JP2023183079A - Training data generation program, training data generation method, and information processing apparatus - Google Patents

Abstract

To generate training data which is unlikely to cause regression in retraining a machine learning model.SOLUTION: A computer is caused to execute processing of: acquiring multiple pieces of first data which are input to a machine learning model to output a correct answer and multiple pieces of second data with which the machine learning model outputs an incorrect answer; selecting multiple pieces of third data from among the multiple pieces of first data on the basis of a probability according to a magnitude of a value of a loss function for the case of inputting output of the machine learning model for each of the multiple pieces of first data; and generating training data including the multiple pieces of second data and the multiple pieces of third data.SELECTED DRAWING: Figure 3

Description

The present invention relates to a training data generation program, a training data generation method, and an information processing device.

In a system using a machine learning model such as a neural network (NN), when an output that is undesirable to the user occurs, the machine learning model may be corrected.

For example, if an automatic driving system using cameras recognizes a stop road sign as a speed limit, it may cause an accident, so the machine learning model needs to be modified so that the system recognizes it correctly.

Machine learning models are created not according to specifications, but according to input training data. Therefore, corrections are made by inputting training data.

However, in such a conventional method of correcting a machine learning model using training data, it is necessary to collect data necessary for correction, but it is not always possible to correct errors using the collected data.

That is, by performing retraining, there is a possibility that an incorrect inference will be made after retraining on data that was correctly inferred before retraining, and degradation may occur. Therefore, it is desired to suppress the occurrence of degradation when retraining a machine learning model.

International Publication No. 2021/064787 Japanese Patent Application Publication No. 2019-174870 US Patent Application Publication No. 2021/0073627 US Patent Application Publication No. 2020/0160207

When retraining a machine learning model, successful data is included in the training data to prevent degradation, but including all successful data increases computational cost. I would like to select this because the data with greater loss is more likely to be degraded. However, data with large losses may cause significant degradation.

In one aspect, the present invention aims to generate training data that is less prone to degradation when retraining a machine learning model.

For this reason, this training data generation program acquires the first plurality of data for which the machine learning model outputs a correct answer when input, and the second plurality of data for which the machine learning model outputs an incorrect answer, and Selecting a third plurality of data from the first plurality of data based on a probability according to the magnitude of the value of a loss function when inputting the output of the machine learning model for each of the plurality of data, A computer is caused to perform a process of generating training data including the second plurality of data and the third plurality of data.

According to one embodiment, training data that is less likely to degrade can be generated when retraining a machine learning model.

FIG. 1 is a diagram schematically showing a functional configuration of an information processing device as an example of an embodiment. FIG. 2 is a diagram for explaining a patch generation method using particle swarm optimization by a training processing unit in an information processing device as an example of an embodiment. FIG. 3 is a diagram for explaining processing by a sampling processing unit of an information processing device as an example of an embodiment. FIG. 2 is a diagram for explaining a method of calculating a fitness value by a fitness calculation unit in an information processing device as an example of an embodiment. 3 is a flowchart for explaining processing of a sampling processing unit in an information processing device as an example of an embodiment. 7 is a flowchart for explaining processing of an optimization processing unit in an information processing device as an example of an embodiment. FIG. 3 is a diagram illustrating a specific example of processing by an optimization processing unit in an information processing device as an example of an embodiment. FIG. 3 is a diagram illustrating a comparison between a case in which a machine learning model is modified by a training processing unit of an information processing device as an example of an embodiment, and a case in which a machine learning model is modified by a conventional method. . 1 is a diagram illustrating a hardware configuration of an information processing device as an example of an embodiment.

Embodiments of the present training data generation program, training data generation method, and information processing device will be described below with reference to the drawings. However, the embodiments shown below are merely illustrative, and there is no intention to exclude the application of various modifications and techniques not specified in the embodiments. That is, this embodiment can be modified in various ways (such as combining the embodiment and each modification) without departing from the spirit thereof. Furthermore, each figure is not intended to include only the constituent elements shown in the figure, but may include other functions.

(A) Configuration FIG. 1 is a diagram schematically showing the functional configuration of an information processing device 1 as an example of an embodiment.

The information processing device 1 implements a model correction function that corrects a trained neural network model (machine learning model). The information processing device 1 has a function as a training processing section 100, as shown in FIG.

The training processing unit 100 performs retraining (machine learning) of a trained machine learning model (trained model). The training processing unit 100 has a function of generating training data, and uses the generated training data to train a machine learning model.

In addition, the training processing unit 100 identifies defective parts in the trained model and identifies weights to be corrected, and then generates a patch for the weight to be corrected and applies it to the trained model. Generate a trained model for .

The machine learning model may be, for example, a classifier for classifying input data into multiple classes. The machine learning model may be, for example, a deep learning model (deep neural network). The neural network may be a hardware circuit or a virtual network using software that connects layers virtually constructed on a computer program by the processor 11 (see FIG. 9) or the like.

Note that defect location identification in the trained model can be realized by a known method, and its explanation will be omitted.

The training processing unit 100 may perform patch generation by, for example, searching for a weight value that reduces failure using metaheuristics. The training processing unit 100 sets weights to be corrected in the neural network model by particle swarm optimization (PSO), and realizes retraining of the machine learning model.

FIG. 2 is a diagram for explaining a patch generation method using particle swarm optimization by the training processing unit 100 in the information processing device 1 as an example of the embodiment.

In patch generation using the particle swarm optimization method, a plurality of (k in the example shown in Fig. 2) particles x ₁ each having a plurality of (n in the example shown in Fig. 2) weights w ₁ to w _n as vectors. ~x _k is used. These plural particles x ₁ to x _k correspond to a particle group.

For each of these particles, the training processing unit 100 applies failure data to be corrected and success data sampled from a plurality of success data to a machine learning model in which weights are replaced using particle values (weights, parameters). data is input, and fitness values are calculated based on the results. The fitness value corresponds to an evaluation value for evaluating each particle.

The training processing unit 100 identifies the particle with the highest fitness value (Global best), and adjusts each of the other particles to approach the value (parameter value) of the global best particle (first particle). (second particle) value (parameter value) is updated.

The training processing unit 100 repeatedly executes such processing until the termination condition is satisfied, and reflects the final global best particle value as the weight of the machine learning model. The termination conditions are that the number of repetitions reaches a predetermined number, and that the global best particles do not change a certain number of times.

The training processing section 100 includes a sampling processing section 101 and an optimization processing section 105, as shown in FIG.

The sampling processing unit 101 samples success data to be used for particle swarm optimization from among a plurality of success data when the training processing unit 100 performs patch generation using particle swarm optimization.

Specifically, the sampling processing unit 101 calculates the magnitude of the loss (value of the loss function) calculated based on the predicted value obtained by inputting the success data into the machine learning model and the correct value in the success data. Make a roulette selection according to the That is, the sampling processing unit 101 performs random selection such that data with a larger loss (successful data) is selected with a higher probability. Sampling success data may also be referred to as success data sampling.

The sampling processing unit 101 includes a loss calculation unit 102, a probability distribution conversion unit 103, and a success data sampling unit 104, as shown in FIG.

The loss calculation unit 102 calculates the value of the loss function L based on the predicted value obtained by inputting the success data into the machine learning model and the correct value in the success data. Hereinafter, the value of the loss function L may simply be referred to as a loss.

The loss calculation unit 102 stores the calculated loss in a predetermined storage area such as the storage device 13 (see FIG. 9).

The probability distribution conversion unit 103 converts the value of the loss function L of each success data calculated by the loss calculation unit 102 into a probability that each success data is selected, so that each success data is selected for a plurality of success data. Generate a distribution of probabilities (probability distribution).

When the value of the loss function L of data s _i is expressed as L(s _i ), the probability distribution conversion unit 103 calculates the probability p( _{s i} ) is calculated.

確率p(s_i)は、データs_iを選択する確率を表す。

The probability p(s _i ) represents the probability of selecting data s _i .

The probability distribution conversion unit 103 converts the probability distribution among the plurality of success data by allocating (proportional allocation) the interval from 0 to 1.000 according to the value of each probability p(s _i ) of the plurality of success data. create. The probability distribution created in this way may be called a successful data selection probability distribution.

The success data selection probability distribution represents the probability according to the magnitude of the value of the loss function L when the output of the machine learning model for each of the plurality of success data (the first plurality of data) is input.

In the success data selection probability distribution, the larger the loss value, the higher the probability of selection. The probability distribution conversion unit 103 may obtain the cumulative probability of the probability p(s _i ) calculated for each success data.

The probability distribution conversion unit 103 may set the probability distribution among the plurality of success data within the numerical range of 0 to 1, for example.

The probability distribution conversion unit 103 stores the calculated probability p(s _i ), information on the success data selection probability distribution, etc. in a predetermined storage area of the storage device 13 or the like.

The success data sampling unit 104 selects success data to be used in the particle swarm optimization method from a plurality of success data based on the success data selection probability distribution generated by the probability distribution conversion unit 103. Hereinafter, success data selected from a plurality of success data and used in the particle swarm optimization method may be referred to as a success data sample.

The success data sampling unit 104 generates a random number according to the success data selection probability distribution set by the probability distribution conversion unit 103, and determines success data with a probability corresponding to the value of this random number as a success data sample. This method of selecting successful data samples can be called roulette selection.

For example, when the success data selection probability distribution is set within the numerical range of 0 to 1, the success data sampling unit 104 prepares random numbers within the numerical range of 0 to 1, and in the success data selection probability distribution, this Success data corresponding to a probability range corresponding to the value of the random number is determined as a success data sample. The success data sampling unit 104 may be prepared by generating random numbers, or may be prepared by obtaining random numbers generated by a random number generation unit (not shown).

The success data sampling unit 104 selects success data samples from a plurality of success data by roulette selection. In this way, by selecting successful data close to the value of the loss function L with a certain degree of randomness, it is possible to search with a gradual trade-off between correction and degradation. .

The success data sample corresponds to the third plurality of data selected from the plurality of success data (the first plurality of data) based on the success data selection probability distribution.

FIG. 3 is a diagram for explaining processing by the sampling processing unit 101 of the information processing device 1 as an example of the embodiment.

FIG. 3 shows an example in which one data set is selected from four data sets (success data). Each data set includes image data and a label (see P1). In the example shown in Figure 3, the image data are traffic signs, and "17. No entry", "14. Stop", "3. Speed limit (60 km/h)" and "34. Turn left ahead". One of the labels is set.

The loss calculation unit 102 inputs each of these data sets (success data) to a machine learning model that reflects the weight of the global best, and calculates each loss (see symbol P2).

The probability distribution conversion unit 103 calculates the probability p(s _i ) of selecting each data for each data set (success data) based on the above equation (1) (see symbol P3).

The probability distribution conversion unit 103 creates a success data selection probability distribution by allocating (proportional allocation) the interval from 0 to 1.000 according to the value of each probability p(s _i ) of a plurality of success data ( (See code P4).

In the success data selection probability distribution indicated by P4 in FIG. 3, the interval from 0.000 to 0.159 corresponds to the success data of the image "17. No entry", and the interval from 0.159 to 0.389 corresponds to the success data of the image "14. Stop". Respond to success data. Furthermore, the interval from 0.389 to 0.727 corresponds to the success data of the image "3. Speed limit (60 km/h)", and the interval from 0.727 to 1.000 corresponds to the success data of the image "34. Turn left ahead".

After that, the success data sampling unit 104 generates a random number within the numerical range of 0 to 1.000, and in the success data selection probability distribution, determines success data with a probability corresponding to the value of this random number as a success data sample (symbol P5 reference).

In the example shown in FIG. 3, the value of the random number prepared by the success data sampling unit 104 is 0.531, and this value corresponds to the success of the image "3. Speed limit (60 km/h)" in the success data selection probability distribution. Respond to data. Therefore, the success data sampling unit 104 determines the success data of the image "3. Speed limit (60 km/h)" as the success data sample.

The success data sampling unit 104 determines a predetermined number of success data samples. The success data sampling unit 104 selects success data (the first plurality of data) is selected as the successful data sample (third plurality of data).

The success data sampling unit 104 stores the success data sample in a predetermined storage area of the storage device 13 or the like.

When the global best weight changes during each iteration of particle swarm optimization, the optimization processing unit 105 newly selects successful data with a large loss using a machine learning model to which the weight is applied. That is, the optimization processing unit 105 samples success data when the global best weight changes during each iteration of particle swarm optimization.

If the weight changes in a machine learning model, the loss of each data will also change, so the successful data with a high loss that should be selected differs depending on the machine learning model.

By selecting (re-selecting) successful data that is likely to cause degradation in a new machine learning model, that is, successful data that is close to the decision boundary, it is possible to guide weights that will prevent degradation from occurring with such successful data. .

As shown in FIG. 1, the optimization processing unit 105 includes a particle initialization processing unit 106, a fitness calculation unit 107, a global best selection unit 108, a resampling determination unit 109, a particle updating unit 110, and a termination determination unit 111.

In particle swarm optimization, a particle swarm having a plurality of particles including a plurality of combinations of weights (parameter values) is used.

The training processing unit 100 selects particles other than the global best (second particle) based on the values (weights, parameter values) of the particle (global best, first particle) with the highest fitness value (evaluation value) in the particle group. The particle group is optimized by updating the values (weights, parameter values) of the particles.

After the optimization of the particle group is completed, the training processing unit 100 reflects the final global best particle value in the weight of the machine learning model. This generates a modified model.

The particle initialization processing unit 106 initializes each particle in the particle swarm optimization method with a normal distribution. Note that the average and variance may be determined from the neighboring weight distribution.

The fitness calculation unit 107 calculates the fitness value (fitness) of each particle.

FIG. 4 is a diagram for explaining a method of calculating a fitness value by the fitness calculation unit 107 in the information processing device 1 as an example of the embodiment.

The fitness value can be used as an evaluation value for evaluating particles. The fitness value may be called a particle evaluation value. The particle with the highest fitness value (global best) may be said to be the best particle among the plurality of particles (particle group).

The fitness calculation unit 107 adds failure data to be corrected and data determined by the success data sampling unit 104 to a machine learning model (replaced model) in which weights identified as correction targets by defect location identification are replaced with particle values. Enter the success data sample.

The success data sample corresponds to the first plurality of data for which the machine learning model will output a correct answer if input. Further, the failure data to be corrected corresponds to the second plurality of data for which the machine learning model outputs an incorrect answer when input.

Hereinafter, failure data to be corrected may be represented by the symbol I _neg , and success data samples may be represented by the symbol I _{pos_sampled} .

The number of correct outputs obtained by inputting failure data I _neg into a replaced model, that is, the number of pieces of data that could be corrected by replacing the weights of the machine learning model, is represented by the symbol N _patched . Furthermore, the loss of failure data I _neg is expressed as L(I _neg ).

On the other hand, the number of correct outputs obtained by inputting the successful data sample I _{pos_sampled} to the replaced model, that is, the number of pieces of data that remain successful even in the replaced model, is represented by the symbol N _intact . Further, the loss of the successful data sample I _{pos_sampled} is expressed as L(I _pos ).

The fitness calculation unit 107 calculates a fitness value based on the following equation (2).

fitness={(N _patched +1)/ (L(I _neg )+1)}+{ (N _intact +1)/ (L(I _pos )+1)}...(2)
The fitness calculation unit 107 stores the calculated fitness value in a predetermined storage area of the storage device 13 or the like.

For example, fitness values of particles x ₁ , . . . , x _k may be expressed as fitness(x _{1 )} , .

For example, for particle x ₁ , when N _patched = 21, L(I _neg ) = 1.184, N _intact = 981, L(I _pos ) = 0.732,
fitness(x ₁ )={(21 + 1)/ 1.184 + 1)}+{(98+1)/ (0.732+1)}=67.23
becomes.

The global best selection unit 108 selects the particle with the highest fitness value from among the plurality of particles as the global best p _g . The global best p _g can be expressed as follows.

グローバルベスト選択部１０８は、選択したグローバルベストp_gの情報を、グローバルベスト履歴として、記憶装置１３等の所定の記憶領域に記憶させる。

The global best selection unit 108 stores information on the selected global best p _g in a predetermined storage area such as the storage device 13 as a global best history.

The resampling determination unit 109 determines to perform resampling of success data used for particle group optimization when the global best particles change.

The resampling determining unit 109 refers to the global best history and confirms whether the particles of the global best selected by the global best selecting unit 108 are the same as the particles of the previous global best.

If the global best particles selected by the global best selection unit 108 are different from the previous global best particles, the resampling determination unit 109 causes the sampling processing unit 101 to acquire a success data sample. That is, when the global best weight changes during the iteration of the particle swarm optimization method, success data is sampled.

The resampling determination unit 109 determines that when the fitness value (evaluation value) of a specific particle (second particle) in the updated particle group is higher than the fitness value of the global best (first particle), that is, when the global best is If the data has changed, the sampling processing unit 101 is caused to execute a process of selecting a successful data sample (a third plurality of data).

The particle update unit 110 updates the value of each particle other than the global best particle so that it approaches the value of the global best particle updated by the global best selection unit 108.

The particle updating unit 110 may update the value of each particle based on the following formula, for example.

x _i (t) = x _i (t-1) + v _i (t-1)
v _i (t) =c ₀ v _i (t-1) + c ₁ r ₁ (p _l - x _i (t))+c ₂ r ₂ (p _g - x _i (t))
c ₀ , c ₁ , c ₂ are constants, r ₁ , r ₂ are random numbers, p _l is the local best, and p _g is the global best. x _i (t) and v _i (t) represent the position and velocity of the particle group.

The end determination unit 111 determines whether the end conditions of the particle optimization process are satisfied. As described above, the termination conditions are, for example, that the number of iterations (the number of repetitions) reaches a predetermined number, or that the global best particles do not change a certain number of times.

The termination determination unit 111 terminates the particle group optimization process when the termination condition is satisfied. When terminating the particle group optimization process, the termination determination unit 111 reflects the value of the global best particle in the weight of the machine learning model. This generates a modified model.

The training data generation unit 100 generates training data including the success data samples (third plurality of data) selected as described above and the plurality of failure data (second plurality of data).

(B) Operation The processing of the sampling processing unit 101 in the information processing device 1 as an example of the embodiment configured as described above will be described according to the flowchart (steps A1 to A3) shown in FIG.

The sampling of success data may be expressed as pos_sampler(I _pos , M). I _pos represents the set of success data and M represents the machine learning model.

In step A1, the loss calculation unit 102 inputs each success data s _i (s _i εI _pos ) to the target machine learning model M, and calculates the value L(s _i ) of the loss function. The loss calculation unit 102 records the calculated loss function value L(s _i ) in a predetermined storage area of the storage device 13 or the like.

In step A2, the probability distribution conversion unit 103 calculates the probability p(s _i ) for each success data s _i based on the value L(s _i ) of the loss function using the above equation (1). Further, the probability distribution conversion unit 103 generates a success data selection probability distribution.

In step A3, the success data sampling unit 104 selects a success data sample to be used in the particle swarm optimization method from among the plurality of success data based on the success data selection probability distribution. The success data sampling unit 104 selects a predetermined number of success data samples.

Let the set of success data samples (selected success data set) be I _{pos_sampled} . The success data sampling unit 104 initializes I _{pos_sampled} using an empty set (Φ). That is, I _{pos_sampled} ={Φ}.

The success data sampling unit 104 prepares a uniform random number r satisfying 0≦r<1. Furthermore, the success data sampling unit 104 initializes the cumulative total d ₀ to d ₀ =0.

The success data sampling unit 104 sets d _i = d _i-1 + p(s _i ), and selects s _i as a success data sample if d _i-1 ≦r<d _i . That is, the success data sampling unit 104 determines success data having a probability corresponding to the value of the random number r in the success data selection probability distribution as a success data sample.

Further, if s _i has not been selected as a successful data sample so far, the success data sampling unit 104 adds the s _i to I _{pos_sampled} . In other words, I _{pos_sampled} = I _{pos_sampled} ∪{s _i }.

When a certain number of successful data samples have been collected, the successful data sampling unit 104 outputs a set of successful data samples I _{pos_sampled} and ends the process. If the number of successful data samples is insufficient to the predetermined value, the successful data sampling unit 104 prepares a new random number r and repeats the same process.

Next, the processing of the optimization processing unit 105 in the information processing apparatus 1 as an example of the embodiment will be described according to the flowchart (steps B1 to B10) shown in FIG. 6 with reference to FIG. FIG. 7 is a diagram showing a specific example of processing by the optimization processing unit 105.

In step B1, the optimization processing unit 105 causes the sampling processing unit 101 to perform successful data sampling.

In step B2, the particle initialization processing unit 106 initializes each particle in the particle swarm optimization method with a normal distribution.

In step B3, a loop process is started in which the control up to step B10 is repeatedly performed until the iteration (iter) reaches a predetermined value (max_iter).

In step B4, the fitness calculation unit 107 calculates the fitness value (fitness) of each particle.

In the example shown in FIG. 7, the fitness calculation unit 107 calculates fitness values for each of 100 particles x ₁ to x _100. For example, in iteration #t, the fitness value of particle x ₁ is 67.23 and the fitness value for particle x ₁₀₀ is 52.11.

In step B5, the global best selection unit 108 selects the particle with the highest fitness value from among the plurality of particles as the global best p _g . The global best selection unit 108 adds information about the selected global best p _g to the global best history. The global best history is represented by the symbol hist.

In the example shown in FIG. 7, for example, in iteration #t, the fitness value of particle x ₁ is 67.23, the fitness value of particle x ₂ is 70.79, and the fitness value of particle x ₁₀₀ is 52.11. At iteration #t, particle x ₂ has the highest fitness value of 70.79, and global best selection unit 108 selects this particle x ₂ as the global best at iteration #t.

In step B6, the particle updating unit 110 updates the value of each particle other than the global best particle so that it approaches the value of the global best particle updated by the global best selecting unit 108.

That is, based on the parameters of the global best (first particle) with the highest fitness value (evaluation value), the particle updating unit 110 updates each particle (second particle) other than the global best so that the parameter value approaches this parameter value. ) update parameter values.

In step B7, the end determination unit 111 checks whether the global best has been updated for a predetermined iteration. If the global best has not been updated for a predetermined iteration (see YES route in step B7), the termination condition is satisfied and the process is terminated.

On the other hand, if the global best is being updated during a predetermined iteration (see NO route in step B7), the process moves to step B8.

In step B8, the resampling determination unit 109 refers to the global best history to check whether the particles of the global best selected by the global best selection unit 108 are the same as the particles of the previous global best. .

Here, the global best particle in the previous iteration is expressed as hist[iter -1]. The resampling determining unit 109 determines whether p _g == hist[iter -1].

If the global best particles selected by the global best selection unit 108 are different from the global best particles in the previous iteration (see NO route in step B8), the process moves to step B9.

In the example shown in FIG. 7, in iteration #t-1, the fitness value of particle x ₁ is 67.23, the fitness value of particle x ₂ is 53.65, and the fitness value of particle x ₁₀₀ is 49.71. In iteration # t-1, particle x ₁ has the highest fitness value of 67.23, and this particle x ₁ is the global best in iteration # t-1. That is, the global best in iteration # t has particle x ₂ , which is different from the global best in the previous iteration (iteration # t-1), which has particle x ₁ .

At iteration # t, the fitness value of the new global best particle x ₂ of 70.79 is higher than the fitness value of particle x ₁ , which was the global best at iteration # t-1, of 67.23.

In step B9, the optimization processing unit 105 causes the sampling processing unit 101 to perform successful data sampling. That is, when the global best weight changes during the iteration of the particle swarm optimization method, success data is sampled. That is, the success data sample is updated. After that, the process moves to step B10.

Furthermore, as a result of the confirmation in step B8, if the global best particles selected by the global best selection unit 108 are the same as the previous global best particles (see YES route in step B8), step B10 to move to.

In step B10, loop end processing corresponding to step B3 is performed. Here, when the iteration (iter) reaches a predetermined value (max_iter), this flow ends.

(C) Effect As described above, according to the information processing device 1 as an example of the embodiment, the sampling processing unit 101 performs probability distribution transformation based on the value of the loss function of each success data calculated by the loss calculation unit 102. The unit 103 generates a probability distribution (probability distribution) for each of the plurality of success data items to be selected.

Then, the success data sampling unit 104 generates a random number according to the success data selection probability distribution set by the probability distribution conversion unit 103, and determines success data with a probability corresponding to the value of this random number as a success data sample. Specifically, the success data sampling unit 104 performs random selection in which success data with a larger loss is selected with a higher probability.

As a result, by selecting successful data close to the value of the loss function L with a certain degree of randomness, it is possible to perform a search with a gradual trade-off between correction and degradation.

In other words, instead of selecting successful data close to the decision boundary layer as in conventional methods, it is possible to select successful data samples with probabilistic variation from among multiple successful data, thereby suppressing the occurrence of degradation. can do.

Further, in the optimization processing unit 105, the resampling determination unit 109 determines to re-sample the success data used for particle swarm optimization when the global best particle changes in particle swarm optimization.

In particle swarm optimization, when the weights of particles change, the loss of each input success data also changes. In a machine learning model whose parameters are changed according to the weight of the global best particle, by selecting success data that is likely to cause degradation, training can be performed to prevent degradation from occurring with the selected success data. , it is possible to generate a machine learning model that is less prone to degradation. In addition, degradation is less likely to occur in local corrections such as for each label.

FIG. 8 compares a case where a machine learning model is modified by the training processing unit 100 of the information processing device 1 as an example of an embodiment, and a case where a machine learning model is modified by a conventional method. FIG.

In FIG. 8, the accuracy, correction rate, and regression rate are compared and shown, and the symbol P1 indicates the case where the machine learning model is modified by the conventional method, and the symbol P2 indicates the case where the machine learning model is modified by the conventional method, and the symbol P2 is the training of the information processing device 1. The cases in which the machine learning model is modified by the processing unit 100 are shown.

In FIG. 8, an example is shown in which VGG16 is used as the machine learning model and CIFAR10 is used as the dataset. The occurrence of degradation caused by modification of a machine learning model is expressed by the regression rate. As shown in FIG. 8, the regression rate was 6.41% in the conventional method, but when the training processing unit 100 of the information processing device 1 corrects the machine learning model, the regression rate becomes 2.66%. It can be seen that this is becoming less likely to occur.

(D) Others FIG. 9 is a diagram illustrating the hardware configuration of the information processing device 1 as an example of the embodiment.

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

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

Then, when the processor 11 executes the control program (training data generation program 13a) for the information processing device 1, the functions of the sampling processing section 101 and the optimization processing section 105 illustrated in FIG. 1 are realized.

Note that the information processing device 1 executes the sampling processing unit 101 and the optimization processing unit 105 by executing the training data generation program 13a and the OS program, which are programs recorded on a computer-readable non-temporary recording medium, for example. Realize the function as

A program that describes the processing content to be executed by the information processing device 1 can be recorded on various recording media. For example, the training data generation program 13a to be executed by the information processing device 1 can be stored in the storage device 13. The processor 11 loads at least a portion of the training data generation program 13a in the storage device 13 into the memory 12, and executes the loaded program.

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

The memory 12 is a storage memory including ROM (Read Only Memory) and RAM (Random Access Memory). The RAM of the memory 12 is used as a main storage device of the information processing device 1. At least a part of the program to be executed by the processor 11 is temporarily stored in the RAM. The memory 12 also stores various data necessary for processing by the processor 11.

The storage device 13 is a storage device such as a hard disk drive (HDD), an SSD (Solid State Drive), or a storage class memory (SCM), and stores various data. The storage device 13 is used as an auxiliary storage device of the information processing device 1. The storage device 13 stores an OS program, a control program, and various data. The control program includes a training data generation program 13a.

Note that a semiconductor storage device such as an SCM or a flash memory can also be used as the auxiliary storage device. Further, a RAID (Redundant Array of Inexpensive Disks) may be configured using a plurality of storage devices 13.

Further, the storage device 13 may store various data generated when the sampling processing section 101 and the optimization processing section 105 described above execute each process.

A monitor 14a is connected to the graphic processing device 14. The graphic processing device 14 displays an image 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), a liquid crystal display device, and the like.

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 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. Other pointing devices include touch panels, tablets, touch pads, trackballs, and the like.

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-temporary recording medium on which data is recorded in a readable manner by the reflection of light. Examples of the optical disc 16a include a DVD (Digital Versatile Disc), a DVD-RAM, a CD-ROM (Compact Disc Read Only Memory), and a CD-R (Recordable)/RW (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-temporary 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 or reads data from the memory card 17c. The memory card 17c is a card-type non-temporary recording medium.

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

The disclosed technology is not limited to the embodiments described above, and can be implemented with various modifications without departing from the spirit of the present embodiments.

Further, based on the above disclosure, this embodiment can be implemented and manufactured by those skilled in the art.

(E) Additional notes Regarding the above embodiments, the following additional notes are further disclosed.

(Additional note 1)
When input, the machine learning model obtains a first plurality of data that outputs a correct answer and a second plurality of data that outputs an incorrect answer,
from the first plurality of data to the third plurality of data based on the probability according to the magnitude of the value of the loss function when the output of the machine learning model for each of the first plurality of data is input. Select
generating training data including the second plurality of data and the third plurality of data;
A training data generation program characterized by causing a computer to perform processing.

(Additional note 2)
The process of selecting the third plurality of data includes selecting the first plurality of data corresponding to a probability range to which the value of the random number corresponds in a probability distribution set according to the magnitude of the value of the loss function. , the training data generation program according to supplementary note 1, characterized in that the training data generation program includes a process of selecting the third plurality of data.

(Additional note 3)
In a particle group having a plurality of particles including a combination of a plurality of parameter values, the parameter value of a second particle other than the first particle is updated based on the parameter value of the first particle with the highest evaluation value. death,
If the updated evaluation value of the second particle is higher than the evaluation value of the first particle, the computer is caused to execute a process of selecting the third plurality of data. , the training data generation program according to appendix 1 or 2.

(Additional note 4)
causing the computer to execute a process of updating some parameters included in the machine learning model using parameter values of a particle with the highest evaluation value among the first particles and the second particles; The training data generation program according to appendix 3, characterized by:

(Appendix 5)
When input, the machine learning model obtains a first plurality of data that outputs a correct answer and a second plurality of data that outputs an incorrect answer,
from the first plurality of data to the third plurality of data based on the probability according to the magnitude of the value of the loss function when the output of the machine learning model for each of the first plurality of data is input. Select
generating training data including the second plurality of data and the third plurality of data;
A training data generation method characterized in that processing is performed by a computer.

(Appendix 6)
The process of selecting the third plurality of data includes selecting the first plurality of data corresponding to a probability range to which the value of the random number corresponds in a probability distribution set according to the magnitude of the value of the loss function. , the training data generation method according to appendix 5, characterized in that the training data generation method includes a process of selecting the third plurality of data.

(Appendix 7)
In a particle group having a plurality of particles including a combination of a plurality of parameter values, the parameter value of a second particle other than the first particle is updated based on the parameter value of the first particle with the highest evaluation value. death,
The computer is characterized in that the computer executes a process of selecting the third plurality of data when the updated evaluation value of the second particle is higher than the evaluation value of the first particle. , the training data generation method according to appendix 5 or 6.

(Appendix 8)
The computer executes a process of updating some parameters included in the machine learning model using parameter values of a particle with the highest evaluation value among the first particles and the second particles. The training data generation method according to appendix 7, characterized by:

(Appendix 9)
When input, the machine learning model obtains a first plurality of data that outputs a correct answer and a second plurality of data that outputs an incorrect answer,
from the first plurality of data to the third plurality of data based on the probability according to the magnitude of the value of the loss function when the output of the machine learning model for each of the first plurality of data is input. Select
generating training data including the second plurality of data and the third plurality of data;
An information processing device characterized by including a control unit that executes processing.

(Appendix 10)
The process of selecting the third plurality of data includes selecting the first plurality of data corresponding to a probability range to which the value of the random number corresponds in a probability distribution set according to the magnitude of the value of the loss function. , the information processing apparatus according to appendix 9, further comprising a process of selecting the data as the third plurality of data.

(Appendix 11)
The control section,
In a particle group having a plurality of particles including a combination of a plurality of parameter values, the parameter value of a second particle other than the first particle is updated based on the parameter value of the first particle with the highest evaluation value. death,
Supplementary note 9, characterized in that when the updated evaluation value of the second particle is higher than the evaluation value of the first particle, a process of selecting the third plurality of data is executed. or the information processing device according to item 10.

(Appendix 12)
The control section,
A process of updating some parameters included in the machine learning model using a parameter value of a particle having the highest evaluation value among the first particle and the second particle is executed. The information processing device according to supplementary note 11.

1 Information processing device 11 Processor (control unit)
12 Memory 13 Storage device 13a Training data generation program 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 Training Processing Unit 101 Sampling Processing Unit 102 Loss Calculation Unit 103 Probability Distribution Conversion Unit 104 Successful Data Sampling Unit 105 Optimization Processing Unit 106 Particle Initialization Processing Unit 107 Fitness Calculation Unit 108 Global Best Selection Unit 109 Resampling Judgment Unit 110 Particle Update Section 111 Completion Judgment Section

Claims (6)

When input, the machine learning model obtains a first plurality of data that outputs a correct answer and a second plurality of data that outputs an incorrect answer,
from the first plurality of data to the third plurality of data based on the probability according to the magnitude of the value of the loss function when the output of the machine learning model for each of the first plurality of data is input. Select
generating training data including the second plurality of data and the third plurality of data;
A training data generation program characterized by causing a computer to perform processing. The process of selecting the third plurality of data includes selecting the first plurality of data corresponding to a probability range to which the value of the random number corresponds in a probability distribution set according to the magnitude of the value of the loss function. The training data generation program according to claim 1, further comprising a process of selecting the third plurality of data as the third plurality of data. In a particle group having a plurality of particles including a combination of a plurality of parameter values, the parameter value of a second particle other than the first particle is updated based on the parameter value of the first particle with the highest evaluation value. death,
If the updated evaluation value of the second particle is higher than the evaluation value of the first particle, the computer is caused to execute a process of selecting the third plurality of data. , The training data generation program according to claim 1 or 2. causing the computer to execute a process of updating some parameters included in the machine learning model using parameter values of a particle with the highest evaluation value among the first particles and the second particles; The training data generation program according to claim 3, characterized in that: When input, the machine learning model obtains a first plurality of data that outputs a correct answer and a second plurality of data that outputs an incorrect answer,
from the first plurality of data to the third plurality of data based on the probability according to the magnitude of the value of the loss function when the output of the machine learning model for each of the first plurality of data is input. Select
generating training data including the second plurality of data and the third plurality of data;
A training data generation method characterized in that processing is performed by a computer. When input, the machine learning model obtains a first plurality of data that outputs a correct answer and a second plurality of data that outputs an incorrect answer,
from the first plurality of data to the third plurality of data based on the probability according to the magnitude of the value of the loss function when the output of the machine learning model for each of the first plurality of data is input. Select
generating training data including the second plurality of data and the third plurality of data;
An information processing device characterized by including a control unit that executes processing.

Images