Information processing device, information processing method, and program

Abstract

An information processing device 100 according to the present disclosure comprises: a generation unit 121 that generates a second model that approximates a prediction regarding a prescribed case by a first model that is a machine learning model; a loss setting unit 122 that sets a loss based on the second model; and an update unit 123 that updates the first model using the loss.

Description

Information processing device, information processing method, and program

[0001] The present disclosure relates to an information processing device, an information processing method, and a program.

[0002] In the field of machine learning, when training a machine learning model, it is important to avoid overfitting and generalize the model, which means that the model can make accurate predictions not only for training data but also for unknown data.

[0003] To generalize a model, it is important to make the model smooth, meaning that its predictions do not change abruptly when small variations in the input are made.

[0004] Here, Non-Patent Document 1 discloses a technique called weight decay. According to this technique, a penalty is applied to the objective variable so that the weight of the linear layer of a neural network becomes smaller. If the weight of the linear layer is small, the fluctuation of the output of the linear layer when the input to the linear layer changes becomes small, making it easier for the model to generalize. This technique is also called L2 regularization or shrinkage.

[0005] Anders Krogh, John A. Hertz, “A simple weight decay can improve generalization”, Proceedings of the 4th International Conference on Neural Information Processing Systems, December 1991, Pages 950-957M. T. Ribeiro, S. Singh, and C. Guestrin, ““why should I trust you?”: Explaining the predictions of any classifier,” in Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2016, pp. 1135-1144.

[0006] However, the technology disclosed in Non-Patent Document 1 has the problem that while it can suppress sudden changes in predictions resulting from linear transformations, it cannot suppress sudden changes in predictions resulting from non-linear transformations. This problem is particularly pronounced in rule-based models based on thresholds, such as gradient boosting decision trees. Rule-based models are non-linear models that determine predictions using thresholds, and predictions tend to change suddenly in response to input fluctuations. However, because they do not have a linear layer, the technology disclosed in Non-Patent Document 1 cannot suppress sudden changes. Furthermore, this problem is not limited to rule-based models and can occur in non-linear models.

[0007] Therefore, an object of the present disclosure is to solve the above-mentioned problem that it is difficult to train a nonlinear model to suppress sudden changes in predictions.

[0008] An information processing device according to an embodiment of the present disclosure is configured to include: a generation unit that generates a second model that approximates a prediction for a predetermined case made by a first model that is a machine learning model; a loss setting unit that sets a loss based on the second model; and an update unit that updates the first model using the loss. An information processing method according to an embodiment of the present disclosure is configured to generate a second model that approximates a prediction for a predetermined case made by a first model that is a machine learning model; set a loss based on the second model; and update the first model using the loss. A program according to an embodiment of the present disclosure is configured to cause a computer to execute processes of generating a second model that approximates a prediction for a predetermined case made by a first model that is a machine learning model; setting a loss based on the second model; and updating the first model using the loss.

[0009] With the above-described configuration, the present disclosure can be trained to suppress sudden changes in predictions in a nonlinear model.

[0010] FIG. 1 is a diagram illustrating an overview of the present disclosure. FIG. 2 is a diagram illustrating an overview of the present disclosure. FIG. 3 is a block diagram illustrating a configuration of an information processing device according to the present disclosure. FIG. 4 is a flowchart illustrating an operation of an information processing device according to the present disclosure. FIG. 5 is a flowchart illustrating an operation of an information processing device according to the present disclosure. FIG. 6 is a diagram illustrating a state of processing by an information processing device according to the present disclosure. FIG. 7 is a block diagram illustrating a hardware configuration of an information processing device according to the present disclosure. FIG. 8 is a block diagram illustrating a configuration of an information processing device according to the present disclosure.

[0011] First Embodiment A first embodiment of the present disclosure will be described with reference to the drawings. Note that the drawings may be relevant to any embodiment.

[0012] [Overview] First, an overview of the present disclosure will be described with reference to Figures 1 and 2. An information processing device in the present disclosure trains a machine learning model using training examples and updates parameters of the machine learning model.

[0013] Here, the machine learning model trained by the information processing device in the present disclosure is referred to as base model f. The base model f may be any machine learning model capable of updating parameters using the gradient of the objective function. For example, a neural network or gradient boosting can be used as such a base model f. Furthermore, the parameters that determine the behavior of the base model f are represented by a vector θ. For example, when the base model f is a neural network, θ is a vector consisting of the weights of the neural network. When the base model f is gradient boosting, θ is the number of weak learners or a parameter of the weak learner. The output of the base model f is determined depending on the value of θ.

[0014] FIG. 1 shows an example of a base model f, whose parameters are updated by an information processing device. The horizontal axis represents input x, and the vertical axis represents output y. Line f in the figure plots the predicted output of base model f for input x. That is, line f is a set of points (x, y) where y = f(x). For the sake of explanation, this figure shows an example in which both input x and output y are one-dimensional, but the present disclosure is not limited to one-dimensional cases. In FIG. 1 (1-1), as an example, base model f is machine-learned using a gradient boosting decision tree and is a nonlinear model. The gray circles on line f each represent a training example and the output of base model f for that training example. For example, in FIG. 1 (1-1), five gray circles correspond to five training examples. If the coordinates of circle C are (x_c, y_c), circle C indicates that when training example x_c is input, base model f outputs y_c, i.e., y_c = f(x_c) holds. When the information processing device is given five training examples corresponding to gray circles, it selects any one of these training examples. For example, suppose that the training example corresponding to Ca (black circle) shown in FIG. 1(1-2) is selected. Furthermore, the information processing device generates multiple neighboring examples, which are similar examples located near the selected training example, and calculates predictions using the base model f. The generated neighboring examples and the predictions made by the base model f for them are plotted in the figure as white circles Cb.

[0015] Next, the information processing device calculates an approximation model g, which is a simple model that approximates the base model f. As the approximation model g, for example, a linear model can be used. Specifically, as shown in FIG. 2 (2-1), the information processing device calculates the approximation model g, which is a linear model that locally approximates predictions for the generated multiple nearby cases. Then, the information processing device uses the calculated approximation model g to train the base model f so as to suppress sudden changes in the predictions in the base model f. At this time, as shown by the arrow in FIG. 2 (2-2), the information processing device performs machine learning to train the base model f so that the base model f approaches the approximation model g, and updates the parameter θ of the base model f.

[0016] Thereafter, the information processing device selects a training case from the training cases that is different from the training case that has already been selected as described above, generates a plurality of neighboring cases for the selected training case in the same manner as described above, calculates an approximate model g, and performs machine learning of the base model f using the approximate model g, thereby repeatedly updating the base model f. In this way, the information processing device can update the entire base model f.

[0017] Here, the machine learning of the base model f is performed using supervised machine learning, and typically a set of training examples and the correct labels associated with each training example included in the set of training examples are input. Then, when each training example is input to the base model f, the parameters are updated so that the difference between the prediction output by the base model f and the correct labels associated with that training example is minimized. The difference between this prediction and the correct label is called the prediction loss.

[0018] In particular, in the present disclosure, the information processing device not only performs machine learning to reduce the prediction loss, but also sets an approximation model loss, which is a loss based on the approximation model g, and updates the base model f by performing machine learning to reduce the approximation model loss. At this time, the information processing device sets the approximation model loss in two ways to suppress sudden changes in the predictions of the base model f, as described below.

[0019] The first method for setting the approximate model loss is to reduce the error between the predictions of the base model f and the approximate model g. This method constrains the base model f to be closer to a simple model such as a linear model, thereby suppressing sudden changes in the predictions of the base model.

[0020] The second method for setting the approximation model loss is to update the base model f so as to reduce the variability of the predictions of the approximation model g. This method indirectly reduces the variability of the predictions of the base model f by reducing the variability of the predictions of the approximation model g calculated from the base model f. Even techniques such as weight decay can be applied to the approximation model g. Therefore, it is possible to suppress sudden changes in the predictions of the base model f using techniques such as weight decay that are not normally applicable to the base model f.

[0021] In the present disclosure, the parameter θ of the model f is updated so as to reduce not only the prediction loss but also the approximation model loss. In particular, it is effective to update the parameter so as to reduce the weighted sum of the prediction loss and the approximation model loss. This makes it possible to balance the prediction loss and the approximation model loss. In addition to the prediction loss, the parameter may be updated so as to reduce the sum of the approximation model loss calculated by the first method and the approximation model loss calculated by the second method.

[0022] [Details] Next, the specific configuration and operation of the information processing device 10 in this embodiment will be described with reference to FIGS.

[0023] The information processing device 10 in this embodiment is configured with one or more information processing devices each including a calculation device and a storage device. As shown in FIG. 3 , the information processing device 10 includes an input unit 11, a base model updating unit 12, a predicted loss calculation unit 13, an approximation model loss calculation unit 14, and an approximation model calculation unit 15. The functions of the input unit 11, the base model updating unit 12, the predicted loss calculation unit 13, the approximation model loss calculation unit 14, and the approximation model calculation unit 15 can be realized by the calculation device executing a program for realizing each function stored in the storage device. The operation of each unit of each configuration will be described below.

[0024] Before describing the overall operation of the information processing device 10 in the first embodiment, the operation of the approximate model calculation unit 15 will be described with reference to the flowchart in Fig. 4. In the description of the approximate model calculation unit 15, the above-mentioned Non-Patent Document 2 will be referred to as appropriate.

[0025] (Step S11) The approximate model calculation unit 15 receives training examples x as input. The training examples x are real-valued vectors of length d that represent examples to be input to the model f. An example of the training examples x is shown in FIG. 7.

[0026] (Step S12) The approximation model calculation unit 15 generates a different representation x' of the training example x. The different representation x' is a vector of length d'. For example, x' can be a binary vector that represents the training example x in terms of the presence or absence of some feature. x' can take various forms as described in Section 3.1 of Non-Patent Document 2. It is also possible to use x as x' without any particular conversion.

[0027] As an example, we will explain a possible method for when the training example x is a vector of continuous values ​​(hereinafter referred to as the threshold method). Two conditions are generated by dividing each of the d elements that make up x using the median as a threshold. For example, if the median of the first element x1 of x is 3, two conditions, "x1 ≥ 3" and "x1 < 3", are generated. This is repeated for the other elements, generating d * 2 conditions. Finally, only the conditions satisfied by x are extracted and used as the features that make up x'. However, the value of each feature is 1 if the condition is satisfied, and 0 otherwise.

[0028] An example of x' generated using the threshold method described above is shown in Figure 8. As shown in this figure, when created using this method, only the conditions satisfied by x are extracted, so all elements of x' will inevitably be 1. Note that the threshold method may also be used to divide into four conditions using quartiles instead of the median. The implementation by the authors of Non-Patent Document 2 (https: / / github.com / marcotcr / limes) implements a threshold method using quartiles.

[0029] (Step S13) In step S13, the approximation model calculation unit 15 generates a set Z of perturbed samples based on x'. The perturbed samples are artificially generated samples and are used as training samples for constructing an approximation model that approximates the local prediction of the base model f around x. The set Z can be generated using the algorithms shown in Section 3.3 and Algorithm 1 of Non-Patent Document 2.

[0030] The parameters for generating the set Z are determined as follows: The number of perturbation examples to be generated is set to N. The function measuring the proximity to x is set to π.ｘ Let π ｘ (z) is an arbitrary function that returns a larger value the closer a vector z of length d is to x, and a smaller value the farther it is from x. For example, the cosine similarity of vectors can be used.

[0031] The method for generating set Z in step S13 of Fig. 4 is shown in the flowchart of Fig. 5. First, set Z is initialized as an empty set (step S21). The variable i is changed from 1 to N and the following is executed (step S22).

[0032] The i-th perturbation case z' ｉ (Step S23). ｉ is a vector of length d', just like x'. Perturbation examples can be generated by any method that can obtain a vector of length d'. For example, they can be obtained by uniformly randomly generating a binary vector of length d'. Figure 8 shows an example of a generated perturbation example. While all values ​​of x' are 1, the perturbation examples take values ​​of 1 and 0. In this example, the perturbation examples are binary vectors, but they are not limited to binary vectors and can be vectors of any real values.

[0033] Perturbation case z' ｉ z, which is the representation in the original space ｉ (Step S24). ｉ is a vector of the same length d as x. In the case of the threshold method described above, for example, the perturbation case z' ｉ From Z ｉ The mean and standard deviation of d elements in the training set are calculated. Then, sampling is performed from d normal distributions with these means and standard deviations as parameters, and z' ｉ The samples that meet the same conditions as ｉ For example, in the example shown in FIG. ２ Since the value satisfies the four conditions "x1 ≥ 3", "x2 ≥ 4", "x3 < 1", and "x4 < 5", we randomly generate values ​​that satisfy these conditions and set z ２ Let's say.

[0034] Next, the base model f is used to predict f(z ｉ ) is obtained (step S25). ｉ By inputting this to the base model f, the prediction f(z ｉ ) is obtained. Next, the proximity π ｘ (z ｉ） (Step S26). Then, the triplet <z' is added to the set Z. ｉ , f(z ｉ ), π ｘ (z ｉ )> is added (step S27).

[0035] The above is repeated N times, and finally the set Z is output (step S28). This completes the explanation of step S13.

[0036] (Step S14) The approximate model calculation unit 15 receives the set Z and calculates an approximate model for x. Specifically, z' ｉ are the training examples, f(z ｉ ) is the correct label, π ｘ (z ｉ ) as weights for the examples to calculate the approximate model g.

[0037] Here, a case where the approximation model g is a linear model will be described. When the approximation model g is a linear model, it can be expressed as in the following equation 1. Note that, for simplicity of explanation, a linear model in which the intercept is omitted is used here, but a linear model in which the intercept is taken into account can be created simply by adding an element to z that is always 1.

[0038] In this case, a design matrix D of N×d′ is defined by the following equation 2. where z' ｉj is z' ｉ represents the j-th element of

[0039] Also, a vector f of length N representing the predictions of model f for N perturbation cases. ｚ is defined by the following equation 3.

[0040] Furthermore, the case weight matrix Π is defined as an N×N diagonal matrix expressed by the following equation 4.

[0041] In this case, the parameter w is a w that minimizes the loss function Lw shown in the following equation (5). The first term of Equation 5 is the predicted value f by the base model f. ｚ The squared error of the predicted Dw by the approximation model g and the proximity is weighted. The second term is a normalization term. The coefficient λ is an arbitrary positive real value.

[0042] The parameter w that minimizes the above loss can be calculated using the following equation (6). I is a d' x d' identity matrix. Here, a d' x N matrix A is defined by the following equation (7). In this case, the parameter w is the predicted value f ｚ As a linear transformation of this, it can be written as in the following equation 8. The above formula 8 is a linear transformation and therefore differentiable.

[0043] Thus, we can obtain the parameter w of the approximate model g in the form of a differentiable function of the prediction of the base model f. By being differentiable, we can calculate the gradient of the approximate model loss with respect to the parameter θ of the base model f, and update the parameter θ of the base model f, as will be described later.

[0044] Furthermore, an N×N matrix B shown in the following equation (9) is calculated in advance for later use in the base model update unit 12.

[0045] The approximation model calculation unit 15 calculates the prediction f of the base model f for the input training example x. ｚ In addition to the matrix A shown in Equation 7 or the matrix B shown in Equation 9, the base model update unit 12 outputs these. As will be described later, the base model update unit 12 updates the base model f using these. Note that it is not necessary to explicitly calculate the parameter w of the approximation model g, and the predicted value f of the base model f is calculated using the ｚ It is sufficient to be able to calculate matrix A or matrix B. This is because the approximate model loss, which will be described later, can be calculated with these.

[0046] Here, differences from Non-Patent Document 2 will be described. The method of calculating an approximation model by the approximation model calculation unit 15 of this embodiment is basically similar to Algorithm 1 in Non-Patent Document 2. However, Non-Patent Document 2 employs a model called K-Lasso, which is difficult to differentiate, as the approximation model g. This is because the ultimate purpose of Non-Patent Document 2 is to generate an explanation, and therefore the parameters of the base model f are not updated using differentiation.

[0047] On the other hand, in this embodiment, the purpose is to update the parameters of the base model f using an approximation model. Therefore, by using a linear model as g instead of K-Lasso, the parameter w is expressed in the form of a differentiable function of the prediction of the base model f. This makes it possible to update the parameters of the base model f using the gradient.

[0048] Although a linear model is used in this embodiment, other models may be used as long as the parameter w can be expressed in the form of a differentiable function of the prediction of the base model f. This concludes the explanation of the operation of the approximation model calculation unit 15.

[0049] Next, the overall operation of the information processing device 10 in this embodiment will be described with reference to FIG.

[0050] First, we will explain the training example set and correct labels that serve as input. The training example set X is a set consisting of M training examples, as shown in Equation 10. Each example is a vector of length d.

[0051] The correct label y is a vector of length M that represents the label to be predicted by the base model f, as shown in Equation 11. The elements of the correct label y are classes in the case of classification, and real values ​​in the case of regression. For the sake of explanation, it is assumed below that regression is used and the correct label y is a real value. The present invention can be applied to both classification and regression.

[0052] The objective function of this embodiment will be described below. The objective function L is given in the form of a weighted sum of the prediction loss and the approximation model loss as shown in the following equation (12). Here, P ｊ is the prediction loss for the jth training example, which is a value that represents the degree to which the prediction output by the base model f for the training example differs from the correct label. ｊ is the approximate model loss for the jth training example. is a coefficient to balance the two losses.

[0053] To update the parameters of the base model for this objective function, P ｊ The gradient of E ｊ In particular, the update formula for the gradient descent method is expressed as the following formula (13). However, θ （ｔ） is the parameter in the t-th update, and η is the learning rate. Here, since Eq. (14) is ｊ The gradient of E ｊ If the gradient of is known, the parameters can be updated. Therefore, in steps S32 and after, for j=1, . . . , M, ｊ The gradient of E ｊ Calculate the gradient of

[0054] (Step S31) The input unit 11 receives a base model f, a set of training examples, and a correct label. (Step S32) The following is repeated for j=1,...,M. (Step S33) The predicted loss calculation unit 13 calculates the j-th training example x ｊ The gradient of the predicted loss with respect to is calculated. For example, in the case of regression, the predicted loss can be a squared error as shown in the following equation (15). In this case, the gradient of the predicted loss for the base model f can be calculated as follows: Once the gradient of the predicted loss with respect to the base model f is known, θ ｋ The gradient of the predicted loss with respect to θ can be calculated as follows: ｋ is the k-th element of θ. In the case of a neural network, the term shown in Equation 18 in Equation 17 can be calculated by the error back propagation method.

[0055] (Step S34) The approximation model calculation unit 15 calculates the j-th training example x ｊ The calculation method is as explained above. In particular, in the case of a linear model, the j-th training example x ｊ , the matrix A shown in Equation 7 or the matrix B shown in Equation 9 is calculated. Note that the approximate model g is calculated for each training example, so it is assumed that the approximate model g is calculated for the j-th training example x ｊ In the following, we will use the subscript j to represent A and B calculated using the input ｊ and B ｊ It is written as follows.

[0056] In addition, the predicted f of the base model f calculated by the approximation model calculation unit 15 ｚ Also received. ｚ Similarly, add a subscript to f ｚ （ｊ） That is, f ｚ （ｊ） is a vector of length N representing the predictions of the base model f on the N perturbed examples generated for the jth training example.

[0057] (Step S35) The approximation model loss calculation unit 14 calculates the j-th training example x ｊ Calculate the gradient of the approximate model loss with respect to . Specifically, two methods are described below.

[0058] The first method of approximate model loss: Approximate model loss E ｊ The first way to set the loss is to set it so that the smaller the approximate model loss, the closer the predictions of the base model and the approximate model become. This updates the base model to produce predictions that are closer to the predictions of the simple approximate model, reducing sudden changes in predictions and improving generalization performance.

[0059] As such a loss, for example, the square error shown in the following equation 19 can be used. This approximate model loss E ｊis the base model prediction f(z ｊ ) is the prediction of the approximation model g(z ｊ ) the greater the deviation.

[0060] When the approximation model g is a linear model, the parameters of the approximation model g are expressed as w ｊ Then, it can be expressed as in Equation 20.

[0061] Furthermore, the matrix B calculated using Equation 9 ｊ When rewritten using the formula, it can be expressed as formula 21. I is an NxN identity matrix.

[0062] f ｚ （ｊ） The gradient of the approximate model loss with respect to can be calculated as follows:

[0063] f ｚ （ｊ） If the gradient of the approximate model loss with respect to can be calculated, the gradient of the predicted loss with respect to the parameter vector θ of the base model f can also be calculated as follows: ｋ The gradient of the approximate model loss with respect to can be written as an inner product of vectors by the chain rule as shown in the following equation (23).

[0064] Here, the term shown in Equation 24 in Equation 23 can be calculated as described above. Regarding the term shown in Equation 25 in Equation 23, f ｚ （ｊ） By the definition of (26), we obtain the formula (26), which means that we need to calculate the gradient of the base model f for N predictions for N perturbation cases. This gradient can be calculated using the backpropagation method in the case of a neural network, for example. Thus, the approximate model loss E ｊ This concludes the description of the first method for setting

[0065] ・Second method of approximate model loss Approximate model loss E ｊThe second method for setting is to set the approximation model loss so that the smaller the approximation model loss, the less abrupt changes occur in the predictions of the approximation model g. In other words, the smoother the approximation model g, the smaller the approximation model loss. Because the approximation model f is calculated from the predictions of the base model f, setting such a loss updates the base model f so that the predictions of the base model f also become smoother, reducing abrupt changes in the predictions and improving generalization performance. The advantage of this method is that even techniques such as weight decay that cannot be directly applied to the base model f can be directly applied to the approximation model g.

[0066] In the following, we will describe a method for applying weight decay to the approximate model g. That is, we will use the L2 norm of the parameters of the approximate model f as the approximate model loss E ｊ This is expressed as the following equation (27).

[0067] When the approximate model g is a linear model, the matrix A output by the approximate model calculation unit 15 is ｊ Using the above, it can be written as in Equation 28.

[0068] f ｚ （ｊ） The gradient of the approximate model loss with respect to can be calculated as follows:

[0069] f ｚ （ｊ） Since the gradient of the approximate model loss with respect to has been calculated, the gradient of the approximate model loss with respect to the parameter vector θ of the base model f can also be calculated by the same method as described in the first method. ｊ This concludes the explanation of the second method for setting

[0070] The two methods described above can be combined. For example, the sum of the loss calculated by the first method and the loss calculated by the second method can be used as the approximate model loss. In that case, the gradient calculated by the two methods can be added together.

[0071] (Step S36) The above is repeated for M cases, and then the process proceeds to the next step. (Step S37) Subsequently, the base model update unit 12 calculates the parameter θ using the gradient. In the case of a neural network, the parameter θ can be updated by the gradient descent method. P ｊ and E ｊ Since the gradients of have already been calculated as above, we can use them to update using the gradient descent update formula.

[0072] (Step S38) The updated parameters are then output. Note that the procedure from steps S32 to S37 may be repeated multiple times. This updates the parameters multiple times, which is expected to reduce losses and lead to better parameters.

[0073] Next, further possible variations of the first embodiment will be described. (In the Case of Gradient Boosting) A case where the base model f is gradient boosting will be described. In gradient boosting, instead of using an update formula for gradient descent, parameters are updated by adding a base learner that uses pseudo-residuals as target values. In gradient boosting, parameters can be considered to be added with each update. Note that with gradient boosting, a combination of multiple base learners corresponds to the base model f described in the present disclosure.

[0074] In conventional gradient boosting, the pseudo-residual r ｊ is calculated as shown in the following formula 30.

[0075] Then, the base learner is trained and added using Equation 31 as a training example set.

[0076] When applied to the present disclosure, N perturbation examples shown in the following equation 32 are added per training example to the training example set when training the base learner. where the pseudo residuals s for the perturbed casesｉ is calculated as shown in the following formula 33. That is, s ｉ is obtained by multiplying the i-th element of equation 24 by (-λ).

[0077] By adding the base learner calculated as above, the parameters are updated so that not only the prediction loss but also the approximate model loss is reduced.

[0078] Second Embodiment Next, a second embodiment of the present disclosure will be described with reference to the drawings. This embodiment shows an outline of the configuration of the information processing device described in the above embodiment. Note that Figures 9 and 10 are diagrams for explaining the configuration, and these diagrams may be relevant to any of the embodiments.

[0079] First, the hardware configuration of the information processing device 100 will be described with reference to Fig. 9. The information processing device 100 is configured as a general information processing device, and is equipped with the following hardware configuration, for example: CPU (Central Processing Unit) 101 (arithmetic unit); ROM (Read Only Memory) 102 (storage device); RAM (Random Access Memory) 103 (storage device); programs 104 loaded into RAM 103; storage device 105 storing programs 104; drive device 106 for reading and writing data from and to a storage medium 110 external to the information processing device; communication interface 107 for connecting to a communication network 111 external to the information processing device; input / output interface 108 for inputting and outputting data; and bus 109 for connecting the various components.

[0080] 9 shows an example of the hardware configuration of the information processing device 100, and the hardware configuration of the information processing device is not limited to the above-described case. For example, the information processing device may be configured with only a part of the above-described configuration, such as excluding the drive device 106. Furthermore, the information processing device may use a GPU (Graphics Processing Unit), a DSP (Digital Signal Processor), an MPU (Micro Processing Unit), an FPU (Floating Point Number Processing Unit), a PPU (Physics Processing Unit), a TPU (Tensor Processing Unit), a quantum processor, a microcontroller, or a combination thereof, instead of the above-described CPU.

[0081] The information processing device 100 can then configure and include the generation unit 121, loss setting unit 122, and update unit 123 shown in FIG. 10 by having the CPU 101 acquire and execute the program group 104. The program group 104 is stored in advance in the storage device 105 or ROM 102, for example, and is loaded into the RAM 103 and executed by the CPU 101 as needed. The program group 104 may be supplied to the CPU 101 via the communication network 111, or may be stored in advance in the storage medium 110, with the drive device 106 reading out the programs and supplying them to the CPU 101. However, the generation unit 121, loss setting unit 122, and update unit 123 described above may be configured using dedicated electronic circuits for realizing such means.

[0082] The generating unit 121 generates a second model that approximates a prediction for a predetermined case made by a first model, which is a machine learning model. The loss setting unit 122 sets a loss based on the second model. The updating unit 123 updates the first model using the loss.

[0083] With the above configuration, the present disclosure updates the first model using a loss based on a second model that approximates the prediction by the first model for a given case, thereby enabling training to suppress sudden changes in predictions in a nonlinear model.

[0084] At least one of the functions of the generation unit 121, loss setting unit 122, and update unit 123 described above may be executed by an information processing device installed and connected anywhere on the network, that is, may be executed by so-called cloud computing.

[0085] The above-described program can be stored and supplied to a computer using various types of non-transitory computer-readable media. Non-transitory computer-readable media include various types of tangible storage media. Examples of non-transitory computer-readable media include magnetic recording media (e.g., flexible disks, magnetic tapes, hard disk drives), magneto-optical recording media (e.g., magneto-optical disks), CD-ROMs (Read Only Memory), CD-Rs, CD-RWs, and semiconductor memories (e.g., mask ROMs, PROMs (Programmable ROMs), EPROMs (Erasable PROMs), flash ROMs, and RAMs (Random Access Memory)). The program can also be supplied to a computer by various types of transitory computer-readable media. Examples of transitory computer-readable media include electrical signals, optical signals, and electromagnetic waves. The transitory computer-readable media can be supplied to a computer via wired communication paths such as electric wires and optical fibers, or via wireless communication paths.

[0086] Although the present disclosure has been described above with reference to the above-described embodiments, the present disclosure is not limited to the above-described embodiments. Various modifications that can be understood by those skilled in the art can be made to the configuration and details of the present disclosure within the scope of the present disclosure. Furthermore, each of the above-described embodiments can be combined with other embodiments as appropriate.

[0087] The following definitions and abbreviations will be used in interpreting the claims and the specification: As used herein, the terms "comprises," "comprising," "includes," "including," "has," "having," "contains," or "containing," or any other variation thereof, are intended to cover a non-exclusive inclusion. For example, a composition, mixture, process, method, article, or device that includes a list of elements is not necessarily limited to only those elements and may include a list of other elements not expressly listed or inherent to such composition, mixture, process, method, article, or device.

[0088] <Supplementary Notes> Some or all of the above embodiments may be described as in the following supplementary notes. Below, an outline of the configurations of an information processing device, an information processing method, and a program according to the present disclosure will be described. However, the present disclosure is not limited to the following configurations. (Supplementary Note 1) An information processing device comprising: a generation unit that generates a second model that approximates a prediction for a predetermined case made by a first model that is a machine learning model; a loss setting unit that sets a loss based on the second model; and an update unit that updates the first model using the loss. (Supplementary Note 2) The information processing device according to Supplementary Note 1, wherein the loss setting unit sets the loss based on a prediction of the second model for the predetermined case, and the update unit updates the first model to reduce the loss. (Supplementary Note 3) The information processing device according to Supplementary Note 2, wherein the loss setting unit sets the loss based on an error between a prediction of the first model and a prediction of the second model for the predetermined case, and the update unit updates the first model to reduce the loss. (Supplementary Note 4) The information processing device according to Supplementary Note 1, wherein the loss setting unit sets the loss based on a weight of the second model, and the updating unit updates the first model to reduce the loss. (Supplementary Note 5) The information processing device according to Supplementary Note 4, wherein the loss setting unit sets the loss based on an absolute value of a weight of the second model, and the updating unit updates the first model to reduce the loss. (Supplementary Note 6) The information processing device according to Supplementary Note 1, wherein the loss setting unit sets the loss based on a prediction of the second model for the specified example and a weight of the second model, and the updating unit updates the first model to reduce the loss. (Supplementary Note 7) The information processing device according to Supplementary Note 1, wherein the generation unit generates the second model that approximates a prediction by the first model for the specified example including a neighboring example located in the vicinity of a training example of the first model.(Supplementary Note 8) The information processing device according to Supplementary Note 7, wherein the generation unit generates the second model, which is a linear model that approximates a prediction by the first model for the specified example including the neighboring example. (Supplementary Note 9) An information processing method, comprising: generating a second model that approximates a prediction for a specified example by a first model that is a machine learning model; setting a loss based on the second model; and updating the first model using the loss. (Supplementary Note 10) The information processing method according to Supplementary Note 9, wherein: setting the loss based on a prediction of the second model for the specified example; and updating the first model to reduce the loss. (Supplementary Note 11) The information processing method according to Supplementary Note 10, wherein: setting the loss based on an error between a prediction of the first model and a prediction of the second model for the specified example; and updating the first model to reduce the loss. (Supplementary Note 12) The information processing method according to Supplementary Note 9, wherein: setting the loss based on a weight of the second model; and updating the first model to reduce the loss. (Supplementary Note 13) The information processing method according to Supplementary Note 12, comprising: setting the loss based on an absolute value of a weight of the second model; and updating the first model to reduce the loss. (Supplementary Note 14) The information processing method according to Supplementary Note 9, comprising: generating the second model that approximates a prediction by the first model for the predetermined case including a nearby case located in the vicinity of a training case of the first model. (Supplementary Note 15) A computer-readable storage medium storing a program that causes a computer to execute processes of generating a second model that approximates a prediction by a first model that is a machine learning model for a predetermined case, setting a loss based on the second model, and updating the first model using the loss.

[0089] REFERENCE SIGNS LIST 10 Information processing device 11 Input unit 12 Base model update unit 13 Predicted loss calculation unit 14 Approximation model loss calculation unit 15 Approximation model calculation unit 100 Information processing device 101 CPU 102 ROM 103 RAM 104 Program group 105 Storage device 106 Drive device 107 Communication interface 108 Input / output interface 109 Bus 110 Storage medium 111 Communication network 121 Generation unit 122 Loss setting unit 123 Update unit

Claims

1. An information processing device comprising: a generation unit that generates a second model that approximates a prediction for a predetermined case made by a first model that is a machine learning model; a loss setting unit that sets a loss based on the second model; and an update unit that updates the first model using the loss.

2. An information processing device according to claim 1, wherein the loss setting unit sets the loss based on the prediction of the second model for the specified case, and the update unit updates the first model so as to reduce the loss.

3. An information processing device according to claim 2, wherein the loss setting unit sets the loss based on an error between the prediction of the first model and the prediction of the second model for the specified case, and the update unit updates the first model so as to reduce the loss.

4. An information processing device according to claim 1, wherein the loss setting unit sets the loss based on the weight of the second model, and the update unit updates the first model so that the loss becomes smaller.

5. An information processing device according to claim 4, wherein the loss setting unit sets the loss based on an absolute value of a weight of the second model, and the update unit updates the first model so that the loss becomes smaller.

6. An information processing device according to claim 1, wherein the loss setting unit sets the loss based on the prediction of the second model for the specified case and a weight of the second model, and the update unit updates the first model so as to reduce the loss.

7. An information processing device according to claim 1, wherein the generation unit generates the second model that approximates a prediction by the first model for the specified case including a nearby case located in the vicinity of a training case of the first model.

8. An information processing device according to claim 7, wherein the generation unit generates the second model, which is a linear model that approximates a prediction by the first model for the predetermined example including the neighboring example.

9. An information processing method comprising: generating a second model that approximates a prediction for a given case made by a first model that is a machine learning model; setting a loss based on the second model; and updating the first model using the loss.

10. An information processing method according to claim 9, further comprising: setting the loss based on the prediction of the second model for the specified case; and updating the first model so that the loss becomes smaller.

11. An information processing method according to claim 10, further comprising: setting the loss based on the error between the prediction of the first model and the prediction of the second model for the specified case; and updating the first model so that the loss becomes smaller.

12. The information processing method according to claim 9, further comprising: setting the loss based on the weight of the second model; and updating the first model so that the loss becomes smaller.

13. An information processing method according to claim 12, further comprising: setting the loss based on the absolute value of the weight of the second model; and updating the first model so that the loss becomes smaller.

14. An information processing method according to claim 9, further comprising generating the second model that approximates the prediction by the first model for the given example, which includes a neighboring example located in the vicinity of the training example of the first model.

15. A computer-readable storage medium storing a program that causes a computer to execute the following processes: generating a second model that approximates a prediction for a given case made by a first model that is a machine learning model; setting a loss based on the second model; and updating the first model using the loss.

Images