Information processing device, terminal device, server, information processing method, and information processing program

Abstract

This information processing device reads the parameter learning process in a model 1, said parameter learning process using a stochastic gradient descent method that uses a data set D, and also reads the initial parameter values of a model 2. The information processing device then finds a permutation transformation that minimizes the distance between the result of permutation transformation of the parameter learning process in the model 1, and the parameter learning process performed in the model 2 when the model 2 is trained using the data set D and the initial parameter values of the model 2. The information processing device then outputs parameter values for the final process of the parameter learning process in the model 2, said parameter values being calculated using the found permutation transformation.

Description

Information processing device, terminal device, server, information processing method, and information processing program

[0001] The present invention relates to an information processing device, a terminal device, a server, an information processing method, and an information processing program for training a machine learning model.

[0002] In deep learning, the computational cost during training is one of the major issues. The computational cost during training increases with the model size and the number of training data.

[0003] One example of a technique for reducing the computational cost during training is transfer learning. Transfer learning is a technique that shortens the training time and improves accuracy for dataset B by using the parameters of model A trained on dataset A as the initial values ​​for training dataset B, which is similar to dataset A.

[0004] Here, model training may involve multiple training sessions using the same dataset. For example, to perform model ensembling, the same dataset is trained using different random initial values ​​for each model. Also, if a pre-trained model on which a model is based is updated after fine-tuning it, the model must be fine-tuned again using the dataset used for fine-tuning.

[0005] On Warm-Starting Neural Network Training, [online], [Retrieved June 5, 2023], Internet <URL: https: / / arxiv.org / abs / 1910.08475> GIT RE-BASIN: MERGING MODELS MODULO PERMUTATION SYMMETRIES, [online], [Retrieved June 5, 2023], Internet <URL: https: / / arxiv.org / abs / 2209.04836>

[0006] As described above, when learning is performed multiple times using the same dataset, each learning process is likely to be similar, but no technology has been proposed to reduce the computational cost of learning.

[0007] Here, in order to reduce the computational cost in training using the same dataset, it is possible to apply the above-mentioned transfer learning (see Non-Patent Document 1). For example, it is possible to train model 2 on dataset D using the parameters of model 1 trained with dataset D as initial values. However, it is known that the above training method results in a deterioration in model accuracy compared to when a model is trained from random initial values.

[0008] Therefore, an object of the present invention is to solve the above-mentioned problems and reduce the calculation cost when similar learning is performed using the same data set.

[0009] In order to solve the above-mentioned problems, the present invention is characterized by comprising a reading unit that reads a parameter learning process performed by stochastic gradient descent using a dataset in a first machine learning model and initial parameter values ​​of a second machine learning model; a calculation unit that calculates a permutation transformation that minimizes the distance between the result of permutation transformation of the parameter learning process in the first machine learning model and the initial parameter values ​​of the second machine learning model and the parameter learning process when the second machine learning model is trained using the dataset; and an output processing unit that outputs parameter values ​​of the final process in the parameter learning process of the second machine learning model calculated by the permutation transformation.

[0010] According to the present invention, it is possible to reduce the calculation cost when similar learning is performed using the same data set.

[0011] FIG. 1 is a conceptual diagram of permutation transformation executed by an information processing apparatus according to a first embodiment. FIG. 2 is a diagram illustrating an example of the configuration of the information processing apparatus according to the first embodiment. FIG. 3 is a flowchart illustrating an outline of processing procedures executed by the information processing apparatus according to the first embodiment. FIG. 4 is a diagram illustrating an example of an algorithm used by the information processing apparatus according to the first embodiment to execute processing. FIG. 5 is a flowchart illustrating the algorithm illustrated in FIG. 4. FIG. 6 is a conceptual diagram of permutation transformation executed by an information processing apparatus according to a second embodiment. FIG. 7 is a diagram illustrating an example of the configuration of the information processing apparatus according to the second embodiment. FIG. 8 is a flowchart illustrating an example of processing procedures executed by the information processing apparatus according to the second embodiment. FIG. 9 is a diagram illustrating an outline of an information processing apparatus according to a third embodiment. FIG. 10 is a diagram illustrating an example of the configuration of the information processing apparatus according to the third embodiment. FIG. 11 is a flowchart illustrating processing procedures executed by the information processing apparatus according to the third embodiment. FIG. 12 is a flowchart illustrating processing procedures executed by the information processing apparatus according to the third embodiment. FIG. 13 is a diagram illustrating an application example 1 of learning transfer executed by the information processing apparatus according to the third embodiment. FIG. 14 is a diagram illustrating an application example 2 of learning transfer executed by the information processing apparatus according to the third embodiment. Fig. 15 is a diagram for explaining application example 3 of learning transfer performed by the information processing apparatus of the third embodiment. Fig. 16 is a diagram illustrating an example of the configuration of a computer that executes an information processing program.

[0012] Hereinafter, with reference to the drawings, a first to third embodiment of the present invention will be described as an example of carrying out the present invention. The present invention is not limited to each of the embodiments.

[0013] First Embodiment Overview First, an overview of an information processing apparatus according to a first embodiment will be described with reference to Fig. 1. The information processing apparatus performs learning of a machine learning model using, for example, a DNN (Deep Neural Network).

[0014] For example, the initial parameter θ1 of the DNN t=0 On the other hand, the learning process of dataset D by SGD (Stochastic Gradient Descent) (θ1 t=0 ,…,θ1 t=T) is given. Here, the information processing device calculates a new initial parameter (initial parameter value) θ2 t=0 When learning a dataset D, the learning process of the given θ1 (θ1 t=0 ,…,θ1 t=T ) is permuted to obtain the learning process of θ2 (θ2 t=0 ,…,θ2 t=T ) is obtained. Then, the information processing device uses the obtained permutation transformation π to obtain the initial parameter θ2 t=0 The learning process of the dataset D for (θ2 t=0 ,…,θ2 t=T ) is calculated. This gives (θ2 t=1 ,…,θ2 t=T ) to reduce the calculation time.

[0015] In other words, the initial parameter θ1 t=0 The learning process of the dataset D for (θ1 t=0 ,…,θ1 t=T ) is considered to contain information useful for learning the dataset D. Therefore, the information processing device utilizes the above information to generate new initial parameters θ2 t=0 This allows the information processing device to efficiently learn the data set D for the initial parameter θ2 t=0 The process of learning the dataset D (θ2 t=1 ,…,θ2 t=T ) to reduce the calculation time.

[0016] [Permutation Symmetry of DNN] Next, the basic concept of learning performed by the information processing device of this embodiment will be described. As a general property of DNN, NN f(x; W1, ..., W L ) has degrees of freedom of permutation symmetry. For example, any permutation matrix σ1,…,σ L-1 The relationship shown in the following equation (1) holds for the DNN parameters θ = (W1, ..., W L ) for the permutation transformation π=(σ1,…,σ L-1The result of applying (the right side of equation (1)) is written as πθ, and this is called a permutation transformation of θ by π.

[0017]

[0018] The permutation matrix is, for example, a permutation matrix as shown in Equation 2. By using the permutation matrix as described above, the information processing device can handle permutations of intermediate dimensions in the NN.

[0019]

[0020] In this embodiment, it is assumed that the learning process of the parameters θ1 and θ2 approximately matches with a certain permutation transformation. In other words, the permutation transformation π=(σ1,...,σ2) satisfies the relationship shown in the following equation (3) for any t. L-1 ) exists.

[0021]

[0022] If a value of π that satisfies the relationship shown in equation (3) is found, the learning process for parameter θ2 can be obtained inductively from the learning process for parameter θ1. In other words, the learning process for parameter θ2 can be obtained by the following equation (4).

[0023]

[0024] Therefore, the information processing device of this embodiment obtains a permutation transformation π that satisfies the relationship of the above formula (3) for any t.

[0025] [Formulation of the Proposed Technology] Based on the above discussion, the desired permutation transformation π can be calculated by solving the optimization problem P T can be considered as a solution to

[0026] Given: θ1 t=0 ,…,θ1 t=T (learning process of θ1) and θ2 t=0 (initial value of θ2)

[0027]

[0028] In addition, || ||2 in equation (5) 2 is the vector L 2 It is the norm.

[0029] However, θ2 in equation (5) t is (θ2 t=0 Since the parameters are unknown (except for T is an incomplete formulation. Therefore, the information processing device calculates each difference θ i t -θ i t-1 is the gradient -∇θ of the loss function L for the dataset D i t-1 Optimization problem P approximated by L T This solves the optimization problem P T can finally be formulated as the following equation (6).

[0030]

[0031] The optimization problem P shown in Equation (6) T is a nested structure (θ 2,π t also contains the unknown variable π), it is difficult to solve directly and requires some ingenuity.

[0032] Therefore, the information processing device solves the above optimization problem P T Instead of T Solve them in order and inductively find the solutions π1, ..., π T-1 , π T = :π where each optimization problem P´ t is P´ t-1 Solution of π t-1 is inductively defined as in the following equation (7).

[0033]

[0034] Each optimization problem P' in the above equation (7) t is a simple quadratic optimization problem without nesting (where clause), and permutation transformation does not change the norm of the parameter vector. tcan be reduced to a linear optimization problem. Therefore, the information processing device can use an existing linear optimization solver to find the solution π t Therefore, the information processing device can calculate the solutions π1, ..., π T-1 , π T By calculating in order, the final value obtained is π T is the optimization problem P T converges to the solution.

[0035] [Configuration Example] Next, a configuration example of the information processing device 10 will be described with reference to Fig. 2. The information processing device 10 includes, for example, an input / output unit 11, a storage unit 12, and a control unit 13.

[0036] The input / output unit 11 is an interface that controls the input and output of various data. For example, the input / output unit 11 controls the learning process of the parameter θ1 of the model 1 (θ1 t=0 ,…,θ1 t=T ), the initial value of the parameter θ2 of model 2 0 , dataset D, etc. are accepted as input.

[0037] The storage unit 12 stores data, programs, etc. that are referenced when the control unit 13 executes various processes. The storage unit 12 is realized by a semiconductor memory element such as a random access memory (RAM) or a flash memory, or a storage device such as a hard disk or an optical disk.

[0038] For example, the memory unit 12 stores the learning process of the parameter θ1 (θ1 t=0 ,…,θ1 t=T ), the initial value of parameter θ2 t=0 , data set D, etc. The learning process of parameter θ1 (θ1 t=0 ,…,θ1 t=T ) is the initial value of the parameter θ1 of Model 1 (θ1 t=0 ) and dataset D.

[0039] The control unit 13 is responsible for overall control of the information processing device 10. The functions of the control unit 13 are realized, for example, by a CPU (Central Processing Unit) executing a program stored in the storage unit 12.

[0040] The control unit 13 includes, for example, an input receiving unit 131, a reading unit 132, a calculation unit 133, and an output processing unit 134. An information generating unit 135 indicated by a dashed line may or may not be provided, and the cases where it is provided will be described later.

[0041] The input receiving unit 131 is a learning process of the parameter θ1 (θ1 t=0 ,…,θ1 t=T ), the initial value of parameter θ2 t=0 , the input of the data set D is accepted and stored in the storage unit 12 .

[0042] The reading unit 132 reads the learning process of the parameter θ1 (θ1 t=0 ,…,θ1 t=T ), the initial value of parameter θ2 t=0 Load.

[0043] The calculation unit 133 performs the learning process of the parameter value θ1 of the model 1 (θ1 t=0 ,…,θ1 t=T ) and the initial parameter value θ2 of Model 2 0 and the learning process of the parameter θ2 using the data set D (θ2 t=0 ,…,θ2 t=T ) to find the permutation transformation that minimizes the distance.

[0044] Here, the calculation unit 133 performs the learning process of the parameter θ1 (θ1 t=0 ,…,θ1 t=T ), and the learning process of parameter θ2 (θ2 t=0 ,…,θ2 t=T ) each difference θ i t -θ i t-1 is the gradient value (-∇θ i t-1L) to find a permutation transformation that minimizes the distance (optimization problem P T (Solve).

[0045] In addition, the calculation unit 133 performs a learning process of the parameter θ1 (θ1 t=0 ,…,θ1 t=T ) and the learning process of the parameter θ2 (θ2 t=0 ,…,θ2 t=T ) is executed in the order of the learning process.

[0046] For example, when the number of learning processes of the parameter θ is T, the calculation unit 133 compares and minimizes partial learning processes up to the tth one, where t≦T, instead of simultaneously comparing all processes and minimizing the distance (the distance of the entire learning process). t is a permutation transformation π that minimizes the distance of the partial learning process up to the t-1th iteration. t-1 The calculation unit 133 repeats this process inductively from t=1 to t=T, and finally obtains the permutation transformation π T By this, the distance of the whole learning process is minimized.

[0047] Then, the calculation unit 133 calculates the learning process of the parameter θ2 (θ2 t=0 ,…,θ2 t=T )

[0048] The output processing unit 134 performs the learning process (θ2 t=0 ,…,θ2 t=T ) the value of the parameter of the final process (θ2 t=T ) is output.

[0049] [Example of Processing Procedure] Next, an example of processing procedure executed by the information processing device 10 will be described with reference to FIG.

[0050] First, the reading unit 132 of the information processing device 10 reads the learning process of Model 1 (θ t=0 ,…,θ1 t=T ) (S1). The reading unit 132 also reads the parameter initial value θ2 of model 2. t=0 Next, the calculation unit 133 calculates π (Permutation transformation) that satisfies the relationship shown in the above-mentioned formula (3) for any t (S3).

[0051] After S3, the calculation unit 133 calculates the learning process (θ t=0 ,…,θ2 t=T ) (S4). After that, the output processing unit 134 returns the value (θ t=T ) is output (S5).

[0052] [Example of Algorithm] An example of an algorithm used by the information processing device when executing the above process is shown in FIG. i t=s θ i s Also, π t is a permutation transform. The algorithm shown in FIG. 4 will be explained using the flowchart shown in FIG.

[0053] First, the reading unit 132 of the information processing device 10 loads the learning process of Model 1 from the storage unit 12. For example, the reading unit 132 reads the learning process of Model 1 with length T (θ1 t=0 ,…,θ1 t=T ) of model 2 from the storage unit 12 (S11). t=0 is read (S12).

[0054] Next, the calculation unit 133 calculates a variable (θ t=0 ,…,θ2 t=T ) are all initialized to zero vectors (S13). Then, a variable π is prepared to store the "permutation transformation for exchanging the components of each intermediate layer of the NN" (S14). Furthermore, the calculation unit 133 initializes a loop index t to 0 (S15).

[0055] After S15, if t=T is not true (No in S16), the calculation unit 133 sets the variable g1 1 ,…,g1 t and g2 1 ,…,g2 t (S17), and initialize the loop index s to 0 (S18). After that, if s=t is not true (No in S19), the calculation unit 133 calculates the mini-batch gradient (S20). For example, the calculation unit 133 calculates g1 s parameter θ1 s-1 Substituting the mini-batch gradient at g2 s parameter θ2 s-1 Then, the calculation unit 133 adds 1 to s (S21) and returns to S19.

[0056] On the other hand, if s=t in S19 (Yes in S19), the calculation unit 133 performs optimization of the subproblem (S22). That is, the calculation unit 133 solves a permutation transformation π that minimizes the sum of squares of the L2 distance of the mini-batch gradient shown in the following equation (8) using a quadratic optimization or linear optimization solver, and substitutes the result for the variable π.

[0057]

[0058] After S22, the calculation unit 133 initializes the loop index s to 0 (S23). If s=t is not satisfied (No in S24), the calculation unit 133 updates the learning process of model 2 (S25). That is, the calculation unit 133 updates the parameter θ s θ2 s-1 +π(θ1 s -θ1 s-1 After that, the calculation unit 133 adds 1 to s (S26) and returns to S24.

[0059] On the other hand, if s=t in S24 (Yes in S24), the calculation unit 133 adds 1 to t (S27) and returns to S16. In S16, if t=T (Yes in S16), the calculation unit 133 calculates the learning process of model 2 (θ ｔ=0 ,…,θ2 ｔ=T) (S28). After that, the output processing unit 134 returns the learning process (θ ｔ=0 ,…,θ2 ｔ=T ) among the parameters of the final process (θ2 ｔ=T ) is output.

[0060] By executing the above process, the information processing device 10 obtains the learned parameters of model 2 (θ ｔ=T ) can be calculated.

[0061] [Modification of the First Embodiment] The information processing device 10 stores in the storage unit 12 the learning process (θ1 t=0 ,…,θ1 t=T ), instead of holding two parameters (e.g., parameter initial value θ1 t=0 , the parameter value θ1 in the final learning process t=T ) may be held.

[0062] In this case, the information processing device 10 includes an information generating unit 135 (see FIG. 2). The information generating unit 135 generates an initial value θ1 of the parameter θ1 of the model 1 stored in the storage unit 12. t=0 and the final process value of parameter θ1 t=T From the above, the intermediate learning process of Model 1 (θ1 t=1 ,…,θ1 t=T-1 For example, the information generating unit 135 linearly interpolates θ t By calculating the following equation (9), the learning process of model 1 with length T (θ1 t=0 ,…,θ1 t=T ) is generated.

[0063]

[0064] This allows the information processing device 10 to reduce the number of parameters θ1 stored in the storage unit 12, thereby reducing storage costs.

[0065] Furthermore, when the calculation unit 133 of the information processing device 10 calculates the mini-batch gradient in S20 of FIG. 5, g1 s , g2 s If you calculate each time, the total is O(T 2It is necessary to perform backward calculation. Here, among s = 0, …, t - 1, when s < t - 1, for the previously calculated g1 s , g2 s , it is considered that they can be approximated. Therefore, when s < t - 1, the calculation unit 133 may reuse the caches of the previous g1 s , g2 s .

[0066] That is, the calculation unit 133 calculates the mini - batch gradient only when s = t - 1, and when s < t - 1, it reuses the previously calculated mini - batch gradient (for example, the initially calculated g1 s , g2 s ). Thereby, the calculation cost when the calculation unit 133 calculates the mini - batch gradient in S20 of FIG. 5 can be reduced.

[0067] That is, in the algorithm shown in FIG. 4, the first mini - batch gradient (∇θ = θ1 s-1 L(f(x;θ)) does not depend on the outer loop variable t. Therefore, by caching and reusing the mini - batch gradient (∇θ = θ1 s-1 L(f(x;θ)) once calculated by the calculation unit 133, the recalculation of the mini - batch gradient (∇θ = θ1 s-1 L(f(x;θ)) can be skipped.

[0068] Also, in the algorithm shown in FIG. 4, the second mini - batch gradient (∇θ = θ2 s-1 L(f(x;θ)) is a value that changes depending on π t , so it cannot be cached and reused in principle. However, in practice, the approximation formula shown in the following formula (10) holds. Therefore, the value of the mini - batch gradient at a certain t and the value of the mini - batch gradient at another t are approximately equal.

[0069]

[0070] Therefore, even if the calculation unit 133 caches and re - uses the value of the mini - batch gradient (∇θ = θ2 s-1 L(f(x;θ)) once calculated, an approximately same result can be obtained.

[0071] [Second embodiment] [Overview] Next, an overview of an information processing device according to a second embodiment will be described with reference to FIG. 6. The algorithm described in the first embodiment calculates the gradient ∇θ2 t L) ≒ π(the gradient of the t-th step of the learning process of Model 1 ∇θ1 t This is an optimization algorithm that assumes the existence of a π that simultaneously satisfies T conditions (T = 1, L = 1, L = 2 ...1, L = 2, L = 1, L = 1, L = 1, L = 1, L = 1, L = 1, L = 1, L = 1, L = 1, L = 1, L = 1, L = 1, L = 1, L = 1, L = 1, L = 1, L = 1, L = 1, L = 1, L = 1, L = 1, L = 1, L = 1, L = 1, L = 1, L = 1, L = 1, L = 1, L = 1, L = 1, L = 1, L = 1, L = 1, L = 1, L = 1, L = 1, L = 1, L = 1, L = 1, L =

[0072] Therefore, the information processing device of the second embodiment divides the entire learning process of the parameters of model 1 into a plurality of (for example, N) segments, as shown in Fig. 6. Then, the information processing device calculates a permutation transformation π corresponding to each segment. p (p = 1, ..., N) to transform (transfer) the learning process of the parameters of each segment of Model 1 to the entire learning process of the parameters of Model 2. p Find (p = 1, ..., N).

[0073] This allows the information processing device to improve the expressive power of the permutation transformation that converts the learning process of the parameters of model 1 into the learning process of the parameters of model 2. As a result, even if the length T of the learning process of the parameters of model 1 is large, the information processing device can easily obtain the permutation transformation that converts the learning process of the parameters of model 1 into the learning process of the parameters of model 2.

[0074] [Transfer of the learning process by multiple permutation transformations] For example, the initial parameter θ1 of the DNN t=0 On the other hand, the learning process of dataset D by SGD (θ1 t=0 ,…,θ1 t=T ) is given. Assume that T is a multiple of integer N. The new initial parameter θ2 t=0Given N permutation transformations π1,…,π N The learning process of Model 2 using (θ2 ｔ=0 ,…,θ2 ｔ=T ) is defined as the following equation (11).

[0075]

[0076] Here, 1≦p≦N is the index of the segment. Also, 1≦q≦T / N is the index of the learning process within each segment. Furthermore, p(T / N) is the T p Corresponds to.

[0077] [Configuration Example] Next, a configuration example of an information processing device 10a according to a second embodiment will be described with reference to Fig. 7. The same components as those in the first embodiment are denoted by the same reference numerals, and their description will be omitted. The information processing device 10a includes a dividing unit 136 and a calculation unit 133a. The dividing unit 136 divides the learning process of the parameter θ1 of model 1 into N segments.

[0078] The calculation unit 133a calculates, for each of the segments, a permutation transformation π that minimizes the distance between the result of permutation transformation of the learning process of the parameters in model 1 and the learning process of the parameters in model 2 corresponding to that segment. p Ask for.

[0079] Then, the calculation unit 133a calculates a permutation transformation π for each segment. p (p=1, ..., N) to calculate the parameter values ​​of the final process in the learning process of the parameters of Model 2. For example, the calculation unit 133a performs a permutation transformation π p (p=1, . . . , N) to calculate the parameter values ​​of the final step in the parameter learning process of Model 2.

[0080] [π1, ..., π N Here, for example, for p = 1, ..., N, the learning process of the parameters of Model 1 (θ1 ｔ=0 , …, θ1ｔ=p（T / N） ) by applying the algorithm described in the first embodiment, and p There is also a way to find it.

[0081] However, in the above method, π p The optimization problem used to find π p-1 This contains the optimization problem used to find π p-1 Conversion result and π p However, as p increases, the size of the optimization problem increases monotonically. p The disadvantage is that the calculation cost is high.

[0082] In addition, for p = 1, ..., N, the learning process of the parameters of Model 1 (θ1t = (p-1)T / N, ..., θ1 ｔ=p（T / N） ) by applying the algorithm described in the first embodiment, and p It is also possible to ask for

[0083] According to the above method, π p This has the advantage that the calculation cost of π can be kept constant. p-1 Conversion result and π p Therefore, we cannot guarantee the similarity of the conversion results with π p-1 Conversion result and π p However, there is a disadvantage in that there may be a large discrepancy between the conversion results by

[0084] Therefore, the calculation unit 133a calculates the learning process of the parameters of model 1 for p = 1, ..., N (θ1t = (p-1)T / N, ..., θ1 ｔ=p（T / N） ) to the objective function (see equation (12)) to which the algorithm described in the first embodiment is applied, and g(π p , π p-1 ) is added to the objective function (see equation (13)) p Ask for.

[0085]

[0086] g (π p , π p-1) is the permutation transformation π of the segment (p) p The transformation result by and the permutation transformation π of the segment just before the segment (p-1) p-1 For example, the calculation unit 133a calculates f~(π p ) to minimize π p However, when p=1, π p-1 Since (π0) is not defined, g(π1, π0):=0. The calculation unit 133a calculates g(π p , π p-1 ) is calculated, for example, as follows:

[0087] (1) g(π p , π p-1 ) Calculation Method 1 For example, the calculation unit 133a calculates π p The transformation result of the p-th segment by (see equation (14)), and π p-1 The Euclidean distance between the transformation result of the p-th segment by (see equation (15)) and the transformation result is expressed as g(π p , π p-1 ) (see equation (16)).

[0088]

[0089] (2) g(π p , π p-1 ) Calculation Method 2 For example, the calculation unit 133a calculates π p The transformation result of the p-1-th segment by (see equation (17)), and π p-1 The Euclidean distance between the transformation result of the p-1th segment by (see equation (18)) and the resulting value is expressed as g(π p , π p-1 ) (see equation (19)).

[0090]

[0091] (3) g(π p , π p-1 ) Calculation Method 3 For example, the calculation unit 133a calculates π p-1The result of transforming the mini-batch gradient by p-1 ∇θ1 t L and π p The transformation result of the gradient by p ∇θ1 t Calculate the Euclidean distance of L and use the value as g(π p , π p-1 ) (see equation (20)).

[0092]

[0093] (4) g(π p , π p-1 ) Calculation Method 4 For example, the calculation unit 133a calculates π p-1 and π p Calculate the Euclidean distance as a matrix between and g(π p , π p-1 ) (see equation (21)).

[0094]

[0095] The calculation unit 133a adds the above g(π p , π p-1 ) is added to the objective function (see equation (13)). p While keeping the computational cost of π constant, p-1 and π p This can minimize the deviation of the conversion results during the learning process.

[0096] [Example of Processing Procedure] Next, an example of processing procedure executed by the information processing device 10a will be described with reference to FIG.

[0097] First, the reading unit 132 of the information processing device 10a reads the learning process of Model 1 (θ t=0 ,…,θ1 t=T ) (S31). The reading unit 132 also reads the parameter initial value θ2 of model 2. t=0 Next, the dividing unit 136 divides the learning process read in S31 into N segments (S32). After that, the calculation unit 133a prepares a loop variable i=1 for the learning process of the N segments (S33).

[0098] The calculation unit 133a then selects the learning process of the i-th segment (S34), and if i==1 (Yes in S35), calculates the permutation transformation π1 that transfers the learning process of the first segment using the same algorithm as described in the first embodiment (S36).Then, the process proceeds to S38.

[0099] On the other hand, if i==1 is not true (No in S35), the calculation unit 133a calculates the permutation transformation π for transferring the learning process of the i-th segment using, for example, Equation (13). i (S37). After that, the calculation unit 133a calculates the obtained permutation transformation π i is stored in the storage unit 12 (S38).

[0100] After S38, if i==N is not true (No in S39), the calculation unit 133a increments the value of i (S40: i=i+1) and returns to S34. On the other hand, if i==N is true after S38 (Yes in S39), the calculation unit 133a calculates N permutation transformations π1,...,π for transferring the learning processes of the N segments, respectively. N (S41). After that, the calculation unit 133a outputs the permutation transformations π1, ..., π N and the initial parameter value θ2 t=0 Using the dataset D, the learning process of the parameters of model 2 (θ2 t=0 ,…,θ2 t=T ) and calculate the final value of the parameter learning of Model 2 (θ2 t=T ) is output.

[0101] By performing the above processing, the information processing device 10a can easily obtain a permutation transformation that converts the learning process of the parameters of Model 1 into the learning process of the parameters of Model 2, even if the length T of the learning process of the parameters of Model 1 is large.

[0102] [Application Example] According to the information processing apparatuses 10 and 10a described above, it is possible to reduce the calculation cost when similar learning is performed using the same data set D.

[0103] For example, when performing model ensembling, the computational cost can be reduced when training the same dataset using different random initial values ​​for each model.

[0104] In addition, for example, when a pre-trained model that is the basis for a model obtained by fine-tuning a pre-trained model is updated, the computational cost of re-fine-tuning the model can be reduced.

[0105] Furthermore, model compression using knowledge distillation, for example, can reduce the computational cost of retraining a smaller model after training a dataset with one model.

[0106] [Third Embodiment] [Overview] An overview of an information processing device according to a third embodiment will be described with reference to Fig. 9 . For example, the information processing device according to the third embodiment trains a model M1 using a data set D1 with a model P1 as an initial model. At this time, the information processing device uses the learning direction or learning process used to train the model M1 to obtain a model (model M2) trained on a data set D2 (D2 may be equal to D1) for the model P2. The model P2 may be, for example, a base model with higher performance than the model P1, a random initial model with a larger number of parameters than the model P1, or the like.

[0107] The learning direction is the difference between the parameters of the model during model learning. For example, the learning direction is the difference between the final parameters obtained by model learning and the initial parameters (θ n -θ1). The learning process is a sequence of intermediate results of the model parameters obtained during model learning (θ 1, …,θ n ) where θ1 corresponds to the initial model P1 or P2, and θ n are parameters corresponding to the final model M1 or M2.

[0108] Here, the information processing device obtains (optimizes) a transformation T such that applying transformation T to the learning direction or learning process from model P1 to model M1 results in a learning direction or learning process from model P2 to model M2. The information processing device then obtains model M2 by updating the parameters of model P2 using the obtained transformation T and the result of applying transformation T to the learning direction or learning process from model P1 to model M1.

[0109] In other words, in the information processing devices of the above-mentioned embodiments, when obtaining model M2, a learning process from model P1 to model M1 was used, but in the information processing device of the third embodiment, model P1 and the learned model M1 are used instead of the above-mentioned learning process.

[0110] The inputs and outputs to the information processing device are summarized below. The inputs to the information processing device are model P1, model M1 obtained by training dataset D1 using model P1 as an initial model, model P2 which is a new initial model, and dataset D2. Dataset D2 is, for example, the same as or similar to dataset D1.

[0111] The model P1 is, for example, a base model, a model in which random initial parameters are set (random initial model), etc. The model M1 is, for example, a base model trained from random initial values, a specialized model specialized for a specific domain or a specific task, etc.

[0112] Based on the above input, the information processing device outputs a model M2 that has the same number of parameters as model P2 and knowledge of dataset D1. Note that this model M2 is expected to combine the expertise of model M1 and the properties (e.g., performance, number of parameters) of model P2.

[0113] [Configuration Example] Next, a configuration example of an information processing device 10b according to a third embodiment will be described with reference to Fig. 10. The same components as those in the above-described embodiments are denoted by the same reference numerals, and their description will be omitted. The information processing device 10b includes, for example, an input / output unit 11b, a storage unit 12b, and a control unit 13b.

[0114] The input / output unit 11b is an interface that controls the input and output of various data, and receives inputs such as model P1 (parameters of model P1), model M1 (parameters of model M1), model P2 (parameters of model P2), and data set D2.

[0115] The storage unit 12b stores data, programs, etc. that are referenced when the control unit 13b executes various processes. For example, the storage unit 12b stores the parameters of the model P1, the parameters of the model M1, the parameters of the model P2, the data set D2, etc. that are received by the input / output unit 11b.

[0116] The control unit 13b controls the entire information processing device 10b. The functions of the control unit 13b are realized, for example, by the CPU executing a program stored in the storage unit 12b.

[0117] The control unit 13 includes, for example, an input receiving unit 131b, a reading unit 132b, a calculation unit 133b, and an output processing unit 134b. An additional learning unit 137 indicated by a dashed line may or may not be provided, and the cases where it is provided will be described later.

[0118] The input receiving unit 131b receives inputs such as parameters of the model P1, parameters of the model M1, parameters of the model P2, and the data set D2, and stores them in the storage unit 12b.

[0119] The reading unit 132b reads the parameters of the model P1, the parameters of the model M1, the parameters of the model P2, and the data set D2 from the storage unit 12b.

[0120] The calculation unit 133b uses the loaded parameters of model P1, model M1, model P2, and dataset D2 to determine the parameter values ​​of the final stage of the learning process of model P2 (final parameter values; parameter values ​​of model M2).

[0121] The calculation unit 133b includes a transformation calculation unit 1331 and a transformation unit 1332. The transformation calculation unit 1331 uses the parameters (initial parameter values) of model P2 and dataset D2 to find a transformation T such that the result of transforming the learning direction or learning process from model P1 to model M1 using transformation T approximates the learning direction or learning process of the parameters when model P2 is trained.

[0122] For example, transformation T is a transformation that minimizes the difference between the result of transforming the parameter learning direction or learning process when model P1 is trained using dataset D2 and the parameter learning direction or learning process when model P2 is trained using dataset D2.

[0123] The conversion unit 1332 converts the learning direction or learning process from the model P1 to the model M1 using the conversion T calculated by the conversion calculation unit 1331, and updates the parameters of the model P2.

[0124] For example, the conversion unit 1332 obtains a "converted learning difference" by converting the learning difference from model P1 to model M1 using a conversion T, and updates the parameters of model P2 using the converted learning difference. In this way, the conversion unit 1332 obtains model M2.

[0125] The output processing unit 134b outputs the parameter values ​​of the model M2 calculated by the calculation unit 133b (the final parameter values ​​of the model P2).

[0126] [Example of Processing Procedure] Next, an example of processing procedure executed by the information processing device 10b will be described with reference to Fig. 11. It is assumed that a setting as to whether or not to use the optimized transformation T when determining the final parameter values ​​of the model P2 (parameters of the model M2) has been input in advance to the information processing device 10b.

[0127] First, the reading unit 132b of the information processing device 10b reads the parameters of the model P1, the parameters of the model M1, the parameters of the model P2, and the data set D2 from the storage unit 12b (S211). If the information processing device 10b uses the optimized transformation T when calculating the final parameter values ​​of the model P2 (Yes in S212), the process proceeds to S213.

[0128] In S213, the conversion calculation unit 1331 uses the data set D2 to calculate the infinitesimal learning direction Δ1 of the parameters of the model P1 and the infinitesimal learning direction Δ2 of the parameters of the model P2.

[0129] For example, the transformation calculation unit 1331 obtains models P1' and P2' by parallel learning models P1 and P2 a relatively small number of times using data set D2. The transformation calculation unit 1331 then calculates Δ1 using the following equation (22). The transformation calculation unit 1331 also calculates Δ2 using the following equation (23).

[0130] Δ1:=(parameter of model P1′)−(parameter of model P1) ...Equation (22) Δ2:=(parameter of model P2′)−(parameter of model P2) ...Equation (23)

[0131] After S213, the transformation calculation unit 1331 optimizes the transformation T for Δ1 using Δ1 and Δ2 calculated in S213 (S214), and then proceeds to S216.

[0132] For example, the transformation calculation unit 1331 calculates a transformation T that minimizes the norm of the difference between Δ1 and Δ2 as a vector (∥T(Δ1)−Δ2∥). The transformation T is calculated using an existing optimization method (for example, optimal transportation).

[0133] On the other hand, in S212, if the information processing device 10b does not use the optimized transformation T when calculating the final parameter value of model P2 (No in S212), it determines that an identity transformation will be used as the transformation T (S215: Transformation T:=Identity transformation), and proceeds to S216.

[0134] In S216, the conversion unit 1332 calculates a "transformed learning difference δ" by converting the difference from the parameters of model P1 to the parameters of model M2 using the conversion T. For example, as shown in the following equation (24), the conversion unit 1332 calculates δ by converting the learning difference (M1-P1) from model P1 to model M1 using the conversion T.

[0135] δ:=T(M1-P1)...Formula (24)

[0136] Next, the conversion unit 1332 uses δ calculated in S216 to calculate the values ​​(final parameter values) of the parameters of the model P2 in the final stage of learning (S217).

[0137] For example, the conversion unit 1332 uses δ calculated in S216 to obtain the final parameter values ​​of model P2 (=parameter values ​​of model M2) based on the following equation (25). Note that λ in equation (25) is a constant for scale adjustment. Then, the conversion unit 1332 updates the parameters of model P2 with the calculated final parameter values. In this way, the conversion unit 1332 obtains model M2.

[0138] M2:=P2+λδ...Formula (25)

[0139] Then, the output processing unit 134b outputs the final parameter values ​​of the model P2 (parameter values ​​of the model M2) calculated in S217 (S218).

[0140] As described above, the transformation calculation unit 1331 calculates minute learning directions for each of the models P1 and P2 using the data set D2, and uses the calculated learning directions to optimize the transformation T. This eliminates the need for the model learning process used in the information processing devices 10 and 10a.

[0141] The conversion calculation unit 1331 converts the parameters from the model P1 to the model M1 in the learning process (θ 1, …,θ n ) may be used to compute the transformation T.

[0142] In this case, the information processing device 10b performs a parameter learning process (for example, θ 1, …,θ n) and stores it in the storage unit 12b. The reading unit 132b then reads the parameter learning process from model P1 to model M1 from the storage unit 12b, and the transformation calculation unit 1331 calculates the transformation T using the parameter learning process from model P1 to model M1. Next, the transformation unit 1332 obtains the final parameter values ​​of model P2 using the result of applying the transformation T to the learning process from model P1 to model M1.

[0143] Furthermore, while the information processing devices 10 and 10a described above use permutation transformation for transformation, the information processing device 10b separates the process of determining transformation T from the process of transferring the content learned by model M1 to model M2. Thus, the information processing device 10b can apply more general transformations when determining model M2.

[0144] Furthermore, in the above-mentioned information processing devices 10 and 10a, the number of optimizations required was at least equal to the length of the learning process, but the information processing device 10b separates the process of determining the transformation T from the process of transferring the content learned by model M1 to model M2, so the number of optimizations can be reduced to one.

[0145] Furthermore, the information processing device 10b can obtain the model M2 even when the data set D1 used for training the model M1 is not available.

[0146] If the accuracy of the model M2 obtained by the information processing device 10b through the above processing (the model M2 after parameter updating using the final parameter values) is not sufficient, additional learning of the model M2 may be performed. In this case, the information processing device 10b further includes an additional learning unit 137 shown in Fig. 10. If the accuracy of the model M2 obtained by the calculation unit 133b is lower than a predetermined accuracy (for example, the accuracy of the model M1), the additional learning unit 137 performs additional learning of the model M2 using the dataset D2.

[0147] For example, as shown in FIG. 12 , the additional learning unit 137 determines whether the accuracy of model M2 is lower than the accuracy of model M1 on dataset D2 (S221). If the additional learning unit 137 determines that the accuracy of model M2 is lower than the accuracy of model M1 (Yes in S221), the additional learning unit 137 additionally learns model M2 on dataset D2 (S222). Thereafter, the output processing unit 134b outputs the parameters of model M2 after the additional learning (S223). On the other hand, if the additional learning unit 137 determines that the accuracy of model M2 is not lower than the accuracy of model M1 (No in S221), the output processing unit 134b outputs the parameters of model M2 (S223).

[0148] By the information processing device 10b performing additional learning of the model M2 as described above, a highly accurate model M2 can be obtained.

[0149] [Application Example] Next, an application example of learning (learning transfer) of the model M2 by the information processing device 10b will be described with reference to FIGS.

[0150] [Application Example 1] For example, learning transfer by the information processing device 10b may be applied to a system including a server and a user terminal device shown in FIG.

[0151] The server shown in Fig. 13 is, for example, a server of a provider that provides each user's terminal device with a base model (base AI) that serves as an initial value for learning. The user's terminal device is a device that constructs a user-specific AI by learning from data using the base AI (Artificial Intelligence) provided by the server as an initial value. The server and the user's terminal device are connected to each other so as to be able to communicate with each other via a network such as the Internet.

[0152] By providing the functions of the information processing device 10b to the user terminal device of the above system, the cost of updating the user-specific AI can be reduced when updating the base AI. Note that, for the sake of simplicity, only one user terminal device (user 1's terminal device) is shown in Figure 13, but multiple user terminal devices may be installed.

[0153] For example, the terminal device of user 1 shown in Figure 13 learns data 1 (e.g., user 1's data) using the base AI (model P1) provided by the server of the base AI provider as the initial value, and constructs AI1 (model M1) dedicated to user 1.

[0154] Thereafter, when User 1's terminal device receives the updated foundation AI (model P2) from the foundation AI provider's server, it trains the updated foundation AI with its own data (for example, data 1 plus an additional element (α)) to obtain a new AI (model M2) dedicated to User 1. At this time, User 1's terminal device can reduce the cost required to update User 1's dedicated AI by applying the learning transfer described above.

[0155] [Application Example 2] In addition, in Application Example 1, a user-specific AI is constructed on the terminal device side of each user, but a user-specific AI may also be constructed on the server side (the server of the AI ​​provision platform). An application example in this case will be described as Application Example 2.

[0156] In this case, the server of the AI ​​provision platform is equipped with the functions of the information processing device 10b. For example, as shown in FIG. 14, the server of the AI ​​provision platform acquires user-specific data (e.g., data 1 for user 1, data 2 for user 2, and data 3 for user 3) from each user's terminal device. The server then constructs an AI (corresponding to model M1) dedicated to each user by training the user-specific data for each user using the base AI (model P1) as the initial value. The server then provides the constructed user-specific AI to each user's terminal device.

[0157] After that, when the server updates the foundation AI, it also reconstructs (trains) each user's dedicated AI in accordance with the updated foundation AI. For example, the server trains the updated foundation AI with User 1's data (Data 1) to obtain a new User 1-dedicated AI1. Similarly, the server trains the updated foundation AI with User 2's data (Data 2) to obtain a new User 2-dedicated AI2. In this case, the server can reduce the cost required to train each user's dedicated AI by applying the above-mentioned learning transfer to the training of each user's dedicated AI.

[0158] In addition, when the server is equipped with the functions of the information processing device 10b as described above, the server's reading unit 132b reads the base AI before the update, the parameters of each user's dedicated AI or its learning process, the updated base AI, and each user's data set.

[0159] In addition, the server's calculation unit 133b calculates a transformation T for each user such that the result of converting the learning direction or learning process of parameters from the pre-update base AI to the user's dedicated AI approximates the learning direction or learning process when the updated base AI is trained using the user's data set.

[0160] The server's calculation unit 133b then calculates the final parameter values ​​of the parameters of the AI ​​dedicated to that user using the transformation T determined for each user, and updates the parameter values ​​of the AI ​​dedicated to that user with the calculated final parameter values.The server then provides the updated AI dedicated to each user to each user's terminal device.

[0161] As described above, by using learning transfer to update the AI ​​dedicated to each user, the server can reduce the cost required to update the AI ​​dedicated to each user.

[0162] [Application Example 3] Furthermore, learning transfer may be applied to updating the platform AI itself provided by the platform AI provider or the AI ​​provision platform. An application example in this case will be described as Application Example 3.

[0163] For example, consider the case where a provider of a foundational AI or a server of an AI provision platform constructs a foundational model (foundational AI) M1 by training an initial model P1 with data D1, as shown in Figure 15. In this case, when the server changes the initial model P1 (for example, when it changes the size of the initial model P1), it is necessary to update the foundational model M1 accordingly.

[0164] For example, if the server changes the size of the initial model P1 to the initial model P2, it is necessary to obtain a base model M2 by training the changed initial model (initial model P2) with data D1+α. In this case, the server can reduce the cost required for training the base model M2 by obtaining the base model M2 through learning transfer.

[0165] The models and AIs described in the above embodiments include, for example, large language models (LLMs), generative AI, etc., which are realized by machine learning models.

[0166] [System Configuration, etc.] The components of each unit shown in the figure are conceptual functional units and do not necessarily have to be physically configured as shown. In other words, the specific form of distribution and integration of each device is not limited to that shown, and all or part of them can be functionally or physically distributed and integrated in any unit depending on various loads, usage conditions, etc. Furthermore, all or any part of the processing functions performed by each device can be realized by a CPU and a program executed by the CPU, or can be realized as hardware using wired logic.

[0167] Furthermore, among the processes described in the above embodiments, all or part of the processes described as being performed automatically can be performed manually, or all or part of the processes described as being performed manually can be performed automatically using a known method.In addition, the information including the processing procedures, control procedures, specific names, various data and parameters shown in the above documents and drawings can be changed as desired unless otherwise specified.

[0168] [Program] The information processing devices 10, 10a, and 10b can be implemented by installing a program (information processing program) as package software or online software on a desired computer. For example, by executing the program on an information processing device, the information processing device can function as the information processing device 10, 10a, or 10b. The information processing device referred to here includes mobile communication terminals such as smartphones, mobile phones, and PHS (Personal Handyphone Systems), as well as terminals such as PDAs (Personal Digital Assistants).

[0169] 16 is a diagram showing an example of a computer that executes an information processing program. The computer 1000 includes, for example, a memory 1010 and a CPU 1020. The computer 1000 also includes a hard disk drive interface 1030, a disk drive interface 1040, a serial port interface 1050, a video adapter 1060, and a network interface 1070. These components are connected by a bus 1080.

[0170] The memory 1010 includes a read-only memory (ROM) 1011 and a random access memory (RAM) 1012. The ROM 1011 stores a boot program such as a basic input / output system (BIOS). The hard disk drive interface 1030 is connected to a hard disk drive 1090. The disk drive interface 1040 is connected to a disk drive 1100. A removable storage medium such as a magnetic disk or optical disk is inserted into the disk drive 1100. The serial port interface 1050 is connected to a mouse 1110 and a keyboard 1120, for example. The video adapter 1060 is connected to a display 1130, for example.

[0171] The hard disk drive 1090 stores, for example, an OS 1091, an application program 1092, a program module 1093, and program data 1094. That is, the programs that define the processes executed by the information processing device 10 are implemented as program modules 1093 in which computer-executable code is written. The program modules 1093 are stored, for example, in the hard disk drive 1090. For example, the program modules 1093 for executing processes similar to those of the functional configuration of the information processing device 10 are stored in the hard disk drive 1090. Note that the hard disk drive 1090 may be replaced with an SSD (Solid State Drive).

[0172] Data used in the processing of the above-described embodiment is stored as program data 1094, for example, in the memory 1010 or the hard disk drive 1090. The CPU 1020 then reads the program module 1093 or the program data 1094 stored in the memory 1010 or the hard disk drive 1090 into the RAM 1012 as necessary and executes them.

[0173] The program module 1093 and program data 1094 may not necessarily be stored in the hard disk drive 1090, but may also be stored in a removable storage medium and read by the CPU 1020 via the disk drive 1100 or the like. Alternatively, the program module 1093 and program data 1094 may be stored in another computer connected via a network (such as a local area network (LAN) or a wide area network (WAN)). The program module 1093 and program data 1094 may then be read by the CPU 1020 from the other computer via the network interface 1070.

[0174] The following additional notes are provided regarding the above-described embodiments.

[0175] (Supplementary Item 1) An information processing device including: a memory; and at least one processor connected to the memory, wherein the processor reads a parameter learning process performed by stochastic gradient descent using a dataset for a first machine learning model and initial parameter values ​​for a second machine learning model; calculates a permutation transformation that minimizes the distance between a result of permutation transformation of the parameter learning process for the first machine learning model and the parameter learning process when the second machine learning model is trained using the initial parameter values ​​of the second machine learning model and the dataset; and outputs parameter values ​​for a final process in the parameter learning process for the second machine learning model calculated by the permutation transformation.

[0176] (Supplementary Item 3) A non-transitory storage medium storing a program executable by a computer to perform information processing, wherein the information processing comprises: reading a parameter learning process performed by stochastic gradient descent using a dataset in a first machine learning model and initial parameter values ​​of a second machine learning model; determining a permutation transformation that minimizes the distance between the result of permutation transformation of the parameter learning process in the first machine learning model and the parameter learning process when the second machine learning model is trained using the initial parameter values ​​of the second machine learning model and the dataset; and outputting parameter values ​​of the final process in the parameter learning process of the second machine learning model calculated by the permutation transformation.

[0177] (Supplementary Item 4) An information processing device including: a memory; and at least one processor connected to the memory, wherein the processor reads a parameter learning process performed by stochastic gradient descent using a dataset in a first machine learning model and initial parameter values ​​of a second machine learning model; divides the parameter learning process into a plurality of segments; calculates, for each segment, a permutation transformation that minimizes the sum of a first distance, which is a distance between a result of permutation transformation of the parameter learning process in the first machine learning model, and the initial parameter values ​​of the second machine learning model and the parameter learning process in the second machine learning model using the dataset, and a second distance, which is a difference between the permutation transformation corresponding to the segment and the permutation transformation corresponding to the segment immediately preceding the segment; calculates parameter values ​​of a final process in the parameter learning process of the second machine learning model using the permutation transformation for each segment; and outputs the calculated parameter values ​​of the final process in the parameter learning process of the second machine learning model.

[0178] (Supplementary Item 6) A non-transitory storage medium storing a program executable by a computer to perform information processing, wherein the information processing comprises: reading a parameter learning process performed by stochastic gradient descent using a dataset in a first machine learning model and initial parameter values ​​of a second machine learning model; dividing the parameter learning process into a plurality of segments; determining, for each segment, a permutation transformation that minimizes the sum of a first distance, which is the distance between a result of permutation transformation of the parameter learning process in the first machine learning model, and the initial parameter values ​​of the second machine learning model and the parameter learning process in the second machine learning model using the dataset, and a second distance, which is the difference between the permutation transformation corresponding to the segment and the permutation transformation corresponding to the segment immediately preceding the segment; calculating parameter values ​​for a final process in the parameter learning process of the second machine learning model using the permutation transformation for each segment; and outputting the calculated parameter values ​​for the final process in the parameter learning process of the second machine learning model.

[0179] (Supplementary Item 7) An information processing device comprising: a memory; and at least one processor connected to the memory, wherein the processor reads initial parameter values ​​and learned parameters or the learning process of a first machine learning model, and initial parameter values ​​of a second machine learning model and a dataset used for training the second machine learning model; determines a transformation such that a result of transforming the learning direction or learning process of the first machine learning model approximates the learning direction or learning process of the parameters when the second machine learning model is trained using the initial parameter values ​​of the second machine learning model and the dataset; and outputs the learning direction of the parameters of the second machine learning model or the final parameter value in the learning process obtained by transforming the learning direction or learning process of the first machine learning model using the transformation.

[0180] (Supplementary Item 8) An information processing device according to Supplementary Item 7, wherein the transformation is a transformation that minimizes the difference between the result of transforming the learning direction or learning process of parameters when the first machine learning model is trained using the dataset and the learning direction or learning process of parameters when the second machine learning model is trained using the dataset.

[0181] (Supplementary Item 9) An information processing device according to Supplementary Item 7, comprising: an additional learning unit that, when it is determined that the accuracy of the second machine learning model after parameter updating using the final parameter value is lower than a predetermined accuracy, performs additional learning of the second machine learning model after parameter updating using the dataset.

[0182] (Supplementary Item 11) A non-transitory storage medium storing a program executable by a computer to perform information processing, wherein the information processing includes: reading initial parameter values ​​and learned parameters or the learning process of a first machine learning model, and initial parameter values ​​of a second machine learning model and a dataset used to train the second machine learning model; determining a transformation such that the result of transforming the learning direction or learning process of the first machine learning model approximates the learning direction or learning process of the parameters when the second machine learning model is trained using the initial parameter values ​​of the second machine learning model and the dataset; and outputting the learning direction of the parameters of the second machine learning model or the final parameter value in the learning process obtained by transforming the learning direction or learning process of the first machine learning model using the transformation.

[0183] (Appendix 12) A terminal device that performs learning using a base model provided from a server that provides a base model that serves as the initial value for learning as an initial value, the terminal device including: a memory; and at least one processor connected to the memory, wherein the processor reads a first base model and parameters of the first base model after learning, a second base model and a dataset used for learning using the second base model as an initial value, determines a transformation such that the result of converting the learning direction or learning process of the parameters of the first base model approximates the learning direction or learning process obtained by learning using the second base model as an initial value using the dataset, and outputs the learning direction or final parameter value in the learning process using the second base model as an initial value obtained by converting the learning direction or learning process of the parameters of the first base model using the transformation.

[0184] (Appendix 13) A server that provides each terminal device with a machine learning model trained for each terminal device using a base model that serves as an initial value for learning, the terminal device including: a memory; and at least one processor connected to the memory, wherein the processor reads a first base model and the parameters or learning process of the machine learning model after learning for each terminal device, and a second base model and a dataset used for learning for each terminal device with the second base model as the initial value, calculates a transformation for each terminal device such that the result of converting the learning direction or learning process of the parameters of the first base model approximates the learning direction or learning process obtained by learning with the second base model as the initial value using the dataset of the terminal device, outputs for each terminal device the final parameter value in the learning direction or learning process with the second base model as the initial value obtained by converting the learning direction or learning process of the parameters of the first base model using the transformation, and provides for each terminal device a machine learning model whose parameters have been updated using the final parameters.

[0185] 10, 10a, 10b Information processing device 11, 11b Input / output unit 12, 12b Storage unit 13, 13b Control unit 131, 131b Input reception unit 132, 132b Reading unit 133, 133a, 133b Calculation unit 134, 134b Output processing unit 135 Information generation unit 136 Division unit 137 Additional learning unit

Claims

1. a reading unit that reads a parameter learning process by a stochastic gradient descent method using a dataset in the first machine learning model and initial parameter values ​​of the second machine learning model; a calculation unit that obtains a permutation transformation that minimizes the distance between a result of permutation transformation of a parameter learning process in the first machine learning model and a parameter learning process when the second machine learning model is trained using initial parameter values ​​of the second machine learning model and the dataset; an output processing unit that outputs the parameter values ​​of the final process in the learning process of the parameters of the second machine learning model calculated by the permutation transformation; An information processing device comprising:

2. An information processing method executed by an information processing device, A step of reading a parameter learning process of a first machine learning model by a stochastic gradient descent method using a dataset and initial parameter values ​​of a second machine learning model; a step of obtaining a permutation transformation that minimizes the distance between a result of permutation transformation of a parameter learning process in the first machine learning model and a parameter learning process when the second machine learning model is trained using the initial parameter values ​​of the second machine learning model and the dataset; outputting the parameter values ​​of the final step in the parameter learning process of the second machine learning model calculated by the permutation transformation; An information processing method comprising:

3. An information processing program for causing a computer to function as the information processing device according to claim 1.

4. a reading unit that reads a parameter learning process by a stochastic gradient descent method using a dataset in the first machine learning model and initial parameter values ​​of the second machine learning model; a division unit that divides the parameter learning process into a plurality of segments; a calculation unit that calculates, for each segment, a permutation transformation that minimizes the sum of a first distance, which is a distance between a result of permutation transformation of a parameter learning process in the first machine learning model and an initial parameter value of the second machine learning model and a parameter learning process in the second machine learning model using the dataset, and a second distance, which is a difference between the permutation transformation corresponding to the segment and the permutation transformation corresponding to the segment immediately preceding the segment, and calculates a parameter value of a final process in the parameter learning process of the second machine learning model using the permutation transformation for each segment; an output processing unit that outputs the calculated parameter values ​​of the second machine learning model in the final step of the learning process; An information processing device comprising:

5. An information processing method executed by an information processing device, A step of reading a parameter learning process of a first machine learning model by a stochastic gradient descent method using a dataset and initial parameter values ​​of a second machine learning model; Dividing the parameter learning process into a plurality of segments; a step of obtaining, for each segment, a permutation transformation that minimizes the sum of a first distance, which is the distance between a result of permutation transformation of a parameter learning process in the first machine learning model and an initial parameter value of the second machine learning model and a parameter learning process in the second machine learning model using the dataset, and a second distance, which is the difference between the permutation transformation corresponding to the segment and the permutation transformation corresponding to the segment immediately preceding the segment; Calculating parameter values ​​of a final step in a parameter learning process of the second machine learning model using the permutation transformation for each segment; outputting the calculated parameter values ​​of the final step in the learning process of the second machine learning model; An information processing method comprising:

6. An information processing program for causing a computer to function as the information processing device according to claim 4.

7. a reading unit that reads initial parameter values ​​and learned parameters of a first machine learning model or the learning process thereof, and initial parameter values ​​of a second machine learning model and a dataset used for learning the second machine learning model; a calculation unit that calculates a transformation such that a result of transforming the learning direction or learning process of the first machine learning model approximates the learning direction or learning process of parameters when the second machine learning model is trained using the initial parameter values ​​of the second machine learning model and the dataset; an output processing unit that outputs the learning direction or final parameter value in the learning process of the parameter of the second machine learning model obtained by converting the learning direction or learning process of the first machine learning model through the conversion; An information processing device comprising:

8. The conversion is The transformation is such that the difference between the result of transforming the learning direction or learning process of the parameters when the first machine learning model is trained using the dataset and the learning direction or learning process of the parameters when the second machine learning model is trained using the dataset is minimized.

8. The information processing apparatus according to claim 7,

9. an additional learning unit that performs additional learning of the second machine learning model after the parameter update using the data set when it is determined that the accuracy of the second machine learning model after the parameter update using the final parameter value is lower than a predetermined accuracy; 8. The information processing apparatus according to claim 7, further comprising:

10. An information processing method executed by an information processing device, a step of reading initial parameter values ​​and learned parameters of a first machine learning model or the learning process thereof, and initial parameter values ​​of a second machine learning model and a dataset used for learning the second machine learning model; determining a transformation such that a result of transforming the learning direction or learning process of the first machine learning model approximates the learning direction or learning process of parameters when the second machine learning model is trained using the initial parameter values ​​of the second machine learning model and the dataset; outputting the learning direction or final parameter value in the learning process of the parameters of the second machine learning model obtained by converting the learning direction or learning process of the first machine learning model through the conversion; An information processing method comprising:

11. An information processing program for causing a computer to function as the information processing device according to claim 7.

12. A terminal device that performs learning using a base model provided from a server that provides a base model that serves as an initial value for learning as an initial value, a reading unit that reads a first base model and parameters of the first base model after learning, and a second base model and a data set used for learning with the second base model as an initial value; a calculation unit that calculates a transformation such that the result of transforming the learning direction or learning process of the parameters of the first foundation model approximates the learning direction or learning process obtained by learning using the second foundation model as an initial value using the data set; an output processing unit that outputs the learning direction or final parameter value in the learning process with the second base model as an initial value, obtained by converting the learning direction or learning process of the parameters of the first base model by the conversion; A terminal device comprising:

13. A server that provides a terminal device with a machine learning model in which a base model serving as an initial value for learning is trained for each terminal device, a reading unit that reads the parameters of the first base model and the machine learning model after learning for each of the terminal devices or the learning process thereof, and the second base model and a dataset used for learning for each of the terminal devices with the second base model as the initial value; a calculation unit that calculates a transformation for each of the terminal devices such that the result of transforming the learning direction or learning process of the parameters of the first basic model approximates the learning direction or learning process obtained by learning using the data set of the terminal device with the second basic model as the initial value; an output processing unit that outputs, for each of the terminal devices, the values ​​of the final parameters in the learning direction or learning process in which the second basic model is an initial value, obtained by converting the learning direction or learning process of the parameters of the first basic model by the conversion; a providing unit that provides the machine learning model of the terminal device, the parameters of which have been updated using the final parameters, to the terminal device; A server comprising:

14. A server that provides AI (Artificial Intelligence) to a terminal device of each user, the AI ​​having a basic model that serves as an initial value for learning trained specifically for each user, A reading unit that reads the first base model and the parameters of the AI ​​after learning for each user or the learning process thereof, and the second base model and a user-specific dataset used for learning the AI ​​for each user using the second base model as an initial value; a calculation unit that calculates a transformation for each user such that the result of transforming the learning direction or learning process of the parameters of the first foundation model approximates the learning direction or learning process obtained by learning using a data set dedicated to that user and with the second foundation model as an initial value; an output processing unit that outputs, for each user, a learning direction or a final parameter value in a learning process in which the second basic model is an initial value, obtained by converting the learning direction or learning process of the parameters of the first basic model by the conversion; a providing unit that provides the AI ​​dedicated to the user, whose parameters have been updated using the final parameters, to the terminal device of the user; A server comprising:

15. The reading unit further reading the initial parameter values ​​and learned parameters of the first base model or the learning process thereof, and the initial parameter values ​​of the second base model and a data set used for learning the second base model; The calculation unit further determining a transformation such that the result of transforming the learning direction or learning process of the first foundational model approximates the learning direction or learning process of the parameters when the second foundational model is trained using the initial parameter values ​​of the second foundational model and the data set; The output processing unit further The first foundation model is updated to the second foundation model using the learning direction or final parameter value in the learning process of the parameters of the second foundation model obtained by converting the learning direction or learning process of the first foundation model through the conversion.

15. The server of claim 14.