JP7666142B2 - Information processing program, information processing method, and information processing device - Google Patents

Description

The present invention relates to an information processing program, an information processing method, and an information processing device.

In learning-type AI (Artificial Intelligence) that mainly uses neural networks, a technique called OOD (Out-of-Distribution Detection) judgment is sometimes used. OOD judgment can also be called OOD detection.

OOD judgment assumes that a machine learning model has been trained by performing empirical risk minimization (ERM) using data samples from a specific domain, and judges whether a given sample given at the time of inference belongs to that domain.

Conventional OOD determination methods utilize the reliability output by a machine learning model for each sample. Specifically, when the reliability is high, the sample is considered to be ID (In-Distribution), and when the reliability is low, it is considered to be OOD. Therefore, conventional OOD determination methods realize OOD determination by setting an appropriate threshold value for the reliability.

However, machine learning models may also output high confidence levels for OOD data. Several methods are known to solve the problem of high confidence levels being output for OOD data, including calibration, parametric assumption, anomaly detection, and outlier exposure.

Calibration prevents an unreasonably high reliability from being output by aligning the reliability output by a machine learning model with the accuracy rate of the machine learning model. Specific calibration methods include, for example, ODIN, FocalLoss, and Confidence Learning.

Parametric Assumption makes statistical assumptions about ID or OOD data, and determines the OOD based on how much the assumptions are violated. Known specific methods of Parametric Assumption include Mahalanobis, PriorNet, and FSSD, for example.

Anomaly detection is a technique that makes OOD decisions by taking advantage of the fact that models are unable to reconstruct unknown data well, mainly in networks with an autoencoder structure. AnoGAN and Efficient GAN are known as specific methods of anomaly detection.

For outlier exposure, it is assumed that OOD data is also available, even if only in small amounts, and the OOD data is regularized during learning so as not to output a high confidence level. Specific methods for outlier exposure include OE and OECC.

JP 2020-123830 A Special Publication No. 2020-533723

However, these conventional OOD determination methods cannot perform OOD determination efficiently for several reasons. For example, the OOD determination accuracy may be low, such as the ODIN and FocalLoss of the calibration described above.

In addition, OOD judgments can have a negative impact on the performance of the original task, for example the PriorNet in the Parametric Assumption mentioned above.

Furthermore, a completely separate model may be required for OOD determination, which increases costs. For example, the AnoGAN used for Anomaly Detection mentioned above falls into this category.

Furthermore, in the above-mentioned Outlier Exposure system, it is necessary to use OOD data for training, but it is difficult to comprehensively prepare OOD data for use in training.
In one aspect, the present invention aims to enable out-of-distribution determination to be performed efficiently.

Therefore, for a machine learning model including a feature extractor that generates a feature vector based on input data and a first network that performs inference on the feature vector , this information processing program inputs the feature vector generated by the feature extractor to a second network, outputs a loss inference result of the machine learning model, and performs an out-of-distribution determination as to whether or not the input data is included in a specific domain based on the loss inference result.

According to one embodiment, out-of-distribution determination can be performed efficiently.

FIG. 1 is a diagram illustrating a functional configuration of an information processing apparatus as an example of an embodiment. FIG. 2 is a diagram illustrating a process of an information processing device as an example of an embodiment. 1 is a flowchart illustrating a process in an information processing device as an example of an embodiment. FIG. 1 is a diagram illustrating a hardware configuration of an information processing apparatus according to an embodiment;

Below, an embodiment of the present information processing program, information processing method, and information processing device will be described with reference to the drawings. However, the embodiments shown below are merely examples, and there is no intention to exclude the application of various modified examples and techniques not explicitly stated in the embodiments. In other words, this embodiment can be implemented with various modifications without departing from its spirit. Furthermore, each figure is not intended to include only the components shown in the figure, but can include other functions, etc.

(A) Configuration FIG. 1 is a diagram showing a schematic functional configuration of an information processing apparatus 1 as an example of an embodiment, and FIG. 2 is a diagram for explaining the processing thereof.

2, the input data is image data, and an example of inferring a classification for this input data is shown. Hereinafter, the input data may be represented by the symbol x. The input data x is data to be inferred by a first inference processing unit 100 (to be described later), and is also data to be subjected to OOD judgment by a second inference processing unit 300 (to be described later).
As shown in FIG. 1, the information processing device 1 includes a first inference processing unit 100, a second inference processing unit 300, and a training processing unit 200.

The first inference processing unit 100 performs inference using the machine learning model 110. Let the inference in the machine learning model 110 be O(F(x)), and the loss function of the original task be Loss(y, y′).
The first inference processing unit 100 inputs input data to the machine learning model 110 and obtains an inference result output by the machine learning model 110.
As shown in FIG. 1, the machine learning model 110 includes a feature extractor 101 and a first network 102 .

The feature extractor 101 extracts features from input data. The feature extractor 101 may extract features from the input data by using, for example, a Convolutional Neural Network (CNN).
The feature extractor 101 generates a feature vector based on the input data (see symbol P1 in FIG. 2). Hereinafter, the feature vector may be represented by symbol z.
The feature vector z generated by the feature extractor 101 is expressed by the following equation (1).
z = F(x; θ _f )...(1)

F is a feature extraction function. A feature vector z is obtained by passing input data x through the feature extraction function F. θ _f represents the weight of the function F, and is updated by the optimization processing unit 204 described later during training of the machine learning model 110.

The feature vector generated by the feature extractor 101 is input to the first network 102, which outputs an inference result (see symbol P2 in FIG. 2). Hereinafter, the inference result may be represented by symbol y. The first network 102 may be configured as, for example, a fully connected layer (FC layer) (see symbol P3 in FIG. 2). The first network 102 may be called the final layer 102.
The inference result y output by the machine learning model 110 is expressed by the following equation (2).
y = O(z; θ _o )...(2)

O is a function. The inference result y is obtained by passing the feature vector z through the function O. θ _o represents the weight of the final layer O, and is updated by the optimization processing unit 204 when the machine learning model 110 is trained.
The second inference processing unit 300 uses the second network 301 to perform loss inference and OOD inference of the machine learning model 110.

In general, a machine learning model will output a correct inference result with a higher probability if a sample is sufficiently trained, i.e., if it is an ID. On the other hand, if the machine learning model is not sufficiently trained, there is a higher probability that it will output an incorrect inference result.

In other words, when a machine learning model outputs a result with a high degree of confidence, the loss is generally small, and when it outputs a result with a low degree of confidence, the loss is large. Also, even if the model has a high degree of confidence, the loss is large when the inference result is incorrect.
In the information processing device 1, such characteristics are utilized to explicitly learn OOD judgment for input data.
As shown in FIG. 1, the second inference processing unit 300 includes a second network 301 and an OOD inference unit 302 .

In the second inference processing unit 300, the feature vector generated by the feature extractor 101 of the machine learning model 110 is input to the second network 301, which outputs the loss inference result.

Hereinafter, the loss inference result by the second network 301 may be represented by the symbol w. In FIG. 2, an example is shown in which the second network 301 performs inference using a three-layer multi-layer fully connected layer (FC layer) (see symbol P4 in FIG. 2).

The second network 301 is a network that receives the feature vector z generated by the feature extractor 101 of the machine learning model 110 as input and outputs one scalar, and is a different network provided separately from the first network 102.
The loss inference result w generated by the second network 301 is expressed by the following equation (3).
w = L(z; _θl )...(3)

L is a function. The loss inference result w is obtained by passing the feature vector z through the function L. θ _l is the weight of the network that infers the loss, and is updated by the optimization processing unit 204 when the second network 301 is trained.
The OOD inference unit 302 infers an OOD judgment based on the loss inference result w inferred by the second network 301. That is, the OOD inference unit 302 performs an out-of-distribution judgment as to whether or not the input data x is included in a specific domain based on the loss inference result w inferred by the second network 301.

The OOD inference unit 302 compares the loss inference result w inferred by the second network 301 with a threshold value T, and if the loss inference result w is greater than the threshold value T, determines that the input data x is not included in the specific domain (OOD). Also, if the loss inference result w is equal to or less than the threshold value T, the OOD inference unit 302 determines that the input data x is included in the specific domain (ID).

The threshold T is a threshold of the loss inference result w that can best separate the ID and the OOD, calculated using the validation set. The threshold T may be arbitrarily set by the user.
The training processing unit 200 performs training (machine learning) of the machine learning model 110.

Training data is used to train the machine learning model 110. The training data includes input data and a correct answer to the input data. The training data is prepared in advance by, for example, a user.
As shown in FIG. 1 , the training processing unit 200 includes a first loss calculation unit 201 , a second loss calculation unit 202 , a total loss calculation unit 203 , and an optimization processing unit 204 .

The first loss calculation unit 201 obtains a correct answer t for the input data x (see symbol P5 in FIG. 2). Then, the first loss calculation unit 201 calculates the loss (distance, error) of the input data x for the correct answer t. The loss of the input data x for the correct answer t may be represented by symbol _l1 (see symbol P6 in FIG. 2). The loss _l1 is expressed by the following formula (4).
l ₁ = Loss(y, t)...(4)

The loss l ₁ may be called the first loss l _1. The first loss l ₁ is a loss calculated from the inference result y output by the first network 102 of the machine learning model 110 and the correct answer data t, and represents a loss that actually occurred in the machine learning model 110.

The second loss calculation unit 202 calculates the L1 distance l2 between the loss inference result w generated by the second network 301 and the first loss _l1 calculated by the first loss calculation unit 201 (see symbol P7 in FIG. 2). The _L1 distance _l2 is expressed by the following formula (5).
l ₂ = |l ₁ - w| ...(5)
The L1 distance _l2 may be called the second loss _l2 . The second loss _l2 represents the loss of the inferred machine learning model 110.

The total loss calculation unit 203 calculates a total loss l total by adding up the first loss l ₁ calculated by the first loss calculation unit 201 and the second loss l ₂ calculated by the second loss calculation unit 202 (see symbol P8 in FIG. 2). The _total loss l _total is expressed by the following formula (6).
l _total = l ₁ + l ₂ ...(6)
The total loss l _total is the sum of the first loss l ₁ and the second loss l ₂ , and indicates the cost to be minimized by the optimization processing unit 204 .

The optimization processing unit 204 calculates the gradient that minimizes the total loss l _total calculated by the total loss calculation unit 203, and performs error backpropagation. For example, the steepest descent method (gradient method) may be used for the error backpropagation, and it can be modified as appropriate. The gradient may be expressed as δl _total /δθ.

The optimization processor 204 calculates weights θ _f , θ _o , and θ _l that minimize the total loss l _total . The calculation of these weights θ _f , θ _o , and θ _l can be realized using various known methods, and the description thereof will be omitted.

The optimization processing unit 204 reflects the calculated weights θ _f , θ _o , and θ _l in the first network 102 and the second network 301 of the machine learning model 110. That is, the optimization processing unit 204 uses these calculated parameters (weights θ _f , θ _o , and θ _l ) to update the parameters (weights θ _f and θ _o ) of the machine learning model 110 (first network 102) and the parameter (weight θ _l ) of the second network 301, respectively.
The weights θ _f , θ _o and θ _l calculated by the optimization processor 204 are stored in a predetermined storage area such as the storage device 13 (see FIG. 4) described later.

(B) Operation The process of the information processing device 1 as one example of the embodiment configured as described above will be described with reference to the flowchart (steps S1 to S12) shown in FIG.

In the flowchart shown in FIG. 3, the processes in steps S1 to S6 represent the inference process for the input data, and the processes in steps S1 to S3 and S7 to S12 represent the training process for the machine learning model 110 (first network 102) and the second network 301.

First, the inference process for input data in the information processing device 1 will be described.
In step S1, the first inference processing unit 100 acquires input data x.
In step S2, the feature extractor 101 of the machine learning model 110 extracts features from the input data x, and calculates the feature vector z using the above-mentioned equation (1).
In step S3, the second network 301 of the second inference processing unit 300 calculates the loss inference result w by using the above-mentioned formula (3).

In step S4, the OOD inference unit 302 of the second inference processing unit 300 compares the loss inference result w inferred by the second network 301 with a threshold value T, and checks whether the loss inference result w is less than or equal to the threshold value T.

If the loss inference result w is equal to or less than the threshold T (see the "True" route in step S4), in step S6, the OOD inference unit 302 determines that the input data x is an ID. Then, the process ends.

On the other hand, if the result of the check in step S4 is that the loss inference result w is greater than the threshold T (see the "False" route in step S4), in step S5, the OOD inference unit 302 determines that the input data x is OOD. Then, the process ends.

Next, the training process in the information processing device 1 will be described.
In step S1, the first inference processing unit 100 acquires input data x.
In step S2, the feature extractor 101 of the machine learning model 110 extracts features from the input data x, and calculates the feature vector z using the above-mentioned equation (1).
In step S7, the first network 102 of the machine learning model 110 calculates the inference result y using the above-mentioned equation (2).

In step S8, the first loss calculation unit 201 of the training processing unit 200 obtains a correct answer t for the input data x.
In step S9, the first loss calculation unit 201 calculates the first loss _l1 by using the above-mentioned formula (4).
In addition, in step S3, the second network 301 of the second inference processing unit 300 calculates the loss inference result w by using the above-mentioned formula (3).

In step S10, the second loss calculation unit 202 calculates the second loss _l2 using the above-mentioned formula (5). Also, the total loss calculation unit 203 calculates the _total loss ltotal using the above-mentioned formula (6).
In step S11, the optimization processor 204 calculates weights θ _f , θ _o and θ _l that minimize the total loss l _total by using backpropagation.

In step S12, the optimization processing unit 204 uses the calculated weights θ _f , θ _o , and θ _l to update the parameters (weights θ _f and θ _o ) of the machine learning model 110 (first network 102) and the parameter (weight θ _l ) of the second network 301. Then, the process returns to step S1.
The training process may be performed, for example, until a preset number of repetitions is reached, or until the inference accuracy of the machine learning model reaches a predetermined threshold or higher, and may be modified as appropriate.

(C) Effects In this way, according to the information processing device 1 as an example of the embodiment, by providing the second network 301 in the second inference processing unit 300, it is possible to directly learn the OOD judgment itself, and high accuracy can be obtained in OOD judgment. In addition, it has been experimentally confirmed that there is almost no degradation in the performance of the original task.

Furthermore, a second network 301 is provided to which the feature vector generated by the feature extractor 101 of the machine learning model 110 is input, and the second network 301 infers the loss of the machine learning model 110. This eliminates the need to provide a new machine learning model dedicated to inferring the OOD, and enables detection of the OOD at low cost.
Furthermore, in the information processing device 1, there is no need to prepare training data for inferring the OOD.

(D) Others FIG. 4 is a diagram illustrating a hardware configuration of the information processing device 1 as an example of an embodiment.

The information processing device 1 is a computer and has as its components a processor 11, a memory 12, a storage device 13, a graphics processing device 14, an input interface 15, an optical drive device 16, a device connection interface 17, and a network interface 18. These components 11 to 18 are configured to be able to communicate with each other via a bus 19.

The processor (processing unit) 11 controls the entire information processing device 1. The processor 11 may be a multiprocessor. The processor 11 may be, for example, any one of a CPU, an MPU (Micro Processing Unit), a DSP (Digital Signal Processor), an ASIC (Application Specific Integrated Circuit), a PLD (Programmable Logic Device), and an FPGA (Field Programmable Gate Array). The processor 11 may also be a combination of two or more types of elements from the CPU, MPU, DSP, ASIC, PLD, and FPGA.

Then, the processor 11 executes a control program (information processing program, not shown) for the information processing device 1, thereby realizing the functions of the first inference processing unit 100, the training processing unit 200, and the second inference processing unit 300 illustrated in FIG. 1.

The information processing device 1 realizes the functions of the first inference processing unit 100, the training processing unit 200, and the second inference processing unit 300 by executing a program (information processing program, OS program) recorded on, for example, a computer-readable non-transitory recording medium.

The program describing the processing contents to be executed by the information processing device 1 can be recorded on various recording media. For example, the program to be executed by the information processing device 1 can be stored in the storage device 13. The processor 11 loads at least a part of the program in the storage device 13 into the memory 12 and executes the loaded program.

The program to be executed by the information processing device 1 (processor 11) can also be recorded on a non-transitory portable recording medium such as an optical disk 16a, a memory device 17a, or a memory card 17c. The program stored on the portable recording medium becomes executable after being installed in the storage device 13, for example, under control of the processor 11. The processor 11 can also read and execute the program directly from the portable recording medium.

Memory 12 is a storage memory including ROM (Read Only Memory) and RAM (Random Access Memory). The RAM of memory 12 is used as the main storage device of information processing device 1. The RAM temporarily stores at least a portion of the programs to be executed by processor 11. In addition, memory 12 stores various data necessary for processing by processor 11.

The storage device 13 is a storage device such as a hard disk drive (HDD), a solid state drive (SSD), a storage class memory (SCM), etc., and stores various data. The storage device 13 is used as an auxiliary storage device of the information processing device 1.
An OS program, a control program, and various data are stored in the storage device 13. The control program includes an information processing program.

As the auxiliary storage device, a semiconductor storage device such as an SCM or a flash memory can also be used. In addition, multiple storage devices 13 may be used to configure a RAID (Redundant Arrays of Inexpensive Disks).

The storage device 13 may store various data generated by the first inference processing unit 100, the second inference processing unit 300, and the training processing unit 200.

The monitor 14a is connected to the graphics processing device 14. The graphics processing device 14 displays images on the screen of the monitor 14a according to instructions from the processor 11. Examples of the monitor 14a include a display device using a CRT (Cathode Ray Tube) and a liquid crystal display device.

A keyboard 15a and a mouse 15b are connected to the input interface 15. The input interface 15 transmits signals sent from the keyboard 15a and the mouse 15b to the processor 11. Note that the mouse 15b is an example of a pointing device, and other pointing devices can also be used. Examples of other pointing devices include a touch panel, a tablet, a touch pad, and a trackball.

The optical drive device 16 uses laser light or the like to read data recorded on the optical disc 16a. The optical disc 16a is a portable, non-transient recording medium on which data is recorded so that it can be read by the reflection of light. Examples of optical discs 16a include DVDs (Digital Versatile Discs), DVD-RAMs, CD-ROMs (Compact Disc Read Only Memory), and CD-Rs (Recordable)/RWs (ReWritable).

The device connection interface 17 is a communication interface for connecting peripheral devices to the information processing device 1. For example, a memory device 17a or a memory reader/writer 17b can be connected to the device connection interface 17. The memory device 17a is a non-transient recording medium equipped with a communication function with the device connection interface 17, such as a USB (Universal Serial Bus) memory. The memory reader/writer 17b writes data to the memory card 17c or reads data from the memory card 17c. The memory card 17c is a card-type non-transient recording medium.

The network interface 18 is connected to a network. The network interface 18 transmits and receives data via the network. Other information processing devices, communication devices, etc. may also be connected to the network.

The disclosed technology is not limited to the above-described embodiment, and can be modified in various ways without departing from the spirit of the present embodiment. Each configuration and process of the present embodiment can be selected as needed, or can be combined as appropriate.

For example, in the above-described embodiment, the OOD inference unit 302 determines whether or not the loss inference result w is an OOD by comparing it with the threshold value T, but this is not limited to this and can be implemented in various modified forms. For example, ensemble learning may be performed to integrate the estimation results of one or more other models such as Mahalanobis, ODIN, and FSSD. When using such ensemble learning, logistic regression or the like may be used, which is expected to further improve accuracy. In the present information processing device 1, ensembling with other methods is easy, and even higher accuracy can be achieved.

In addition, in the above-described embodiment, the input data is image data, and an example is shown in which a classification for this input data is inferred, but the present invention is not limited to this and can be implemented with appropriate modifications.

In addition, in the above-described embodiment, an example has been shown in which the feature extractor 101 extracts features of the input data using a Convolutional Neural Network (CNN), but this is not limited to this and can be modified as appropriate.

In the above embodiment, the second network 301 performs inference using a three-layer fully connected layer (FC layer), but the present invention is not limited to this. The second network 301 can be modified as appropriate.
Moreover, the above disclosure enables a person skilled in the art to implement and manufacture the present embodiment.

(E) Supplementary Notes The following supplementary notes are further disclosed with respect to the above-described embodiments.
(Appendix 1)
A machine learning model including a feature extractor that generates feature data based on input data, and a first network that performs inference on the feature data,
Input the feature vector generated by the feature extractor into a second network, and output the loss inference result of the machine learning model;
An information processing program that causes a processor to execute a process of determining whether the input data is included in a specific domain or not based on the loss inference result.

(Appendix 2)
The information processing program according to claim 1, characterized in that the process of making the out-of-distribution determination determines that the input data is not included in the specific domain when the loss inference result is greater than a threshold, and determines that the input data is included in the specific domain when the loss inference result is equal to or less than the threshold.

(Appendix 3)
When a machine learning model including a feature extractor that generates feature data based on input data and a first network that performs inference on the feature data is trained using training data including sample data and correct answer data included in a specific domain,
causing the feature extractor to generate feature data based on the sample data;
causing the first network to perform inference on the feature data based on the sample data and output an inference result;
performing a first calculation to calculate a first loss of the inference result for the ground truth data;
a second network that outputs a loss inference result of the machine learning model based on the feature vector generated by the feature extractor, and causes the second network to perform a second calculation to calculate an L1 distance between the loss inference result output by performing inference on the feature data based on the sample data and the first loss calculated by the first calculation;
An information processing program that causes a processor to execute a process of calculating each parameter of the machine learning model and the second network by obtaining a gradient that minimizes the first loss and the L1 distance and performing backpropagation of the error.

(Appendix 4)
A machine learning model including a feature extractor that generates feature data based on input data, and a first network that performs inference on the feature data,
Input the feature vector generated by the feature extractor into a second network, and output the loss inference result of the machine learning model;
An information processing method, comprising: a processor executing a process of determining whether the input data is included in a specific domain or not based on the loss inference result.

(Appendix 5)
The information processing method according to claim 4, wherein the process of determining whether the input data is out of distribution determines that the input data is not included in the specific domain when the loss inference result is greater than a threshold, and determines that the input data is included in the specific domain when the loss inference result is equal to or less than the threshold.

(Appendix 6)
When a machine learning model including a feature extractor that generates feature data based on input data and a first network that performs inference on the feature data is trained using training data including sample data and correct answer data included in a specific domain,
The feature extractor generates feature data based on the sample data;
The first network performs inference on the feature data based on the sample data and outputs an inference result;
performing a first calculation to calculate a first loss of the inference result for the ground truth data;
A second network that outputs a loss inference result of the machine learning model based on the feature vector generated by the feature extractor performs a second calculation to calculate an L1 distance between the loss inference result output by performing inference on the feature data based on the sample data and the first loss calculated by the first calculation;
An information processing method, characterized in that a processor executes a process of calculating each parameter of the machine learning model and the second network by finding a gradient that minimizes the first loss and the L1 distance and backpropagating the error.

(Appendix 7)
A machine learning model including a feature extractor that generates feature data based on input data, and a first network that performs inference on the feature data,
Input the feature vector generated by the feature extractor into a second network, and output the loss inference result of the machine learning model;
The information processing device further comprises a processing unit that performs an out-of-distribution determination as to whether the input data is included in a specific domain based on the loss inference result.

(Appendix 8)
The information processing device of claim 7, wherein the process of performing the out-of-distribution determination determines that the input data is not included in the specific domain when the loss inference result is greater than a threshold, and determines that the input data is included in the specific domain when the loss inference result is equal to or less than the threshold.

(Appendix 9)
When a machine learning model including a feature extractor that generates feature data based on input data and a first network that performs inference on the feature data is trained using training data including sample data and correct answer data included in a specific domain,
causing the feature extractor to generate feature data based on the sample data;
causing the first network to perform inference on the feature data based on the sample data and output an inference result;
performing a first calculation to calculate a first loss of the inference result for the ground truth data;
A second network that outputs a loss inference result of the machine learning model based on the feature vector generated by the feature extractor performs a second calculation to calculate an L1 distance between the loss inference result output by performing inference on the feature data based on the sample data and the first loss calculated by the first calculation;
An information processing device comprising: a processing unit that calculates each parameter of the machine learning model and the second network by obtaining a gradient that minimizes the first loss and the L1 distance and backpropagating the error.

1 Information processing device 11 Processor (processing unit)
12 Memory 13 Storage device 14 Graphic processing device 14a Monitor 15 Input interface 15a Keyboard 15b Mouse 16 Optical drive device 16a Optical disk 17 Device connection interface 17a Memory device 17b Memory reader/writer 17c Memory card 18 Network interface 18a Network 19 Bus 100 First inference processing unit 101 Feature extractor 102 First network 110 Machine learning model 200 Training processing unit 201 First loss calculation unit 202 Second loss calculation unit 203 Total loss calculation unit 204 Optimization processing unit 300 Second inference processing unit 301 Second network 302 OOD inference unit

Claims (7)

A machine learning model including a feature extractor that generates a feature vector based on input data, and a first network that performs inference on the feature vector ,
Input the feature vector generated by the feature extractor into a second network, and output a loss inference result of the machine learning model;
An information processing program that causes a processor to execute a process of determining whether the input data is included in a specific domain or not based on the loss inference result. The information processing program according to claim 1, characterized in that the process of performing the out-of-distribution determination determines that the input data is not included in the specific domain when the loss inference result is greater than a threshold, and determines that the input data is included in the specific domain when the loss inference result is equal to or less than the threshold. When a machine learning model including a feature extractor that generates a feature vector based on input data and a first network that performs inference on the feature vector is trained using training data including sample data and correct answer data included in a specific domain,
causing the feature extractor to generate a feature vector based on the sample data;
causing the first network to perform inference on the feature vector based on the sample data and output an inference result;
performing a first calculation to calculate a first loss of the inference result for the ground truth data;
a second network that outputs a loss inference result of the machine learning model based on the feature vector generated by the feature extractor, and causes the second network to perform a second calculation to calculate an L1 distance between the loss inference result output by performing inference on the feature vector based on the sample data and the first loss calculated by the first calculation;
An information processing program that causes a processor to execute a process of calculating each parameter of the machine learning model and the second network by obtaining a gradient that minimizes the first loss and the L1 distance and performing backpropagation of the error. A machine learning model including a feature extractor that generates a feature vector based on input data, and a first network that performs inference on the feature vector ,
Input the feature vector generated by the feature extractor into a second network, and output a loss inference result of the machine learning model;
An information processing method, comprising: a processor executing a process of determining whether the input data is included in a specific domain or not based on the loss inference result. When a machine learning model including a feature extractor that generates a feature vector based on input data and a first network that performs inference on the feature vector is trained using training data including sample data and correct answer data included in a specific domain,
The feature extractor generates a feature vector based on the sample data;
The first network performs inference on the feature vector based on the sample data and outputs an inference result;
performing a first calculation to calculate a first loss of the inference result for the ground truth data;
A second network that outputs a loss inference result of the machine learning model based on the feature vector generated by the feature extractor performs a second calculation to calculate an L1 distance between the loss inference result output by performing inference on the feature vector based on the sample data and the first loss calculated by the first calculation;
By finding a gradient that minimizes the first loss and the L1 distance and backpropagating the error,
An information processing method, characterized in that a processor executes a process of calculating parameters of the machine learning model and the second network. A machine learning model including a feature extractor that generates a feature vector based on input data, and a first network that performs inference on the feature vector ,
Input the feature vector generated by the feature extractor into a second network, and output a loss inference result of the machine learning model;
The information processing device further comprises a processing unit that performs an out-of-distribution determination as to whether the input data is included in a specific domain based on the loss inference result. When a machine learning model including a feature extractor that generates a feature vector based on input data and a first network that performs inference on the feature vector is trained using training data including sample data and correct answer data included in a specific domain,
causing the feature extractor to generate a feature vector based on the sample data;
causing the first network to perform inference on the feature vector based on the sample data and output an inference result;
performing a first calculation to calculate a first loss of the inference result for the ground truth data;
A second network that outputs a loss inference result of the machine learning model based on the feature vector generated by the feature extractor performs a second calculation to calculate an L1 distance between the loss inference result output by performing inference on the feature vector based on the sample data and the first loss calculated by the first calculation;
An information processing device comprising: a processing unit that calculates each parameter of the machine learning model and the second network by obtaining a gradient that minimizes the first loss and the L1 distance and backpropagating the error.

Images