the functions of embodiments of the application.

[0064] Program / utility 716 having a set (at least one) of program modules 716 can be stored in memory 704 by way of example, and not limitation, as well as an operating system, one or more application programs, other program modules, and program data. Each of the operating system, one or more application programs, other program modules, and program data or some combination thereof, can include an implementation of a networking environment. Program modules 716 generally carry out the functions and / or methodologies of embodiments of the application as described herein.

[0065] Computer system / server 700 can also communicate with one or more external devices 718 such as a keyboard, a pointing device, a display 720, etc.; one or more devices that enable a user to interact with computer system / server 700; and / or any devices (e.g., network card, modem, etc.) that enable computer system / server 700 to communicate with one or more other computing devices. Such communication can occur via Input / Output (I / O) interface(s) 714. Still yet, computer system / server 700 can communicate with one or more networks such as a local area network (LAN), a general wide area network (WAN), and / or a public network (e.g., the Internet) via network adapter 722. As depicted, network adapter 722 can communicate with the other components of computer system / server 700 via bus system 706. It should be understood that although not specifically shown, other hardware and / or software components could also be used, such as a microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data archival storage systems, etc.

[0066] Additionally, the data label verification system 600 for verifying labels of records of a dataset can be attached to the bus system 706.

[0067] The description of various embodiments of the present application has been presented for purposes of illustration, but is not intended to be exhaustive or limited to the embodiments disclosed. Many modifications and variations will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the described embodiments. The terminology used herein was chosen to best explain the principles of the embodiments, the practical application or technical improvement over technology found in the marketplace, or to enable others skilled in the art to understand the embodiments disclosed herein.

[0068] The present application can be implemented as a system, a method, and / or a computer program product. A computer program product can include a computer readable storage medium (or media) having computer readable program instructions thereon for causing a processor to carry out aspects of the present application.

[0069] The medium can be an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system (or apparatus) for propagating signals. Examples of a computer-readable medium include a semiconductor or solid state memory, magnetic tape, a removable computer diskette, a random access memory (RAM), a read-only memory (ROM), a rigid magnetic disc, and an optical disk. Current examples of optical disks include compact disk - read only memory (CD-ROM), compact disk - read / write (CD-R / W), DVD, and Blu-ray disk.

[0070] The computer readable storage medium can be a tangible device that can retain and store instructions for use by an instruction execution device. The computer readable storage medium can be, for example, but is not limited to, an electronic storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the foregoing. A non-exhaustive list of more specific examples of the computer readable storage medium includes the following: a portable computer diskette, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), a static random access memory (SRAM), a portable compact disc read-only memory (CD-ROM), a digital versatile disk (DVD), a memory stick, a floppy disk, a mechanically encoded device such as punch-cards or raised structures in a groove having instructions recorded thereon, and any suitable combination of the foregoing. A computer readable storage medium, as used herein, is not to be construed as being transitory signals per se, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through a waveguide or other transmission media (e.g., light pulses passing through a fiber-optic cable), or electrical signals transmitted through a wire.

[0071] Computer readable program instructions described herein can be downloaded to respective computing / processing devices from a computer readable storage medium or to an external computer or external storage device via a network, for example, the Internet, a local area network, a wide area network and / or a wireless network. The network can comprise copper transmission cables, optical transmission fibers, wireless transmission, routers, firewalls, switches, gateway computers and / or edge servers. A network adaptation card or network interface in each computing / processing device receives computer readable program instructions from the network and forwards the computer readable program instructions to storage media within the respective computing / processing device for execution.

[0072] Computer readable program instructions for carrying out operations of the present application can be assembly instructions, instruction-set-architecture (ISA) instructions, machine- related instructions, microcode, firmware instructions, state-setting data, or source code or object code written in any combination of one or more programming languages, including an object oriented programming language such as Smalltalk, C++ or the like, and conventional procedural programming languages, such as the "C" programming language or similar programming languages. The computer readable program instructions can execute entirely on the user's computer, partly on the user's computer, as a stand-alone software package, partly on the user's computer and partly on a remote computer or entirely on the remote computer or server. In the latter scenario, the remote computer can be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or the connection can be made to an external computer (for example, through the Internet using an Internet Service Provider). In some embodiments, electronic circuitry including, for example, programmable logic circuitry, field-programmable gate arrays (FPGA), or programmable logic arrays (PLA) can execute the computer readable program instructions by utilizing state information of the computer readable program instructions to personalize the electronic circuitry, in order to perform aspects of the present application.

[0073] Aspects of the present application are described herein with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems) and computer program products according to embodiments of the application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer readable program instructions.

[0074] These computer readable program instructions can be provided to a processor of a general purpose computer, special purpose computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions / acts specified in the flowchart and / or block diagram block or blocks. These computer readable program instructions can also be stored in a computer readable storage medium that can direct a computer, a programmable data processing apparatus, and / or other devices to function in a particular manner, such that the computer readable storage medium having instructions stored therein comprises an article of manufacture including

[0075] The computer readable program instructions can also be loaded onto a computer, other programmable data processing apparatus, or another device to cause a series of operational steps to be performed on the computer, other programmable apparatus or other device to produce a computer implemented process, such that the instructions which execute on the computer, other programmable data processing apparatus, or other device implement the functions / acts specified in the flowchart and / or block diagram block or blocks.

[0076] The flowchart and / or block diagram in the FIGURE(s) illustrate the architecture, functionality, and operation of possible implementations of systems, methods and computer program products according to various embodiments of the present application. In this regard, each block in the flowchart and / or block diagram can represent a module, segment, or portion of instructions, which comprises one or more executable instructions for implementing the specified logical functions (s). In some alternative implementations, the functions noted in the blocks can occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently, or the blocks can sometimes be executed in the reverse order, depending upon the functionality involved. It will also be noted that each block of the block diagrams and / or flowchart illustrations, and combinations thereof, can be implemented by special purpose hardware-based systems that perform the specified functions or acts, or combinations of special purpose hardware and computer instructions.

[0077] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the present application. As used herein, the singular forms "a", "an" and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will be further understood that the terms "comprises" and / or "comprising," when used in this specification, specify the presence of stated features, integers, steps, operations, elements, and / or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0078] All means or steps plus function elements in the following claims are meant to encompass the corresponding structural, material, acts or acts with equivalent functions. The description of the application has been presented for purposes of illustration and description, but is not exhaustive or limited to the disclosure provided. Many modifications and variations will be apparent to those of ordinary skill in the art, without departing from the scope and spirit of the application. The embodiments are chosen and described in order to best explain the principles of the application and its practical application, and to enable others skilled in the art to understand various embodiments with various modifications as are suited to the particular use contemplated.

Claims

1. A computer-implemented method for validating a label of a record of a dataset, wherein the record includes sample data and a relevant label of a plurality of labels, the method comprising: partitioning, by one or more processors, the dataset into a training dataset and an inference dataset, the training dataset including records relevant to a selected label, the inference dataset including records having sample data relevant to the selected label and all other labels of the plurality of labels, wherein the dataset is an image dataset; partitioning, by the one or more processors, the training dataset into a plurality of learner training datasets, each learner training dataset training a corresponding label-specific few-shot learner, wherein a learner training dataset includes at least one sample relevant to a selected label; training, by the one or more processors, a plurality of label-specific few-shot learners, wherein the few-shot learners are trained using respective ones of the plurality of learner training datasets; performing, by the one or more processors, inference on the inference dataset by the plurality of trained label-specific few-shot learners; generating, by the one or more processors, a plurality of sets of predicted label output values based on the inference, wherein respective ones of the sets of predicted label output values correspond to respective ones of the trained label-specific few-shot learners; aggregating, by the one or more processors, predicted label output values within each of the sets of predicted label output values to generate respective score values for each of the plurality of label-specific few-shot learners, the respective score values reflecting variability in predicting the selected label; and in response to determining that at least a predefined number of the score values are below a predefined variability threshold, validating, by the one or more processors, the selected label for the dataset.

2. The method of claim 1, wherein at least one of the few-shot learners is a one-shot learner.

3. The method of one of claims 1-2, wherein at least one of the few-shot learners uses a Bayesian few-shot learning algorithm.

4. The method of one of claims 1-2, further comprising: training, by the one or more processors, one of the plurality of label-specific few-shot learners using records of the dataset relevant to a selected label.

5. The method of one of claims 1-2, wherein each of the plurality of label-specific few-shot learners has been pre-trained with a data sample relevant to a selected label.

6. The method of one of claims 1-2, further comprising: in response to determining that the plurality of learner training datasets does not result in a confidence level above a confidence threshold level, dividing, by the one or more processors, the dataset into a second training dataset and a second inference dataset, the second training dataset comprising records related to a second selected label, the second inference dataset comprising records having sample data related to the second selected label and all other labels in the plurality of labels.

7. The method of one of claims 1-2, further comprising: increasing, by the one or more processors, the number of the plurality of learner training datasets from one repetition cycle to the next repetition cycle until improvement in the measured outcome is below a predefined threshold.

8. The method of one of claims 1-2, further comprising: reducing, by the one or more processors, the number of the training datasets by removing training datasets having a low confidence level and retaining learner training datasets in the plurality of learner training datasets that result in a high confidence level of a related label for the next repetition cycle.

9. The method of one of claims 1-2, wherein each of the plurality of label-specific few-shot learners is adapted to classify into two classes comprising a fitting label class and a non-fitting label class.

10. The method of one of claims 1-2, wherein dividing the dataset into a training dataset and an inference dataset comprising records further comprises: ordering, by the one or more processors, the sample data according to known confidence levels of the respective related labels; and selecting, by the one or more processors, a portion of the sample data of the training dataset such that the portion of sample data has a higher confidence level of the related label than a remaining portion of the dataset.

11. The method of one of claims 1-2, further comprising: adding, by the one or more processors, respective records to the training dataset upon determining records in the dataset having a related confidence level.

12. A computer program product for verifying labels of records of a dataset, wherein the records comprise sample data and a related label of a plurality of labels, the computer program product comprising: one or more computer-readable storage media and program instructions stored on the one or more computer-readable storage media, the program instructions comprising: program instructions to divide the dataset into a training dataset and an inference dataset, the training dataset comprising records related to a selected label, the inference dataset comprising records having sample data related to the selected label and all other labels in the plurality of labels, wherein the dataset is an image dataset; program instructions to divide the training dataset into a plurality of learner training datasets, each learner training dataset training a corresponding label-specific few-shot learner, wherein a learner training dataset comprises at least one sample related to the selected label; ​ program instructions for training a plurality of label-specific few-shot learners, wherein the few-shot learners are trained using respective ones of the plurality of learner training data sets; program instructions for performing inference on the inference data set by the plurality of trained label-specific few-shot learners to generate a plurality of sets of predicted label output values; program instructions for aggregating the predicted label output values within each of the sets of predicted label output values to generate respective score values for each of the plurality of label-specific few-shot learners, the respective score values reflecting variability in the predicted selected label; and program instructions for validating the selected label for the data set in response to determining that at least a predefined number of the score values are below a predefined variability threshold.

13. The computer program product of claim 12, wherein at least one of the few-shot learners is a one-shot learner.

14. The computer program product of one of claim 12, wherein at least one of the few-shot learners uses a Bayesian few-shot learning algorithm.

15. The computer program product of one of claims 12-13, further comprising program instructions stored on the one or more computer-readable storage media for: training one of the plurality of label-specific few-shot learners using records of the data set that are related to the selected label.

16. A computer system for validating labels of records of a data set, wherein the records include sample data and a related one of a plurality of labels, the computer system comprising: one or more computer processors; one or more computer-readable storage media; and program instructions stored on the computer-readable storage media for execution by at least one of the one or more processors, the program instructions comprising: program instructions for dividing the data set into a training data set and an inference data set, the training data set including records related to a selected label, the inference data set including records having sample data related to the selected label and all other ones of the plurality of labels, wherein the data set is an image data set; program instructions for dividing the training data set into a plurality of learner training data sets, each learner training data set training a corresponding label-specific few-shot learner, wherein a learner training data set includes at least one sample related to the selected label; program instructions for training a plurality of label-specific few-shot learners, wherein the few-shot learners are trained using respective ones of the plurality of learner training data sets; program instructions for performing inference on the inference data set by the plurality of trained label-specific few-shot learners to generate a plurality of sets of predicted label output values; program instructions for aggregating the predicted label output values within each of the sets of predicted label output values to generate respective score values for each of the plurality of label-specific few-shot learners, the respective score values reflecting variability in the predicted selected label; and program instructions for validating the selected label for the data set in response to determining that at least a predefined number of the score values are below a predefined variability threshold. program instructions for generating a plurality of sets of predicted label output values based on the inferring, wherein respective ones of the sets of predicted label output values correspond to respective ones of the trained label-specific few-shot learners; program instructions for aggregating the predicted label output values within each of the sets of predicted label output values to generate, for each of the plurality of label-specific few-shot learners, a respective score value that reflects variability in the predicted selected label; and program instructions for validating the selected label for the data set in response to determining that at least a predefined number of the score values are below a predefined variability threshold.

17. The computer system of claim 16, wherein at least one of the few-shot learners is a one-shot learner.

Images