KR20230115143A - Apparatus for generating a learning model using log scaling loss function and method therefor - Google Patents

Abstract

A method for generating a learning model of the present invention comprises: a step of preparing, by a data processing part, learning data of input data for learning and label corresponding to the input data; a step of inputting, by the data processing part, the input data for learning into a learning model wherein learning has not been completed; a step of calculating, by the learning model, output data for learning through a plurality of operations wherein a weight for which learning between a plurality of strata has not been completed are applied for the input data for learning; a step of calculating, by a loss deriving part, a loss representing a difference of the output data for learning and the label using a log mean square error function wherein a mean square error function is converted to a log scale; and a step of performing, by an optimization part, optimization to modify the weight of the learning model so that the loss is minimized.

Description

Apparatus for generating a learning model using log scaling loss function and method therefor}

The present invention relates to a loss function, and more particularly, to an apparatus and method for generating a learning model using a logarithmic scaling loss function.

In the case of machine learning algorithms, detailed features of the data must be defined in order to train the model. If the machine learning system operator has advanced knowledge or experience on the problem, the feature information for machine learning model training will be more elaborately defined. However, there is a limitation that it takes a lot of time, effort, and money to reach this level. In addition, this increases the system's dependence on machine learning operators.

On the other hand, when using a deep learning technique, a high-performance computing environment such as a graphic processing unit is required for training. Instead, deep learning techniques can replace delicate tasks based on domain knowledge such as decision rules or feature definitions in the aforementioned existing methods. In these deep learning techniques, the operator only needs to define the model structure and loss function to train the neural network better. Various effective loss functions and neural network architectures have already been proposed.

In other words, deep learning-based algorithms are widely adopted due to the advantage of being able to create an artificial neural network model with no or minimal domain knowledge about the task. Instead, it is desirable to define an appropriate neural network structure or loss function in order to train artificial neural networks more stably.

Korean Patent Publication No. 2020-0116225 (published on October 12, 2020)

An object of the present invention is to provide an apparatus and method for generating a learning model using a logarithmic scaling loss function.

In the method for generating a learning model of the present invention, when the learning model calculates output data for learning with respect to the input data for learning, the loss derivation unit converts the mean square error function to a logarithmic scale using a log mean square error function for the learning output. Calculating a loss representing the difference between the data and the label, and performing optimization by an optimization unit to modify the weight of the learning model so that the loss is minimized.

Calculating the loss is the log mean square error function Calculate the loss using , wherein y is the label, and is the output data for learning.

The log mean square error function converts the structure of the mean square error function into a log function in negative form, applies the 1-X form, and to X It is characterized in that it is generated by substituting.

A method for generating a learning model of the present invention includes the steps of preparing learning data including input data for learning and a label corresponding to the input data by a data processing unit, The step of inputting data to a model, the learning model calculating output data for learning through a plurality of operations to which weights for which learning between a plurality of layers is not completed are applied to the input data for learning, and a loss derivation unit calculating the mean square error Calculating a loss representing a difference between the training output data and the label by using a log mean square error function obtained by converting a function to a log scale; Optimization in which an optimizer modifies the weight of the learning model so that the loss is minimized. It includes the steps of performing

Calculating the loss is a log mean square error function Calculate the loss using , where y is a label, It is characterized in that is output data for learning.

When the learning model is a generative network, the label is input data for learning, and when the learning model is a classified network, the label is a target value representing a classification to which the input data for learning belongs.

The log mean square error function converts the structure of the mean square error function into a log function in negative form, applies the 1-X form, and to X It is characterized in that it is generated by substituting.

An apparatus for generating a learning model of the present invention, when the learning model calculates output data for learning with respect to input data for learning, uses a log mean square error function converted from a mean square error function to a logarithmic scale to obtain the output data for learning and the output data for learning. It includes a loss derivation unit that calculates a loss representing the difference between labels, and an optimization unit that performs optimization to modify the weights of the learning model so that the loss is minimized.

The loss derivation unit uses the log mean square error function Calculate the loss using , wherein y is the label, and is the output data for learning.

The log mean square error function converts the structure of the mean square error function into a log function in negative form, applies the 1-X form, and to X It is characterized in that it is generated by substituting.

An apparatus for generating a learning model of the present invention includes a data processing unit that prepares learning input data for learning and a label corresponding to the input data, and inputs the learning input data to a learning model in which learning has not been completed, and the learning model. When output data for learning is calculated through a plurality of operations to which weights for which learning between a plurality of layers are not completed are applied to the input data for learning, a log mean square error function converted from a mean square error function to a log scale is used. It includes a loss derivation unit that calculates a loss representing the difference between the training output data and the label, and an optimization unit that performs optimization to modify the weight of the learning model so that the loss is minimized.

The loss derivation unit uses the log mean square error function Calculate the loss using , where y is a label, It is characterized in that is output data for learning.

When the learning model is a generative network, the label is input data for learning, and when the learning model is a classified network, the label is a target value representing a classification to which the input data for learning belongs.

The log mean square error function converts the structure of the mean square error function into a log function in negative form, applies the 1-X form, and to X It is characterized in that it is generated by substituting.

Optimization by the gradient descent method can be performed stably by using the log mean square error function, which is the loss function of the present invention.

1 is a diagram for explaining the configuration of an apparatus for generating a learning model using a logarithmic scaling loss function according to an embodiment of the present invention.
2 is a flowchart illustrating a method for generating a learning model according to an embodiment of the present invention.
3 is a graph for explaining the advantages of the Log Mean Square Error (LMSE) function according to an embodiment of the present invention.
4 is an exemplary diagram of a hardware system for implementing an apparatus for generating a learning model using a log scaling loss function according to an embodiment of the present invention.

In order to clarify the characteristics and advantages of the problem solving means of the present invention, the present invention will be described in more detail with reference to specific embodiments of the present invention shown in the accompanying drawings.

However, detailed descriptions of well-known functions or configurations that may obscure the gist of the present invention will be omitted in the following description and accompanying drawings. In addition, it should be noted that the same components are indicated by the same reference numerals throughout the drawings as much as possible.

The terms or words used in the following description and drawings should not be construed as being limited to a common or dictionary meaning, and the inventor may appropriately define the concept of terms for explaining his/her invention in the best way. It should be interpreted as a meaning and concept consistent with the technical spirit of the present invention based on the principle that there is. Therefore, the embodiments described in this specification and the configurations shown in the drawings are only one of the most preferred embodiments of the present invention, and do not represent all of the technical ideas of the present invention. It should be understood that there may be equivalents and variations.

In addition, terms including ordinal numbers, such as first and second, are used to describe various components, and are used only for the purpose of distinguishing one component from other components, and to limit the components. Not used. For example, a second element may be termed a first element, and similarly, a first element may be termed a second element, without departing from the scope of the present invention.

Additionally, when an element is referred to as being “connected” or “connected” to another element, it means that it is logically or physically connected or capable of being connected. In other words, it should be understood that a component may be directly connected or connected to another component, but another component may exist in the middle, or may be indirectly connected or connected.

In addition, terms used in this specification are only used to describe specific embodiments, and are not intended to limit the present invention. Singular expressions include plural expressions unless the context clearly dictates otherwise. In addition, terms such as "include" or "having" described in this specification are intended to designate that there is a feature, number, step, operation, component, part, or combination thereof described in the specification, but one or the It should be understood that the above does not preclude the possibility of the presence or addition of other features, numbers, steps, operations, components, parts, or combinations thereof.

In addition, terms such as “… unit”, “… unit”, and “module” described in the specification mean a unit that processes at least one function or operation, which may be implemented as hardware or software or a combination of hardware and software. there is.

Also, "a or an", "one", "the" and similar words in the context of describing the invention (particularly in the context of the claims below) indicate otherwise in this specification. may be used in the sense of including both the singular and the plural, unless otherwise clearly contradicted by the context.

In addition, embodiments within the scope of the present invention include computer-readable media having or conveying computer-executable instructions or data structures stored thereon. Such computer readable media can be any available media that can be accessed by a general purpose or special purpose computer system. By way of example, such computer readable media may be in the form of RAM, ROM, EPROM, CD-ROM or other optical disk storage, magnetic disk storage or other magnetic storage, or computer executable instructions, computer readable instructions or data structures. physical storage media such as, but not limited to, any other medium that can be used to store or convey any program code means in a computer system and which can be accessed by a general purpose or special purpose computer system. .

In the following description and claims, a "network" is defined as one or more data links that enable the transfer of electronic data between computer systems and/or modules. When information is transmitted or provided to a computer system over a network or other (wired, wireless, or combination of wired or wireless) communication connection, the connection may be understood as a computer-readable medium. Computer readable instructions include, for example, instructions and data that cause a general purpose or special purpose computer system to perform a particular function or group of functions. Computer executable instructions may be, for example, binary, intermediate format instructions, such as assembly language, or even source code.

In addition, the present invention relates to personal computers, laptop computers, handheld devices, multiprocessor systems, microprocessor-based or programmable consumer electronics, network PCs, minicomputers, mainframe computers, mobile phones, PDAs, pagers It can be applied in a network computing environment having various types of computer system configurations including (pager) and the like. The invention may also be practiced in distributed system environments where tasks are performed by both local and remote computer systems linked by wired data links, wireless data links, or a combination of wired and wireless data links through a network. In a distributed system environment, program modules may be located in local and remote memory storage devices.

First, an apparatus for generating a learning model using a logarithmic scaling loss function according to an embodiment of the present invention will be described. 1 is a diagram for explaining the configuration of an apparatus for generating a learning model using a logarithmic scaling loss function according to an embodiment of the present invention.

Referring to FIG. 1, a device (10, hereinafter abbreviated as 'learning device') for generating a learning model according to an embodiment of the present invention is a learning model (LM: machine learning model or deep learning model) through learning data. ) to generate. The learning device 10 of the present invention includes a data processing unit 100, a loss calculation unit 200, and an optimization unit 300.

The learning model LM includes a plurality of layers, and each of the plurality of layers performs a plurality of operations. Each calculation result of a plurality of calculation modules of any one layer is transferred to the next layer after a weight is applied. This means that a weight is applied to the calculation result of the current layer and input to the calculation of the next layer. That is, the learning model LM performs a plurality of operations to which weights of a plurality of layers are applied.

The plurality of layers of the learning model (LM) include one or more of a fully-connected layer, a convolutional layer, a recurrent layer, a graph layer, and a pooling layer. contains a combination The plurality of operations may include a convolution operation, a down sampling operation, an up sampling operation, an operation using an activation function, and the like. Here, the activation function may include sigmoid, hyperbolic tangent (tanh), exponential linear unit (ELU), rectified linear unit (ReLU), leaky ReLU, Maxout, Minout, Softmax, and the like. .

The learning model (LM) may be a generative network or a classification network.

When the learning model LM is a generative network, when input data is input, the learning model LM performs a plurality of operations to which weights of a plurality of layers are applied to simulate the input data and outputs output data. That is, the output data may be simulated input data. Such a generative network may exemplify, for example, a Restricted Boltzmann Machine (RBM), an Auto-Encoder (AE), a Generative Adversarial Network (GAN), and the like.

When the learning model (LM) is a classification network, when input data is input, the learning model (LM) performs a plurality of operations to which weights of a plurality of layers are applied, and outputs probabilities for a plurality of classifications as output data. . That is, the output data is a probability that the input data belongs to each of a plurality of classifications. Such a classification type network may exemplify, for example, a Convolution Neural Network (CNN), a Recurrent Neural Network (RNN), and the like.

The data processing unit 100 is for preparing learning data. The learning data includes input data for learning input to the learning model LM and labels corresponding to the input data for learning. When the learning model (LM) is a generative network, labels may be input data for learning. Meanwhile, when the learning model (LM) is a classified network, the label may be a target value indicating a classification to which input data for learning belongs. When the learning data is prepared, the data processing unit 100 inputs input data among the learning data to the learning model LM. Then, the learning model LM will calculate output data for learning through a plurality of calculations to which a weight for which learning between a plurality of layers has not been completed is applied to the input data for learning.

The loss deriving unit 200 is for calculating a loss using a loss function. Here, the loss represents the difference between the output data for learning and the label. A loss function according to an embodiment of the present invention is a log scale based loss function.

More specifically, the loss function according to an embodiment of the present invention is generated by converting a mean squared error (MSE) function to a logarithmic scale, and is referred to as a 'logarithmic mean squared error (LMSE)' function. to name This log mean square error (LMSE) function is shown in Equation 1 below.

here, Is the learning output data of the learning model, represents a label corresponding to the learning input data of the learning model. Here, the label may be input data for learning when the learning model LM is a generative network. Meanwhile, when the learning model LM is a classified network, the label may be a target value representing a classification to which input data for learning belongs.

The log mean square error (LMSE) function of the present invention is converted to a logarithmic scale while maintaining the structure of the mean square error (MSE) function. and When the range of is normalized between 0 and 1, it has a logarithmic scale The range of is -1 to 0. In order to apply the Log Mean Square Error (LMSE) function, which is a log scale to the task for minimizing the loss, i.e. optimization, the range of the loss of the Log Mean Square Error (LMSE) function is given in the form of 0 to C (positive finite value). need to adjust

In order to utilize the logarithmic function while maintaining the structure of the mean square error (MSE) function, it is converted into a logarithmic function in negative form so that the expression range of the logarithmic function is inverted. Accordingly, the output range is changed from 1 to 0. Then, apply the 1-X format to flip the values (1 to 0) at the beginning and end of the output range. Then, for X of the form 1-X Substitute

However, since the loss function ends up with infinity, we add a floating point value ε to the 1-X form. This gives the range of loss of the Log Mean Square Error (LMSE) function from 0 to C (positive finite value).

The optimizer 300 performs optimization to modify the parameters of the learning model so that the loss derived through the log mean square error function is minimized. At this time, optimization is performed according to Equation 2 below.

here, represents the loss derived from the tth iteration, represents the slope of the loss, represents the learning rate.

Next, a method for learning the learning model (LM) according to an embodiment of the present invention will be described. 2 is a flowchart illustrating a method for generating a learning model according to an embodiment of the present invention.

Referring to FIG. 2 , the learning unit 100 prepares learning data for the learning model LM in step S110. Here, the learning data includes input data for learning and labels. Here, the label may be input data for learning when the learning model LM is a generative network. Meanwhile, when the learning model LM is a classified network, the label may be a target value representing a classification to which input data for learning belongs.

In step S120, the learning unit 100 inputs the input data for learning to the learning model LM where learning has not been completed. Then, the learning model (LM) calculates output data for learning through a plurality of calculations to which weights for which learning between a plurality of layers is not completed are applied in step S130.

Subsequently, the learning unit 100 calculates a loss representing a difference between the predicted value for learning and the label through a loss function, that is, a log mean square error (LMSE) function of Equation 1 in step S140.

Next, the learning unit 100 performs optimization of modifying the weight of the learning model LM so that the loss derived through the loss function is minimized in step S150. Here, the optimization algorithm may use a gradient descent method.

Steps S120 to S150 described above are repeatedly performed using a plurality of different learning data, and the weights of the learning model LM are repeatedly updated according to these repetitions. And this repetition is performed until the loss is less than the predetermined target value. Therefore, the learning unit 100 determines whether or not the loss calculated earlier (S140) is less than or equal to a preset target value in step S160, and if the loss is less than or equal to the preset target value, completes learning for the learning model (LM) in step S170. .

Next, a log mean square error (LMSE) function and a mean square error (MSE) function according to an embodiment of the present invention will be compared and described. 3 is a graph for explaining the advantages of the Log Mean Square Error (LMSE) function according to an embodiment of the present invention.

In order to apply the gradient descent method for optimization, the gradient (slope) of the loss must be obtained, and for this purpose, the differentiation of the loss function is required.

First, the derivative of the gradient mean square error (MSE) function representing the slope of the loss of the mean square error (MSE) function is shown in Equation 3 below.

3(A) visualizes the slope of the loss of the mean square error (MSE) function.

In addition, the derivative of the log mean square error (LMSE) function representing the slope of the loss of the log mean square error (LMSE) function is as shown in Equation 4 below.

Figure 3 (B) is a visualization of the slope of the loss of the log mean square error (LMSE) function.

Comparing (A) and (B) of FIG. 3, in the case of the log mean square error (LMSE) function, the learning output data ( ), it shows a steeper slope when the spacing between labels (y) increases.

These characteristics indicate that the log mean square error (LMSE) function can perform stable optimization compared to the mean square error (MSE) function.

4 is an exemplary diagram of a hardware system for implementing an apparatus for generating a learning model using a log scaling loss function according to an embodiment of the present invention.

As shown in FIG. 4 , a hardware system 2000 according to an embodiment of the present invention has a configuration including a processor unit 2100, a memory interface unit 2200, and a peripheral device interface unit 2300. can

Each component in the hardware system 2000 may be an individual part or may be integrated into one or more integrated circuits, and each component may be coupled by a bus system (not shown).

where, in the case of a bus system, any one or more individual physical buses, communication lines/interfaces, and/or multi-drop or point-to-point communication lines connected by suitable bridges, adapters, and/or controllers; It is an abstraction representing point-to-point connections.

The processor unit 2100 performs a role of executing various software modules stored in the memory unit 2210 by communicating with the memory unit 2210 through the memory interface unit 2200 to perform various functions in the hardware system. .

Here, in the memory unit 2210, each configuration including the learning model (LM), the data processing unit 100, the loss derivation unit 200, and the optimization unit 300 described above with reference to FIG. 2 is stored in the form of a software module. and other operating systems (OS) may be additionally stored. A configuration including the learning model (LM), the data processing unit 100, the loss derivation unit 200, and the optimization unit 300 may be loaded into the processor unit 2100 and executed.

Each component including the above-described learning model (LM), data processing unit 100, loss deduction unit 200, and optimization unit 300 is implemented in the form of a software module or hardware module executed by a processor, or a software module and It may also be implemented in the form of a combination of hardware modules.

As such, a software module executed by a processor, a hardware module, or a combination of software modules and hardware modules may be implemented as an actual hardware system (eg, a computer system).

For operating systems (e.g. embedded operating systems such as I-OS, Android, Darwin, RTXC, LINUX, UNIX, OS X, WINDOWS, or VxWorks), general system tasks (e.g. memory management, storage control) , power management, etc.) and includes various procedures, command sets, software components and/or drivers that control and manage, and plays a role in facilitating communication between various hardware modules and software modules.

For reference, the memory unit 2210 may include a memory hierarchy including, but not limited to, a cache, a main memory, and a secondary memory. In the case of such a memory hierarchy, for example, RAM (eg, SRAM, DRAM, DDRAM), ROM, FLASH, magnetic and/or optical storage devices such as disk drives, magnetic tapes, compact disks (CDs), and digital video discs (DVDs).

The peripheral device interface unit 2300 serves to enable communication between the processor unit 2100 and peripheral devices.

Here, in the case of a peripheral device, it is for providing different functions to the hardware system 2000, and in one embodiment of the present invention, for example, a communication unit 2310 may be included.

Here, the communication unit 2310 serves to provide a communication function with other devices, and for this purpose, for example, an antenna system, an RF transceiver, one or more amplifiers, a tuner, one or more oscillators, a digital signal processor, a codec ( CODEC) chipset, memory, etc., but are not limited thereto, and may include a known circuit that performs this function.

Such communication protocols supported by the communication unit 2310 include, for example, Wireless LAN (WLAN), Digital Living Network Alliance (DLNA), Wireless Broadband (Wibro), and World Interoperability for Microwave Access (Wimax). ), GSM (Global System for Mobile communication), CDMA (Code Division Multi Access), CDMA2000 (Code Division Multi Access 2000), EV-DO (Enhanced Voice-Data Optimized or Enhanced Voice-Data Only), WCDMA (Wideband CDMA) , HSDPA (High Speed Downlink Packet Access), HSUPA (High Speed Uplink Packet Access), IEEE 802.16, Long Term Evolution (LTE), LTE-A (Long Term Evolution-Advanced), 5G communication system, broadband wireless Wireless Mobile Broadband Service (WMBS), Bluetooth, Radio Frequency Identification (RFID), Infrared Data Association (IrDA), Ultra-Wideband (UWB), ZigBee, Near Field Communication ( Near Field Communication (NFC), Ultra Sound Communication (USC), Visible Light Communication (VLC), Wi-Fi, Wi-Fi Direct, and the like may be included. In addition, wired communication networks include wired local area network (LAN), wired wide area network (WAN), power line communication (PLC), USB communication, Ethernet, serial communication, optical/coaxial A cable may be included, and all protocols capable of providing a communication environment with other devices may be included, which are not now limited.

In the hardware system 2000 according to an embodiment of the present invention, each component stored in the form of a software module in the memory unit 2210 is in the form of a command executed by the processor unit 2100, and the memory interface unit 2200 and An interface with the communication unit 2310 is performed via the peripheral device interface unit 2300.

As set forth above, this specification contains many specific implementation details, but these should not be construed as limiting on the scope of any invention or claimables, but rather may be specific to a particular embodiment of a particular invention. It should be understood as a description of the features in Certain features that are described in this specification in the context of separate embodiments may also be implemented in combination in a single embodiment. Conversely, various features that are described in the context of a single embodiment can also be implemented in multiple embodiments individually or in any suitable subcombination. Further, while features may operate in particular combinations and are initially depicted as such claimed, one or more features from a claimed combination may in some cases be excluded from that combination, and the claimed combination is a subcombination. or sub-combination variations.

Similarly, while actions are depicted in the drawings in a particular order, it should not be construed as requiring that those actions be performed in the specific order shown or in the sequential order, or that all depicted actions must be performed to obtain desired results. In certain cases, multitasking and parallel processing can be advantageous. Further, the separation of various system components in the embodiments described above should not be understood as requiring such separation in all embodiments, and the program components and systems described may generally be integrated together into a single software product or packaged into multiple software products. You have to understand that you can.

Specific embodiments of the subject matter described herein have been described. Other embodiments are within the scope of the following claims. For example, the actions recited in the claims can be performed in a different order and still achieve desirable results. As an example, the processes depicted in the accompanying figures do not necessarily require the particular depicted order or sequential order in order to obtain desirable results. In certain implementations, multitasking and parallel processing may be advantageous.

The present description presents the best mode of the invention and provides examples to illustrate the invention and to enable those skilled in the art to make and use the invention. The specification thus prepared does not limit the invention to the specific terms presented. Therefore, although the present invention has been described in detail with reference to the above-described examples, those skilled in the art may make alterations, changes, and modifications to the present examples without departing from the scope of the present invention.

Therefore, the scope of the present invention should not be defined by the described embodiments, but by the claims.

The present invention relates to an apparatus and method for generating a learning model using a logarithmic scaling loss function. When optimization by the gradient descent method is performed using the log mean square error function, which is the loss function of the present invention, optimization can be performed stably. Therefore, the present invention has industrial applicability because it can be clearly practiced in reality as well as having a sufficient possibility of commercialization or business.

10: learning device
100: data processing unit
200: loss deduction unit
300: optimization unit

Claims (14)

When the learning model calculates output data for learning with respect to the input data for learning, the loss derivation unit calculates a loss representing the difference between the output data for learning and the label by using a log mean square error function obtained by converting the mean square error function to a logarithmic scale. doing; and
performing optimization by an optimizer to modify weights of the learning model so that the loss is minimized;
characterized in that it includes
A method for creating a learning model.
According to claim 1,
The step of calculating the loss is
The log mean square error function

Calculate the loss using
Wherein y is the label,
remind Is the output data for learning, characterized in that
A method for creating a learning model.
According to claim 2,
The log mean square error function is
Convert the structure of the mean square error function to a logarithmic function in negative form,
Apply the 1-X format,
to X Characterized in that it is generated by substituting
A method for creating a learning model.
Calculating output data for learning through a plurality of calculations to which a weight for which learning between a plurality of layers is not completed is applied to the input data for learning by the learning model;
calculating, by a loss derivation unit, a loss representing a difference between the output data for training and the label by using a log mean square error function obtained by converting the mean square error function into a log scale; and
performing optimization by an optimizer to modify weights of the learning model so that the loss is minimized;
characterized in that it includes
A method for creating a learning model.
According to claim 4,
The step of calculating the loss is
The log mean square error function

Calculate the loss using
Wherein y is the label,
remind Is the output data for learning, characterized in that
A method for creating a learning model.
According to claim 5,
If the learning model is a generative network,
The label is the input data for learning,
If the learning model is a classification network,
Characterized in that the label is a target value indicating the classification to which the learning input data belongs
A method for creating a learning model.
According to claim 6,
The log mean square error function is
Convert the structure of the mean square error function to a logarithmic function in negative form,
Apply the 1-X format,
to X Characterized in that it is generated by substituting
A method for creating a learning model.
When the learning model calculates output data for learning with respect to the input data for learning, deriving a loss that calculates a loss representing the difference between the output data for learning and the label using a log mean square error function converted to a logarithmic scale of the mean square error function wealth; and
an optimization unit that performs optimization to modify the weight of the learning model so that the loss is minimized;
characterized in that it includes
A device for generating a learning model.
According to claim 8,
The loss derivation unit
The log mean square error function

using
Calculate the loss;
Wherein y is the label,
remind Is the output data for learning, characterized in that
A device for generating a learning model.
According to claim 8,
The log mean square error function is
Convert the structure of the mean square error function to a logarithmic function in negative form,
Apply the 1-X format,
to X Characterized in that it is generated by substituting
A device for generating a learning model.
a data processing unit that prepares learning data including input data for learning and a label corresponding to the input data for learning, and inputs the input data for learning to a learning model in which learning has not been completed;
When the learning model calculates output data for learning through a plurality of operations to which a weight for which learning between a plurality of layers is not completed is applied to the input data for learning,
a loss derivation unit that calculates a loss representing a difference between the output data for training and the label by using a log mean square error function obtained by converting the mean square error function to a logarithmic scale; and
an optimization unit that performs optimization to modify the weight of the learning model so that the loss is minimized;
characterized in that it includes
A device for generating a learning model.
According to claim 11,
The loss derivation unit
The log mean square error function

using
Calculate the loss;
Wherein y is the label,
remind Is the output data for learning, characterized in that
A device for generating a learning model.
According to claim 12,
If the learning model is a generative network,
The label is the input data for learning,
If the learning model is a classification network,
Characterized in that the label is a target value indicating the classification to which the learning input data belongs
A device for generating a learning model.
According to claim 12,
The log mean square error function is
Convert the structure of the mean square error function to a logarithmic function in negative form,
Apply the 1-X format,
to X Characterized in that it is generated by substituting
A device for generating a learning model.

Images