JP7469341B2 - Method and system for classifying data using parameter size invariant classifiers for unbiased classification - Patents.com - Google Patents

Description

The present invention relates to a method and system for classifying data using a parameter size invariant classifier for unbiased classification.

Standard classifiers, which are widely used in the fields of conventional artificial intelligence and machine learning, have a problem of having biased classification performance because the parameter size of the classifier is affected by the bias inherent in the training data set. For example, when classifying input data into one of multiple categories, bias occurs due to differences in the amount of each category in the training data.

To solve these problems, there is a conventional technique that forms groups for each category containing roughly the same amount of training data, and processes the learning in each group. However, this conventional technique has the problem that it is specialized for only a specific task (detection task) and cannot be applied to classification problems of other tasks (instance segmentation task or classification task). Furthermore, this conventional technique requires the search for many hyperparameters, so it requires a lot of resources and time to actually apply it.

Korean Patent No. 10-2180054

We provide a data classification method and system that can resolve the problem of classifier bias by removing the bias with respect to the size of the classifier's parameter vector.

A data classification method for a computer device including at least one processor is provided, the data classification method including a step of generating an embedding vector of input data by the at least one processor, a step of calculating an inner product of the embedding vector and a parameter vector of a trained classifier by the at least one processor, and a step of calculating a logit with no bias for the size of the parameter vector by applying a norm for the parameter vector to the result of the inner product by the at least one processor.

According to one aspect, the step of calculating the logit may be characterized by dividing the result of the dot product by the norm for the parameter vector to remove bias with respect to the size of the parameter vector.

In another aspect, the step of calculating the logit may be characterized by including a step of biasing the result of the norm-applied dot product by applying a hyperbolic tangent.

A computer program is provided for causing a computer device to execute the method.

A computer-readable recording medium is provided on which a program for causing a computer device to execute the method is recorded.

A computer device is provided, comprising at least one processor implemented to execute computer-readable instructions, the at least one processor generating an embedding vector for input data, the at least one processor calculating an inner product of the embedding vector and a parameter vector of a trained classifier, and the at least one processor applying a norm for the parameter vector to the result of the inner product to calculate a logit that is debiased with respect to the size of the parameter vector.

The problem of classifier bias can be eliminated by removing the bias with respect to the size of the classifier's parameter vector.

FIG. 2 is a block diagram illustrating an example computing device according to an embodiment of the present invention. 1 is a flow chart illustrating an example method for classifying data in accordance with an embodiment of the present invention. 1 is a chart comparing the performance of using a standard logit representation and a norm-invariant logit representation according to an embodiment of the present invention; 1 is a chart comparing the performance of using a standard logit representation and a norm-invariant logit representation according to an embodiment of the present invention;

The following describes the embodiments in detail with reference to the attached drawings.

The data classification system according to the embodiment of the present invention may be realized by at least one computer device, and the data classification method according to the embodiment of the present invention may be executed by at least one computer device included in the data classification system. A computer program according to an embodiment of the present invention may be installed and implemented in the computer device, and the computer device may execute the data classification method according to the embodiment of the present invention according to the control of the executed computer program. The above-mentioned computer program may be recorded on a computer-readable recording medium in order to be combined with the computer device and cause the computer to execute the data classification method.

1 is a block diagram showing an example of a computer device in an embodiment of the present invention. Such a computer device 100 may include a memory 110, a processor 120, a communication interface 130, and an input/output interface 140, as shown in FIG. 1. The memory 110 is a computer-readable recording medium and may include a RAM (random access memory), a ROM (read only memory), and a persistent mass storage device such as a disk drive. Here, a persistent mass storage device such as a ROM or a disk drive may be included in the computer device 100 as a separate persistent storage device separate from the memory 110. In addition, the memory 110 may record an operating system and at least one program code. Such software components may be loaded into the memory 110 from a computer-readable recording medium separate from the memory 110. Such a separate computer-readable recording medium may include a computer-readable recording medium such as a floppy drive, a disk, a tape, a DVD/CD-ROM drive, or a memory card. In other embodiments, the software components may be loaded into the memory 110 through a communication interface 130 that is not a computer-readable recording medium. For example, the software components may be loaded into the memory 110 of the computer device 100 based on a computer program that is installed by a file received via the network 160.

The processor 120 may be configured to process computer program instructions by performing basic arithmetic, logic, and input/output operations. The instructions may be provided to the processor 120 by the memory 110 or by the communication interface 130. For example, the processor 120 may be configured to execute received instructions according to program code recorded in a storage device such as the memory 110.

The communication interface 130 may provide a function for the computer device 100 to communicate with other electronic devices (for example, the storage device described above) via the network 160. As an example, requests, commands, data, files, etc. generated by the processor 120 of the computer device 100 according to program code recorded in a storage device such as the memory 110 may be transmitted to other devices via the network 160 under the control of the communication interface 130. Conversely, signals, commands, data, files, etc. from other devices may be received by the computer device 100 through the communication interface 130 of the computer device 100 via the network 160. The signals, commands, data, etc. received through the communication interface 130 may be transmitted to the processor 120 or the memory 110, and the files, etc. may be recorded on a recording medium (the persistent storage device described above) that the computer device 100 may further include.

The input/output interface 140 may be a means for interfacing with the input/output device 150. For example, the input device may include devices such as a microphone, keyboard, or mouse, and the output device may include devices such as a display, speaker, etc. As another example, the input/output interface 140 may be a means for interfacing with a device in which functions for input and output are integrated into one, such as a touch screen. The input/output device 150 may be configured as a single device together with the computer device 100.

Also, in other embodiments, the computing device 100 may include fewer or more components than those of FIG. 1. However, most of the prior art components need not be explicitly shown in the figures. For example, the computing device 100 may be implemented to include at least some of the input/output devices 150 described above, and may further include other components such as a transceiver, a database, etc.

Figure 2 is a flow chart showing an example of a data classification method according to an embodiment of the present invention. The data classification method according to this embodiment may be executed by a computer device 100. In this case, the processor 120 of the computer device 100 may be implemented to execute control instructions according to the operating system code and at least one computer program code contained in the memory 110. Here, the processor 120 may control the computer device 100 so that the computer device 100 executes steps 210 to 230 included in the method of Figure 2 according to the control instructions provided by the code recorded in the computer device 100.

At step 210, the computer device 100 may generate an embedding vector for the input data. As an example, if the input data is an image, the computer device 100 may generate a feature vector of the image as the embedding vector.

In step 220, the computer device 100 may calculate an inner product of the embedding vector and the parameter vector of the trained classifier. Training the classifier using the training data may correspond to training the parameter vector of the classifier. In this case, the computer device 100 may calculate an inner product of the parameter vector (or classifier weight vector) of the classifier previously trained using the training data of each category and the embedding vector generated in step 210.

In step 230, the computer device 100 may apply a norm for the parameter vector to the result of the dot product to calculate a logit (logistic probit) that removes bias with respect to the size of the parameter vector. For example, the computer device 100 may divide the result of the dot product by the norm for the parameter vector to remove bias with respect to the size of the parameter vector. Since the norm refers to the size of a vector, by dividing the result of the dot product of the embedding vector and the parameter vector by the norm, only the directionality remains in the parameter vector, and the size of the parameter vector is removed. Therefore, bias that appears due to differences in the amounts of each category in the training data can be removed.

In some embodiments, the computer device 100 may bias the result of the normed dot product by applying a hyperbolic tangent (tanh). The bias may correspond to the bias of the classifier.

The mathematical meaning of the data classification method in an embodiment of the present invention is explained below.

In general, the standard logit representation of a classifier may be expressed as follows:

Here, l _i may mean the logit representation of the i-th category for the classifier. Also, f may mean a parameter vector and an embedding vector having C dimensions, respectively. In other words, the standard logit representation of the classifier may be mathematically generated based on the inner product of the parameter vector of the classifier and the embedding vector generated from the input data. At this time, bias caused by the difference in the amount of each category of the training data may be inherent in the parameter vector of the classifier, and may be reflected in the standard logit representation generated using such a parameter vector. Reflection of such bias may lead to bias in the classification performance of the classifier. Also, b _i may mean the bias of the classifier.

As described above, in an embodiment of the present invention, the problem of such bias can be solved by using the norm of the parameter vector of the classifier, as shown in the following equation (2), for example.

here,

may mean the norm of the parameter vector of the classifier. In other words, by dividing the inner product of the parameter vector of the classifier and the embedding vector of the input data by the norm indicating the size of the vector (the norm of the parameter vector of the classifier), it is possible to remove the bias due to size from the result of the inner product. In addition, by applying a hyperbolic tangent (tanh) to the bias of the classifier and limiting the value of the applied bias to real numbers between -1 and 1, it is possible to limit the influence of the bias on the classifier. In this specification, the logit representation according to Equation (2) is referred to as a norm-invariant logit representation.

Figures 3 and 4 are charts comparing the performance of using a standard logit representation and a norm-invariant logit representation according to one embodiment of the present invention.

3 shows the performance of an instance segmentation task, which is a task of searching for the location of an object in an input image and classifying the object category, when using a standard logit representation and when using a norm-invariant logit representation according to an embodiment of the present invention. In this case, in the instance segmentation task, the location may be represented by a box and a pixel-level mask. The performance test was performed using the LVISv0.5benchmark dataset. In the table of FIG. 3, AP indicates the average precision, which is a measurement item widely used to measure the detection accuracy. In this case, AP_S, AP_M, and AP_L are APs related to the object size (i.e., Small, Medium, and Large), and AP_r, AP_c, and AP_f are APs related to the sample frequency (i.e., rarecommon, frequent) of the training dataset. In this case, the graph in Figure 3 shows that the norm-invariant logit representation significantly improved performance in the low-shot cases (AP_r and AP_c, for example) for both boxes and masks, effectively resolving the imbalance of each category of training data and the resulting bias problem. In addition, in the case of AP_f, the performance when using the norm-invariant logit representation was similar to that when using the standard logit representation.

Figure 4 shows the performance of a classification task for classifying the category of an input image when using a standard logit representation and when using a norm-invariant logit representation according to an embodiment of the present invention. The performance test was performed using the Long-tailed CIFAR-10 benchmark dataset. In the diagram of Figure 4, the higher the imbalance ratio, the higher the imbalance of the training dataset. In this case, the diagram of Figure 4 shows that the norm-invariant logit representation significantly improved performance in the case of low-shot (for example, an imbalance ratio of 100), effectively solving the problem of imbalance and the resulting bias in each category of training data. Even when the imbalance ratio was low (for example, an imbalance ratio of 50 and 10), the performance of the norm-invariant logit representation was similar to that of the standard logit representation.

Thus, according to an embodiment of the present invention, the problem of classifier bias can be resolved by removing the bias with respect to the size of the classifier's parameter vector.

The above-described systems or devices may be realized by hardware components or a combination of hardware and software components. For example, the devices and components described in the embodiments may be realized using one or more general-purpose or special-purpose computers, such as, for example, a processor, a controller, an arithmetic logic unit (ALU), a digital signal processor, a microcomputer, a field programmable gate array (FPGA), a programmable logic unit (PLU), a microprocessor, or various devices capable of executing and responding to instructions. The processing device may execute an operating system (OS) and one or more software applications that run on the OS. The processing device may also respond to the execution of the software and access, record, manipulate, process, and generate data. For ease of understanding, the description may be given as one processing device being used, but those skilled in the art will understand that the processing device may include multiple processing elements and/or multiple types of processing elements. For example, a processing unit may include multiple processors, or one processor and one controller. Other processing configurations, such as parallel processors, are also possible.

The software may include computer programs, codes, instructions, or a combination of one or more of these, and may configure or instruct the processing device to operate as desired, either independently or collectively. The software and/or data may be embodied in any type of machine, component, physical device, virtual device, computer storage medium, or device to be interpreted based on the processing device or to provide instructions or data to the processing device. The software may be distributed and stored or executed in a distributed manner on computer systems connected by a network. The software and data may be stored on one or more computer-readable storage media.

The method according to the embodiment may be realized in the form of program instructions executable by various computer means and recorded on a computer-readable medium. The computer-readable medium may include program instructions, data files, data structures, etc., alone or in combination. The medium may be one that continuously records a computer-executable program, or one that temporarily records it for execution or download. The medium may be one that is a recording or storage means in the form of a single or multiple hardware devices combined, and is not limited to a medium that is directly connected to a certain computer system, but may be one that is distributed over a network. Examples of the medium include magnetic media such as hard disks, floppy disks, and magnetic tapes, optical media such as CD-ROMs and DVDs, magneto-optical media such as floptical disks, and ROMs, RAMs, flash memories, etc., configured to record program instructions. Other examples of the medium may include recording media or storage media managed by application stores that distribute applications, or sites, servers, etc. that supply or distribute various other software. Examples of program instructions include not only machine code, such as that produced by a compiler, but also high-level language code that is executed by a computer using an interpreter or the like.

Although the embodiments have been described above based on limited embodiments and drawings, those skilled in the art will appreciate that various modifications and variations can be made from the above description. For example, the described techniques may be performed in a different order than described, and/or the components of the described systems, structures, devices, circuits, etc. may be combined or combined in a different manner than described, or may be counterbalanced or replaced by other components or equivalents, and still achieve suitable results.

Therefore, different embodiments that are equivalent to the scope of the claims are within the scope of the attached claims.

100: Computer device 110: Memory 120: Processor 130: Communication interface 140: Input/output interface 150: Input/output device 160: Network

Claims (4)

1. A method of data classification in a computer device, the method comprising:
the at least one processor generating an embedding vector for the input data;
said at least one processor calculating an inner product of said embedding vector and a parameter vector of a trained classifier; and said at least one processor calculating a logit by applying a norm for said parameter vector to the result of said inner product to remove bias with respect to the size of said parameter vector and leaving only the directionality of said parameter vector .
Including,
The step of calculating the logit comprises:
Dividing the result of the dot product by the norm for the parameter vector, and then:
Add a bias with a hyperbolic tangent applied to limit the value between -1 and 1;
1. A data classification method comprising : A computer program comprising a plurality of computer executable instructions,
When the instructions are executed by a processor in a computing device,
2. A method for causing the computer device to perform the method of claim 1 .
Computer program. A computer-readable recording medium having a computer program recorded thereon,
When the computer program is executed by a computer device,
2. A method for causing the computer device to perform the method of claim 1 .
A computer-readable recording medium. 1. A computing device comprising: at least one processor implemented to execute computer readable instructions,
The at least one processor
Generate an embedding vector for the input data;
Calculate the dot product of the embedding vector and a parameter vector of a trained classifier; and
A norm for the parameter vector is applied to the result of the inner product to remove bias with respect to the size of the parameter vector, and a logit is calculated that leaves only the direction of the parameter vector;
To calculate the logits, the at least one processor:
Dividing the result of the dot product by the norm for the parameter vector, and then:
Add a bias with a hyperbolic tangent applied to limit the value between -1 and 1;
A computer device comprising: