Information processing device, information processing system, information processing method, and program

Abstract

This information processing device is provided with: an acquisition unit that acquires first data; a first generation unit that generates second data by adding a prompt to the first data; a calculation unit that refers to the output of an inference model to which the second data is input and in which a parameter perturbation is added to at least one of a plurality of parameters, and thereby calculates a loss value of the inference model; and a second generation unit that refers to the loss value of the inference model, and generates at least one of the prompt and a prompt generation model that generates the prompt, so as to improve the flatness of the loss value.

Description

Information processing device, information processing system, information processing method, and program

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

[0002] In inference processing using a machine learning model, improving generalization performance is known to be one of the challenges. For example, Patent Literature 1 discloses a technology for improving the generalization performance of classification processing by generating augmented training data from training data for first training and performing second training using the augmented training data in a training method for a machine learning model in which a core tensor is generated.

[0003] Japanese Patent Application Publication No. 2020-119044

[0004] Even if the technology described in Patent Document 1 is used, it is difficult to obtain a model with sufficient generalization performance.

[0005] The present disclosure has been made in consideration of the above-mentioned problems, and one exemplary purpose thereof is to provide a technology that can suitably improve the generalization performance of inference processing using a machine learning model.

[0006] An information processing device according to an exemplary aspect of the present disclosure comprises an acquisition means for acquiring first data, a first generation means for generating second data by adding a prompt to the first data, a calculation means for calculating a loss value of an inference model to which the second data is input by referring to the output of the inference model in which a parameter perturbation has been added to at least one of a plurality of parameters possessed by the inference model, and a second generation means for generating at least one of the prompt and a prompt generation model that generates the prompt by referring to the loss value of the inference model so as to improve the flatness of the loss value.An information processing device according to an exemplary aspect of the present disclosure comprises an acquisition means for acquiring first data, a generation means for generating second data by adding a prompt to the first data, and an inference means for performing inference processing by inputting the second data into the inference model, and the prompt is a prompt generated by referring to the output of the inference model to which the second data is input so as to improve the flatness of the loss value obtained.

[0007] An information processing method according to one exemplary aspect of the present disclosure includes acquiring first data, generating second data by adding a prompt to the first data, calculating a loss value of an inference model to which the second data is input by referring to the output of the inference model in which a parameter perturbation has been added to at least one of a plurality of parameters possessed by the inference model, and generating the prompt and at least one of a prompt generation model that generates the prompt by referring to the loss value of the inference model so as to improve the flatness of the loss value.

[0008] An information processing method according to one exemplary aspect of the present disclosure includes acquiring first data, generating second data by adding a prompt to the first data, and performing inference processing by inputting the second data into an inference model, wherein the prompt is generated so as to improve the flatness of a loss value obtained by referring to the output of the inference model to which the second data is input.

[0009] In addition, the information processing device according to each aspect may be realized by a computer. In this case, a program for realizing the information processing device on a computer by causing the computer to operate as each means provided in the information processing device, and a computer-readable recording medium on which the program is recorded, also fall within the scope of the present invention.

[0010] According to one exemplary aspect of the present disclosure, an exemplary effect is achieved in that the generalization performance of inference processing using a machine learning model can be suitably improved.

[0011] FIG. 1 is a block diagram showing a configuration of an information processing device according to the present disclosure. FIG. 2 is a flow diagram showing a flow of an information processing method according to the present disclosure. FIG. 3 is a block diagram showing a configuration of an information processing device according to the present disclosure. FIG. 4 is a flow diagram showing a flow of an information processing method according to the present disclosure. FIG. 5 is a block diagram showing a configuration of an information processing device according to the present disclosure. FIG. 6 is a diagram for explaining processing by an information processing device according to the present disclosure. FIG. 7 is a diagram for explaining processing by an information processing device according to the present disclosure. FIG. 8 is a block diagram showing a configuration of an information processing device according to the present disclosure. FIG. 9 is a block diagram showing a configuration of a computer that functions as an information processing device according to the present disclosure.

[0012] The following are examples of embodiments of the present invention. However, the present invention is not limited to the exemplary embodiments shown below, and various modifications are possible within the scope of the claims. For example, embodiments obtained by appropriately combining the technical means employed in the exemplary embodiments shown below may also be included in the scope of the present invention. Furthermore, embodiments obtained by appropriately omitting some of the technical means employed in the exemplary embodiments shown below may also be included in the scope of the present invention. Furthermore, the effects mentioned in the exemplary embodiments shown below are examples of effects expected in the exemplary embodiments, and do not define the scope of the present invention. In other words, embodiments that do not exhibit the effects mentioned in the exemplary embodiments shown below may also be included in the scope of the present invention.

[0013] [First Exemplary Embodiment] A first exemplary embodiment, which is an example of an embodiment of the present invention, will be described in detail with reference to the drawings. This exemplary embodiment is a basic form for each of the exemplary embodiments described below. Note that the scope of application of each technical means employed in this exemplary embodiment is not limited to this exemplary embodiment. That is, each technical means employed in this exemplary embodiment can also be employed in other exemplary embodiments included in the present disclosure, to the extent that no particular technical obstacles arise. Furthermore, each technical means shown in the drawings referenced to explain this exemplary embodiment can also be employed in other exemplary embodiments included in the present disclosure, to the extent that no particular technical obstacles arise.

[0014] (Configuration of information processing device 1) The configuration of the information processing device 1 according to this exemplary embodiment will be described below with reference to Fig. 1. Fig. 1 is a block diagram showing the configuration of the information processing device 1 according to this exemplary embodiment. As shown in Fig. 1, the information processing device 1 includes an acquisition unit 11, a first generation unit 12, a calculation unit 13, and a second generation unit 14.

[0015] (Acquisition Unit 11) The acquisition unit 11 acquires first data. The first data is data that is the target of learning processing and inference processing in the information processing device 1, and may be expressed as data for learning or data for inference, for example. The type of the first data is not limited to this exemplary embodiment, and may be image data, text data (natural language data), sensing data obtained by a microphone or a sensor, or other data.

[0016] (First Generator 12) The first generator 12 generates second data by adding a prompt to the first data acquired by the acquirer 11. Here, the term "prompt" refers to data of any format that is added to the first data, and the present exemplary embodiment is not limited by this term.

[0017] As an example, if the first data is an image, the prompt may be a set of pixel values ​​that are superimposed on at least a portion of the image represented by the first data, or may be a set of pixel values ​​that are added to the outside of the image.

[0018] Furthermore, when the first data is text data (natural language data), the prompt may be a character string or a symbol string added serially to the front or back of the text data. Furthermore, when the first data is sensing data, the prompt may be a numeric string added serially to the front or back of the sensing data, or a numeric string added superimposed on the sensing data.

[0019] The prompt added by the first generator 12 may be generated (updated) using a prompt generation model having one or more parameters, or may be generated (updated) without using such a model. Furthermore, the prompt added by the first generator 12 may be obtained by combining a prompt generated (updated) using a prompt generation model and a prompt generated (updated) without using a prompt generation model.

[0020] Furthermore, in this exemplary embodiment, the degrees of freedom of the prompt added by the first generation unit 12 (e.g., the total number of data points that the prompt has) and the degrees of freedom of the prompt generation model (e.g., the total number of parameters that the prompt generation model has) are smaller than the degrees of freedom of the inference model described below (e.g., the total number of parameters that the inference model has).

[0021] (Calculation unit 13) The calculation unit 13 is an inference model to which the second data generated by the first generation unit 12 is input, and calculates the loss value of the inference model by referring to the output of the inference model to which a parameter perturbation has been added to at least one of the multiple parameters possessed by the inference model.

[0022] As an example, the calculation unit 13 performs the following processing: inputting second data generated by the first generation unit 12 into an inference model having multiple parameters; obtaining the output of the inference model as a pre-perturbation output; calculating a pre-perturbation loss value by inputting the pre-perturbation output into a loss function; applying a perturbation (the perturbation in this process is also called a "parameter perturbation") to at least one of multiple parameters of the inference model, and then inputting the second data generated by the first generation unit 12 into the inference model; obtaining the output of the inference model to which the perturbation has been applied as a post-perturbation output; and calculating a post-perturbation loss value by inputting the post-perturbation output into a loss function.

[0023] (Second generation unit 14) The second generation unit 14 generates the prompt and / or a prompt generation model that generates the prompt by referring to the loss value of the inference model so as to improve the flatness of the loss value.

[0024] As an example, the second generation unit 14 generates (updates) at least one of the prompt and the parameters of a prompt generation model that generates the prompt, by referring to the pre-perturbation loss value and the post-perturbation loss value, so as to improve the flatness of the loss values ​​of the inference model. Here, "flatness" can be expressed, for example, as the flatness of a graph in which the horizontal axis represents the values ​​of the parameters of the inference model and the vertical axis represents the loss values, but this does not limit the present exemplary embodiment. The "flatness" may be expressed as "small changes in the loss values ​​when the parameters of the inference model fluctuate (for example, small fluctuations)," or as "insensitivity of the loss values ​​to fluctuations (for example, small fluctuations) of the parameters of the inference model." Furthermore, the above-mentioned "flatness" may be expressed as "flatness of the loss values," "flatness of the loss function," "flatness of the inference model," "flatness of the inference process," or the like.

[0025] (Effects of information processing device 1) As described above, the information processing device 1 according to this exemplary embodiment adopts the following configuration: - acquire first data; - generate second data by adding a prompt to the first data; - calculate a loss value of an inference model to which the second data is input, by referring to the output of an inference model in which a parameter perturbation has been added to at least one of a plurality of parameters possessed by the inference model; - generate (update) the prompt and at least one of the prompt generation models that generate the prompt, by referring to the loss value of the inference model, so as to improve the flatness of the loss value.

[0026] Furthermore, as described above, in the information processing device 1, the degrees of freedom of the prompt added by the first generation unit 12 (e.g., the total number of data points that the prompt has) and the degrees of freedom of the prompt generation model (e.g., the total number of parameters that the prompt generation model has) are smaller than the degrees of freedom of the above-mentioned inference model (e.g., the total number of parameters that the inference model has).

[0027] Therefore, with the above configuration, it is possible to achieve improved flatness while suppressing increases in computational costs (computational resources and computational time) compared to configurations that attempt to improve flatness by generating (updating) parameters of an inference model.

[0028] Furthermore, according to the inventor's findings, improving the flatness of the loss value improves the generalization performance. Therefore, the information processing device 1 configured as described above can improve the generalization performance while suppressing an increase in the calculation cost (calculation resources and calculation time).

[0029] (Flow of Information Processing Method S1) Next, the flow of the information processing method S1 according to this exemplary embodiment will be described with reference to Fig. 2. Fig. 2 is a flow diagram showing the flow of the information processing method S1. As shown in Fig. 2, the information processing method S1 includes a process (step, process) S11 of acquiring first data, a process (step, process) S12 of generating second data by adding a prompt to the first data, a process (step, process) S13 of calculating a loss value, and a process (step, process) S14 of generating (updating) a prompt or a prompt generation model.

[0030] (Step S11) In step S11, the acquisition unit 11 acquires first data. Specific examples of the first data have been described above, and therefore will not be described here.

[0031] (Step S12) Subsequently, in step S12, the first generator 12 generates second data by adding a prompt to the first data acquired by the acquisition unit 11 in step S11. The specific processing by the first generator 12 has been described above, and therefore will not be described here. Note that the degrees of freedom of the prompt added by the first generator 12 in step S12 (e.g., the total number of data points included in the prompt) and the degrees of freedom of the prompt generation model (e.g., the total number of parameters included in the prompt generation model) are smaller than the degrees of freedom of the inference model according to this exemplary embodiment (e.g., the total number of parameters included in the inference model).

[0032] (Step S13) Subsequently, in step S13, the calculation unit 13 calculates a loss value of an inference model, which is an inference model to which the second data generated by the first generation unit 12 in step S12 is input, by referring to the output of the inference model in which a parameter perturbation has been added to at least one of a plurality of parameters of the inference model. The specific processing by the calculation unit 13 has been described above, and therefore a description thereof will be omitted here.

[0033] (Step S14) Subsequently, in step S14, the second generation unit 14 generates (updates) at least one of the prompt and the prompt generation model that generates the prompt, by referring to the loss value of the inference model calculated in step S13, so as to improve the flatness of the loss value. The specific processing by the second generation unit 14 has been described above, and therefore will not be described here. After the processing of step S14, for example, it is determined whether a predetermined termination condition is satisfied (for example, whether the flatness satisfies a predetermined condition). If the predetermined termination condition is satisfied, the information processing method S1 ends. If not, the process returns to step S12, where second data is generated using the generated (updated) prompt, and the process proceeds to step S13.

[0034] (Effects of information processing method S1) As described above, the information processing method S1 according to this exemplary embodiment adopts the following configuration: - Acquire first data; - Generate second data by adding a prompt to the first data; - Calculate a loss value of an inference model to which the second data is input, by referring to the output of an inference model in which a parameter perturbation has been added to at least one of a plurality of parameters possessed by the inference model; - Generate (update) the prompt and at least one of the prompt generation models that generate the prompt, by referring to the loss value of the inference model, so as to improve the flatness of the loss value.

[0035] Furthermore, as described above, in the information processing method S1, the degrees of freedom of the prompt added by the first generation unit 12 in step S12 (e.g., the total number of data points that the prompt has) and the degrees of freedom of the prompt generation model (e.g., the total number of parameters that the prompt generation model has) are smaller than the degrees of freedom of the inference model (e.g., the total number of parameters that the inference model has).

[0036] Therefore, with the above configuration, it is possible to achieve improved flatness while suppressing increases in computational costs (computational resources and computational time) compared to configurations that attempt to improve flatness by generating (updating) parameters of an inference model.

[0037] Furthermore, according to the inventor's findings, improving the flatness of the loss value improves the generalization performance. Therefore, according to the information processing method S1 configured as described above, it is possible to improve the generalization performance while suppressing an increase in the calculation cost (calculation resources and calculation time).

[0038] (Configuration of information processing device 2) The configuration of the information processing device 2 according to this exemplary embodiment will be described below with reference to Fig. 3. Fig. 3 is a block diagram showing the configuration of the information processing device 2 according to this exemplary embodiment. As shown in Fig. 3, the information processing device 2 includes an acquisition unit 21, a generation unit 22, and an inference unit 23.

[0039] (Acquisition unit 21) The acquisition unit 21 acquires first data. The first data is data that is the target of inference processing in the information processing device 2, and may be expressed as inference data, for example. As described in the information processing device 1, the type of the first data may be image data, text data (natural language data), sensing data obtained by a microphone or a sensor, or other data.

[0040] (Generation unit 22) The generation unit 22 generates second data by adding a prompt to the first data acquired by the acquisition unit 21. Here, the prompt is a prompt generated (updated) to improve the flatness of the loss value obtained by referring to the output of the inference model to which the second data is input. In addition, in this exemplary embodiment, the degrees of freedom of the prompt added by the generation unit 22 (e.g., the total number of data points included in the prompt) and the degrees of freedom of the prompt generation model (e.g., the total number of parameters included in the prompt generation model) are smaller than the degrees of freedom of the inference model according to this exemplary embodiment (e.g., the total number of parameters included in the inference model). As an example, the prompt is a prompt generated (updated) by the generation (updating) process performed by the second generation unit 14 of the information processing device 1 described above.

[0041] (Inference unit 23) The inference unit 23 performs inference processing by inputting the second data generated by the generation unit 22 into the inference model. The result of the inference processing by the inference unit 23 is presented to the user via output means (not shown), for example.

[0042] (Effects of information processing device 2) As described above, the information processing device 2 according to this exemplary embodiment has the following configuration: - acquire first data; - generate second data by adding a prompt to the first data; - perform inference processing by inputting the second data into an inference model; - the prompt is generated (updated) so as to improve the flatness of the loss value obtained by referring to the output of the inference model into which the second data is input.

[0043] Furthermore, as described above, in the information processing device 2, the degrees of freedom of the prompt added by the generation unit 22 (e.g., the total number of data points that the prompt has) and the degrees of freedom of the prompt generation model (e.g., the total number of parameters that the prompt generation model has) are smaller than the degrees of freedom of the above-mentioned inference model (e.g., the total number of parameters that the inference model has).

[0044] Therefore, with the above configuration, it is possible to achieve improved flatness while suppressing increases in computational costs (computational resources and computational time) compared to configurations that attempt to improve flatness by generating (updating) parameters of an inference model.

[0045] Furthermore, according to the inventor's findings, improving the flatness of the loss values ​​improves the generalization performance. Therefore, the information processing device 2 configured as described above can execute an inference process with favorably improved generalization performance while suppressing an increase in the calculation cost (calculation resources and calculation time).

[0046] (Flow of Information Processing Method S2) Next, the flow of the information processing method S2 according to this exemplary embodiment will be described with reference to Fig. 4. Fig. 4 is a flow diagram showing the flow of the information processing method S2. As shown in Fig. 4, the information processing method S2 includes a process (step, process) S21 of acquiring first data, a process (step, process) S22 of generating second data by adding a prompt to the first data, and a process (step, process) S23 of executing inference.

[0047] (Step S21) In step S21, the acquisition unit 21 acquires the first data. Specific examples of the first data have been described above, and therefore will not be described here.

[0048] (Step S22) Subsequently, in step S22, the generation unit 22 generates second data by adding a prompt to the first data acquired by the acquisition unit 21 in step S21. Here, the prompt is a prompt generated (updated) so as to improve the flatness of the loss value obtained by referring to the output of the inference model to which the second data is input. Note that the degrees of freedom of the prompt added by the generation unit 22 in step S22 (e.g., the total number of data points possessed by the prompt) and the degrees of freedom of the prompt generation model (e.g., the total number of parameters possessed by the prompt generation model) are smaller than the degrees of freedom of the inference model according to this exemplary embodiment (e.g., the total number of parameters possessed by the inference model). As an example, the prompt is a prompt generated (updated) by the generation (updating) process in step S14 of the information processing device S1 described above.

[0049] (Step S23) In step S23, the inference unit 23 performs inference processing by inputting the second data generated by the generation unit 22 in step S22 into the inference model. The result of the inference processing by the inference unit 23 in step S23 is presented to the user via output means (not shown), for example.

[0050] (Effects of information processing method S2) As described above, the information processing method S2 according to this exemplary embodiment has the following configuration: - Acquire first data; - Generate second data by adding a prompt to the first data; - Perform inference processing by inputting the second data into an inference model; - The prompt is generated (updated) so as to improve the flatness of the loss value obtained by referring to the output of the inference model into which the second data is input.

[0051] Furthermore, as described above, in the information processing method S2, the degrees of freedom of the prompt added by the generation unit 22 in step S22 (e.g., the total number of data points that the prompt has) and the degrees of freedom of the prompt generation model (e.g., the total number of parameters that the prompt generation model has) are smaller than the degrees of freedom of the inference model (e.g., the total number of parameters that the inference model has).

[0052] Therefore, with the above configuration, it is possible to achieve improved flatness while suppressing increases in computational costs (computational resources and computational time) compared to configurations that attempt to improve flatness by generating (updating) parameters of an inference model.

[0053] Furthermore, according to the inventor's findings, improving the flatness of the loss values ​​improves the generalization performance. Therefore, according to the information processing method S2 configured as described above, it is possible to execute an inference process with favorably improved generalization performance while suppressing an increase in the calculation cost (computational resources and calculation time).

[0054] Second Exemplary Embodiment A second exemplary embodiment, which is an example of an embodiment of the present invention, will be described in detail with reference to the drawings. Components having the same functions as those described in the above exemplary embodiment will be denoted by the same reference numerals, and their description will be omitted as appropriate. The scope of application of each technical means employed in this exemplary embodiment is not limited to this exemplary embodiment. That is, each technical means employed in this exemplary embodiment can also be employed in other exemplary embodiments included in the present disclosure, to the extent that no particular technical hindrance occurs. Furthermore, each technical means shown in each drawing referenced to describe this exemplary embodiment can also be employed in other exemplary embodiments included in the present disclosure, to the extent that no particular technical hindrance occurs.

[0055] (Configuration of Information Processing Device 1A) The configuration of the information processing device 1A according to this exemplary embodiment will be described with reference to Fig. 5. Fig. 5 is a block diagram showing the configuration of the information processing device 1A. As shown in Fig. 5, the information processing device 1A includes a control unit 10A, a storage unit 15A, a communication unit 16A, and an input / output unit 17A.

[0056] The communication unit 16A communicates with devices external to the information processing device 1A. The communication unit 16A transmits data supplied from the control unit 10A to the external devices, and supplies data received from the external devices to the control unit 10A. The information processing device 1A and the external devices are configured to be able to communicate with each other via a network, as an example. Here, the specific configuration of the network does not limit the present exemplary embodiment, but examples include a wireless local area network (LAN), a wired LAN, a wide area network (WAN), a public line network, a mobile data communication network, or a combination of these networks.

[0057] The input / output unit 17A is configured to include at least one of input / output devices such as a keyboard, a mouse, a display, a printer, and a touch panel. Alternatively, the input / output unit 17A may be configured to have input / output devices such as a keyboard, a mouse, a display, a printer, and a touch panel connected to it. In this configuration, the input / output unit 17A accepts various types of information input to the information processing device 1A from the connected input devices. Furthermore, the input / output unit 17A outputs various types of information to the connected output devices under the control of the control unit 10A. An example of the input / output unit 17A is an interface such as a USB (Universal Serial Bus).

[0058] (Storage unit 15A) The storage unit 15A stores various types of data referenced by the control unit 10A and various types of data generated by the control unit 10A. As an example, the storage unit 15A stores: First data D1; Prompt PR; Second data D2; Inference model MD (for example, multiple parameters that the inference model MD has); Loss function LF (for example, one or multiple parameters that define the loss function). Details of the first data D1, the prompt PR, the second data D2, the inference model MD, and the loss function LF will be described later.

[0059] The storage unit 15A may also be configured to store: a prompt generation model PGM; and an additional model AM. The prompt generation model PGM and the additional model AM will be described in detail later.

[0060] 5, the control unit 10A includes an acquisition unit 11, a first generation unit 12, a calculation unit 13, a second generation unit 14, and an output unit 15. The acquisition unit 11, the first generation unit 12, the calculation unit 13, and the second generation unit 14 have the functions described in the first exemplary embodiment as an example, but are not limited to this.

[0061] (Acquisition unit 11) The acquisition unit 11 acquires first data D1. As in the first exemplary embodiment, the first data D1 is data that is the subject of learning processing and inference processing in the information processing device 1A. For example, the first data D1 may be expressed as data for learning or data for inference. Furthermore, the type of the first data D1 is not limited to this exemplary embodiment, and may be image data, text data (natural language data), sensing data obtained by a microphone or a sensor, or other data. In this exemplary embodiment, the first data D1 may also be expressed as first data x, etc.

[0062] (First Generator 12) The first generator 12 generates second data D2 by adding a prompt PR to the first data D1 acquired by the acquirer 11. Here, the prompt PR refers to data of any format that is added to the first data, and the present exemplary embodiment is not limited by this term.

[0063] As an example, if the first data D1 is an image, as in exemplary embodiment 1, the prompt PR may be a set of pixel values ​​that are superimposed on at least a portion of the image represented by the first data D1, or may be a set of pixel values ​​that are added outside the image.

[0064] Furthermore, when the first data D1 is text data (natural language data), the prompt PR may be a character string or a symbol string added serially to the front or back of the text data. Furthermore, when the first data D1 is sensing data, the prompt may be a numeric string added serially to the front or back of the sensing data, or a numeric string added superimposed on the sensing data.

[0065] As in the first exemplary embodiment, the prompt PR added by the first generator 12 may be generated (updated) by a prompt generation model PGM having one or more parameters, or may be generated (updated) without using the prompt generation model PGM. Furthermore, the prompt PR added by the first generator 12 may be a prompt obtained by combining a prompt generated (updated) by the prompt generation model PGM with a prompt generated (updated) without using the prompt generation model PGM.

[0066] Furthermore, in this exemplary embodiment, the degrees of freedom of the prompt PR added by the first generator 12 (e.g., the total number of data points that the prompt PR has) and the degrees of freedom of the prompt generation model PGM (e.g., the total number of parameters that the prompt generation model PGM has) are smaller than the degrees of freedom of the inference model MD according to this exemplary embodiment (e.g., the total number of parameters that the inference model MD has). Note that in this exemplary embodiment, the prompt PR may also be expressed as a prompt p, etc. Furthermore, the second data D2 may also be expressed as second data x+p, etc.

[0067] 6 is a diagram illustrating an example of generation of second data D2 by the first generator 12. Fig. 6 shows, as an example, a case where the first data D1 is image data. As shown in Fig. 6, as an example, the first generator 12 generates second data D2 by superimposing a prompt PR, which is a set of pixel values, on the outer edge of the image represented by the first data D1.

[0068] (Calculation unit 13) The calculation unit 13 is an inference model MD to which the second data D2 generated by the first generation unit 12 is input, and calculates the loss value of the inference model MD by referring to the output of the inference model MD to which a parameter perturbation has been added to at least one of the multiple parameters possessed by the inference model MD.

[0069] As an example, the calculation unit 13 performs the following processing: inputting the second data D2 generated by the first generation unit 12 into an inference model MD having a plurality of parameters; obtaining the output of the inference model MD as a pre-perturbation output; calculating a pre-perturbation loss value by inputting the pre-perturbation output into a loss function LF; adding a perturbation (the perturbation in this process is also called a "parameter perturbation") to at least one of a plurality of parameters of the inference model MD, and then inputting the second data D2 generated by the first generation unit 12 into the inference model MD; obtaining the output of the inference model MD to which the perturbation has been applied as a post-perturbation output; and calculating a post-perturbation loss value by inputting the post-perturbation output into the loss function LF.

[0070] (Second generation unit 14) The second generation unit 14 generates (updates) at least one of the prompt PR and the prompt generation model PGM that generates the prompt PR by referring to the loss value of the inference model MD so as to improve the flatness of the loss value.

[0071] As an example, the second generation unit 14 generates (updates) at least one of the prompt PR and the parameters of the prompt generation model PGM that generates the prompt PR by referring to the pre-perturbation loss value and the post-perturbation loss value so as to improve the flatness of the loss values ​​of the inference model. Here, as in the first exemplary embodiment, "flatness" can be expressed, for example, as the flatness of a graph in which the horizontal axis represents the values ​​of the parameters of the inference model and the vertical axis represents the loss values, but this does not limit the present exemplary embodiment. The "flatness" may be expressed as "small changes in the loss values ​​when the parameters of the inference model fluctuate (for example, small fluctuations)" or "insensitivity of the loss values ​​to fluctuations of the parameters of the inference model (for example, small fluctuations)." Furthermore, the above-mentioned "flatness" may be expressed as "flatness of the loss values," "flatness of the loss function," "flatness of the inference model," "flatness of the inference process," or the like.

[0072] 7 is a diagram for explaining an example of processing by the second generating unit 14. In FIG. 7, x is the first data D1, y is the correct label for the first data, and F θ (x) denotes the inference model MD obtained by inputting first data x into the inference model MD having a parameter θ, or the output of the inference model MD. L denotes the loss function LF. The loss function L is a function that expresses the difference between the output of the inference model MD and the correct label y for that output. For example, the greater the difference between x and y, the larger the value of the loss function L.

[0073] The upper part of Fig. 7 shows a process according to a comparative example. In the process according to the comparative example, The output F of the inference model MD with perturbation ε added to the parameter θ is θ＋ε (x) and the correct label y. θ＋ε (x), y) and the output F of the inference model MD without adding perturbation ε to the parameter θ. θ(x) and the correct label y. θ (x), y) and the difference L(F θ＋ε (x),y)-L(F θ (x), y) is maximized by determining the perturbation ε, and the difference of the loss function determined by the determined ε (in other words, the fluctuation of the loss value due to the perturbation ε) L(F θ＋ε (x),y)-L(F θ The parameter θ of the inference model MD is determined (updated) so as to minimize (x, y).

[0074] In this way, in the processing of the comparative example, the parameter θ is determined (updated) so as to minimize the fluctuation of the loss value, which is the value of the loss function L, in response to the perturbation ε to the parameter θ of the inference model MD (in other words, so as to improve the flatness of the loss value).

[0075] In the upper part of FIG. 7, the fluctuation ΔL of the loss value at step t ｔ However, by going through the above process, the fluctuation ΔL of the loss value at step t+1 ｔ＋１ , and the flatness of the loss value is improved.

[0076] In this comparative example, the parameter θ of the inference model MD is updated in order to improve the flatness of the loss values. Generally, the inference model MD has a large number of parameters (in other words, there are a large number of the above-mentioned parameters θ), and therefore, the process of improving the flatness of the loss values ​​by updating the parameters of the inference model MD as described above requires a large computational cost (computational resources and computational time).

[0077] On the other hand, the lower part of Fig. 7 shows a processing example by the information processing device 1A according to this exemplary embodiment. In the lower part of Fig. 7, p indicates a prompt added to the first data D1 (first data x), and x + p indicates second data D2 (second data x + p) obtained by adding prompt p to the first data D1. The loss function L according to this exemplary embodiment is also a function that expresses the difference between the output of the inference model MD and the correct label y for that output. As an example, the greater the difference between x and y, the greater the value of the loss function L, but this does not limit this exemplary embodiment. In the processing example according to this exemplary embodiment, First, the calculation unit 13 inputs the second data x+p to the inference model MD to which the perturbation ε is added to the parameter θ, and calculates the output F θ＋ε (x + p) and the correct label y. θ＋ε Furthermore, the calculation unit 13 calculates the output F of the inference model MD obtained by inputting the second data x+p to the inference model MD in which the perturbation ε is not added to the parameter θ. θ (x + p) and the correct label y. θ (x), y) are calculated.

[0078] Here, L(F θ＋ε (x+p), y) corresponds to the above-mentioned "perturbation loss value", and L(F θ (x+p), y) corresponds to the above-mentioned "pre-perturbation loss value", for example. Subsequently, the second generator 14 calculates the difference L(F θ＋ε (x+p),y)-L(F θ (x+p), y) is determined so as to maximize the difference in the loss function defined by the determined ε (in other words, the fluctuation in the loss value due to the perturbation ε) L(F θ＋ε (x+p),y)-L(F θ At least one of the prompt p and the prompt generation model PGM that generates the prompt p is determined (updated) so as to minimize (x+p), y).

[0079] The lower part of Figure 7 shows that the fluctuation in loss value ΔL(x+p) by the inference model MD with the second data x+p as input is smaller than the fluctuation in loss value ΔL(x) by the inference model MD with the first data x as input (in other words, the flatness is improved).

[0080] Thus, in this exemplary embodiment, in order to improve the flatness of the loss value, the parameter θ of the inference model MD is not generated (updated), but rather the prompt p and / or the prompt generation model PGM that generates the prompt p are generated (updated). As described above, the degrees of freedom of the prompt p (e.g., the total number of data points of the prompt p) and the degrees of freedom of the prompt generation model PGM that generates the prompt p (e.g., the total number of parameters that the prompt generation model PGM has) are smaller than the degrees of freedom of the inference model MD according to this exemplary embodiment (e.g., the total number of parameters that the inference model MD has).

[0081] Furthermore, according to the inventor's findings, improving the flatness of the loss values ​​improves generalization performance. Therefore, the information processing device 1A that executes the processing according to this exemplary embodiment can reduce computational costs (computational resources and computation time) compared to a configuration that improves the flatness of the loss values ​​by generating (updating) parameters of an inference model.

[0082] It is generally known that a model trained by machine learning may suffer from a degradation in performance due to domain shift. Here, domain shift refers, for example, to a difference in data distribution (domain) between training data (learning data) and predicted data (inference data). To improve the versatility of a machine learning model, a technology is needed to improve the performance (generalization performance) of the machine learning model for data belonging to an unknown domain. The information processing device 1A described above can reduce computational costs (computational resources and computation time) while suitably improving the generalization performance of an inference process using a machine learning model (inference model DM) post-hoc.

[0083] (Output Unit 15) The output unit 15 presents each piece of data referenced by each unit included in the control unit 10A described above or at least a portion of each piece of data derived by each unit included in the control unit 10A described above to the user via the input / output unit 17A or transmits it to an external device via the communication unit 16A. As an example, the output unit 15 may present the prompt p generated by the first generation unit 12 or the second data x+p to which the prompt is added to the user via the input / output unit 17A.

[0084] <Specific Processing Example by First Generator 12, Calculator 13, and Second Generator 14> A more specific processing example by the first generator 12, calculator 13, and second generator 14 will be described below. In the following example, as an example, assume that the acquisition unit 11 acquires a training dataset D is acquired, and the learning process is performed by the first generating unit 12, the calculating unit 13, and the second generating unit 14 by referring to the training dataset D. Note that in the following description, explanations of notations that have already been explained will be omitted.

[0085] (Processing Example 1) In the processing according to this example, the first generation unit 12 generates second data x+p by adding a prompt p to first data x, and the calculation unit 13 calculates an output F of the inference model MD to which the second data x+p generated by the first generation unit 12 is input, the output F being a parameter perturbation ε added to at least one of a plurality of parameters θ of the inference model MD. θ＋ε The loss value L(F θ＋ε (x+p), y) is calculated. The second generating unit 14 then refers to the fluctuation of the loss value given by the following formula, and generates (updates) the prompt p so as to improve the flatness of the loss value. Here, the parameter ρ is a parameter that defines the upper limit of the perturbation ε, and its specific value can be set appropriately depending on the dataset D, the allowable calculation cost, etc. Furthermore, as an example, a gradient descent method can be applied to the above calculation. Furthermore, when the first generation unit 12 sets the initial value of the prompt p, as an example, the initial value of the prompt p may be set using a random number.

[0086] In this way, in this example, the second generation unit 14 refers to the fluctuations in the loss value and generates (updates) the prompt p so as to improve the flatness of the loss value, thereby suppressing an increase in computational costs while suitably improving the generalization performance of the inference processing using the inference model MD.

[0087] (Processing Example 2) In the processing according to this example, the first generation unit 12 generates second data x+p by adding a prompt p to first data x, and the calculation unit 13 calculates an output F of the inference model MD to which the second data x+p generated by the first generation unit 12 is input, where the perturbation ε is not added to multiple parameters θ of the inference model MD. θ The loss value L(F θ (x+p), y). The calculation unit 13 also calculates the following: - The calculation unit 13 calculates an inference model MD to which the second data x+p generated by the first generation unit 12 is input, and for at least one of a plurality of parameters θ of the inference model MD, a parameter perturbation ε and a gradient term indicating the gradient of the loss function The loss value of the inference model MD to which Here, ∇ in the above formula represents an operator for calculating the gradient. The second generator 14 then refers to the fluctuation of the loss value given by the following formula, generates (updates) a prompt p so that the flatness of the loss value is improved, and generates (updates) model parameters so that the loss value becomes smaller. Here, the parameter ρ is a parameter that defines the upper limit of the perturbation ε, and its specific value can be set appropriately depending on the dataset D, the allowable calculation cost, etc. Furthermore, as an example, a gradient descent method can be applied to the above calculation. Furthermore, when the first generation unit 12 sets the initial value of the prompt p, as an example, the initial value of the prompt p may be set using a random number.

[0088] In this way, the loss values ​​calculated by the calculation unit 13 and referenced by the second generation unit 14 include: a first loss value (the first term in the above formula 7) determined according to the output of the inference model MD; and a second loss value (the second term in the above formula 7) determined according to the gradient of the first loss value.

[0089] In the above process, the model parameters generated (updated) by the second generating unit 14 are a part θ′ of the multiple parameters θ of the inference model MD. In addition to or instead of this, a part of the multiple parameters θ″ included in the additional model AM may be included. The former configuration can be expressed as a configuration in which the second generation unit 14 refers to the loss value of the inference model MD and further generates (updates) at least one of the multiple parameters θ (θ') possessed by the inference model MD.

[0090] Furthermore, in the latter configuration in which some of the multiple parameters θ'' held by the latter additional model AM are updated, the loss value L described above may be configured to include the loss value of the additional model AM in addition to, or instead of, the loss value of the inference model MD.

[0091] In other words, the calculation unit 13 may perform the following processing: - the calculation unit 13 is a model (additional model AM) to be added to the inference model MD, and further calculates a loss value of the additional model AM by referring to the output of the additional model AM to which the second data x+p is input; - the second generation unit 14 further generates (updates) at least one of one or more parameters θ'' possessed by the additional model AM by referring to at least one of the loss value of the inference model MD calculated by the calculation unit 13 and the loss value of the additional model AM.

[0092] FIG. 8 is a diagram showing an example of the configuration of an additional model AM referenced by an information processing device 1A performing processing according to this example. As shown in FIG. 8, as an example, the additional model AM is configured in parallel with the inference model MD, and input to the inference model MD is also input to the additional model AM. Here, the total number of parameters possessed by the additional model AM is smaller than the total number of parameters possessed by the inference model MD. As an example, the additional model AM shown in FIG. 8 is composed of: a first rank decomposition matrix A that generates r-dimensional (r<d) data from d-dimensional input data x; and a second rank decomposition matrix B that generates an output of the same dimension as the output of the inference model MD from the r-dimensional data. With this configuration, the total number of parameters possessed by the additional model AM can be made smaller than the total number of parameters possessed by the inference model MD. Furthermore, in the process of generating (updating) model parameters, the parameters of the inference model MD are frozen, and only the parameters of the additional model AM are generated (updated). By performing such processing, the inference accuracy of the inference model composed of the inference model MD and the additional model AM can be suitably improved.

[0093] In the processing of this example, as described above, the loss values ​​referenced by the second generation unit 14 include: a first loss value (corresponding to the first term in the above formula 7) determined according to the output of the inference model MD (and the additional model AM); and a second loss value (corresponding to the second term in the above formula 7) determined according to the gradient of the first loss value, and the second generation unit 14: generates (updates) a prompt p so as to improve the flatness of the loss values, and generates (updates) at least one of the parameter θ' of the inference model MD or the parameter θ'' of the additional model AM so as to reduce the loss value.

[0094] Therefore, according to the processing of this example, it is possible to suitably improve the generalization performance of the inference processing using the inference model MD (or an inference model composed of the inference model MD and the additional model AM) while suppressing an increase in computational costs, and it is also possible to suitably improve the inference accuracy using the inference model MD (or an inference model composed of the inference model MD and the additional model AM).

[0095] Furthermore, as described above, if the parameters of the additional model AM are generated (updated) instead of the parameters of the inference model MD, the calculation cost can be suitably reduced while suitably improving the inference accuracy of the inference model composed of the inference model MD and the additional model AM.

[0096] (Processing Example 3) In the processing according to this example, the first generation unit 12 generates second data x+p+δ by adding a prompt p and a data perturbation δ to the first data x, and the calculation unit 13 calculates an output F of the inference model MD to which the second data x+p+δ generated by the first generation unit 12 is input, the output F being obtained by adding a parameter perturbation ε to at least one of a plurality of parameters θ of the inference model MD. θ＋ε The loss value L(F θ＋ε(x+p+δ), y) is calculated. The second generation unit 14 then refers to the fluctuations in the loss value given by the following formula, and generates (updates) at least one of the prompt p and the multiple parameters of the inference model MD so as to improve the flatness of the loss value. Here, the parameters ρ and γ are parameters that define the upper limits of the perturbations ε and δ, respectively, and their specific values ​​can be set appropriately depending on the dataset D, the allowable calculation cost, etc. Furthermore, as an example, a gradient descent method can be applied to the above calculation. Furthermore, when the first generation unit 12 sets the initial value of the prompt p, as an example, the initial value of the prompt p may be set using a random number.

[0097] In this example, the following configuration is adopted: the first generator 12 generates second data x+p+δ by adding a prompt p and a data perturbation δ to the first data x, and the second generator 14 generates (updates) the prompt and at least one of the multiple parameters of the inference model. Here, the generation (update) process is performed, for example, to improve the flatness of the loss value (in other words, to improve robustness against the data perturbation).

[0098] Therefore, the above configuration has the following effects: It is possible to suitably improve the generalization performance of the inference process using the inference model MD while suppressing an increase in calculation cost, and It is also possible to improve the flatness of the space defined by the input data (input space). In other words, as an example, it is possible to construct an inference model MD that is more robust against adversarial perturbations to the input data.

[0099] (Processing Example 4) In the above-mentioned processing example 1, an example was described in which the second generation unit 14 generates (updates) a prompt p so as to improve the flatness of the loss value. However, in this example, instead of the process of generating (updating) the prompt in processing example 1, a prompt generation model PGM that generates the prompt p is generated (updated).

[0100] More specifically, the second generating unit 14 generates (updates) one or more parameters (^θ: hatched theta) of the prompt generation model PGM. As an example, the second generating unit 14 generates (updates) one or more parameters (^θ: hatched theta) of the prompt generation model PGM. Here, g ＾θ (x) indicates the output of the prompt generation model PGM to which input data x has been input. In this way, in this example, one or more parameters (^θ) of the prompt generation model PGM are generated (updated) so as to improve the flatness of the loss value. Regarding other processing related to this example, explanations of content similar to that of processing example 1 will be omitted.

[0101] As in this example, by generating (updating) the prompt generation model PGM, a wider variety of prompts p can be generated and added to the first data. Therefore, according to the processing of this example, it is possible to more suitably improve the generalization performance of the inference processing using the inference model MD while suppressing an increase in calculation cost.

[0102] (Processing Example 5) In the above-mentioned processing example 2, an example was described in which the second generation unit 14 generates (updates) a prompt p so as to improve the flatness of the loss value, and also generates (updates) model parameters so as to reduce the loss value. However, in this example, instead of the process of generating (updating) the prompt in processing example 2, a prompt generation model PGM that generates the prompt p is generated (updated).

[0103] More specifically, the second generating unit 14 generates (updates) one or more parameters (^θ: hatched theta) of the prompt generation model PGM. As an example, the second generating unit 14 generates (updates) one or more parameters (^θ: hatched theta) of the prompt generation model PGM. Here, g ＾θ(x) indicates the output of the prompt generation model PGM to which input data x has been input. In this way, in this example, one or more parameters (^θ) of the prompt generation model PGM are generated (updated) so as to improve the flatness of the loss value, and model parameters are generated (updated) so as to reduce the loss value. Regarding other processing related to this example, explanations of content similar to that of processing example 2 will be omitted.

[0104] As in this example, by generating (updating) the prompt generation model PGM, it is possible to generate a wider variety of prompts p and add them to the first data.

[0105] Therefore, according to the processing of this example, it is possible to more suitably improve the generalization performance of the inference processing using the inference model MD (or an inference model composed of the inference model MD and the additional model AM) while suppressing an increase in computational costs, and it is also possible to suitably improve the inference accuracy using the inference model MD (or an inference model composed of the inference model MD and the additional model AM).

[0106] (Processing Example 6) In the above-mentioned processing example 3, an example was described in which the second generation unit 14 generates (updates) at least one of the multiple parameters of the prompt p and the inference model MD so as to improve the flatness of the loss value.However, in this example, instead of the process of generating (updating) the prompt in processing example 3, a prompt generation model PGM that generates the prompt p is generated (updated).

[0107] More specifically, the second generating unit 14 generates (updates) one or more parameters (^θ: hatched theta) of the prompt generation model PGM. As an example, the second generating unit 14 generates (updates) one or more parameters (^θ: hatched theta) of the prompt generation model PGM. Here, g ＾θ(x) indicates the output of the prompt generation model PGM to which input data x has been input. In this way, in this example, one or more parameters (^θ) possessed by the prompt generation model PGM and at least one of the multiple parameters possessed by the inference model MD are generated (updated) so as to improve the flatness of the loss value. Regarding other processing related to this example, explanations of content similar to that of processing example 3 will be omitted.

[0108] As in this example, by generating (updating) the prompt generation model PGM, a wider variety of prompts p can be generated and added to the first data. Therefore, the above configuration provides the following advantages: It is possible to more suitably improve the generalization performance of the inference process using the inference model MD while suppressing an increase in computational cost, and It also provides improved flatness with respect to the space (input space) defined by the input data. In other words, as an example, it is possible to construct an inference model MD that is more robust against adversarial perturbations of the input data.

[0109] <Processing Example 7> The generation (update) process (learning process) of prompt p in the above-described processing examples 1 to 6 may be performed for each class. θ In other words, if the output of the inference model F is k classes (k is a natural number equal to or greater than 2), θ is a model that solves k-class problems. If so, the second generator 14 generates a prompt p More specifically, the first generating unit 12 may generate (update) (learn) a prompt p for each class. １ , ..., p ｋ The second data x+p for each class is obtained by adding the first data x to the １ , ..., x+p ｋa calculation unit 13 inputs the second data for each class into an inference model MD to which a parameter perturbation ε has been added to at least one of a plurality of parameters θ, and calculates a loss value of the inference model MD by referring to the output of the inference model MD; a second generation unit 14 calculates the loss value of the prompt p by referring to the loss value of the inference model MD so as to improve the flatness of the loss value. １ , ..., p ｋ and each of the prompts p １ , ..., p ｋ At least one of the prompt generation models PGM that generate the following is generated (updated).

[0110] Here, as described in processing examples 2, 3, 5, and 6, in addition to generating (updating) the prompt, a portion θ' of the parameters θ possessed by the inference model MD, or a parameter θ'' possessed by the additional model AM added to the inference model, may also be generated (updated).

[0111] Thus, in this example, the inference model MD is a model that makes inferences regarding classification problems into multiple classes, the prompt p is set for each class, and the second generation means generates (updates) at least one of the prompt p and the prompt generation model PGM that generates the prompt for each class.

[0112] According to the process of this example, by generating (updating) a prompt for each class individually, it is possible to increase the effect of improving the flatness of the loss value by the prompt.

[0113] In addition, the inference process (prediction process) using the inference model MD learned by the process according to this example may be configured to be executed by, for example, identifying the class for which the output of the inference model MD is the maximum. Here, denotes the score by the inference model MD for class i.

[0114] Furthermore, in the processing of this example, in addition to or instead of performing the generation (update) processing (learning processing) of prompt p for each class, the generation (update) processing (learning processing) of prompt p may be configured to be performed for each domain.

[0115] (Configuration of Information Processing Device 2A) Next, the configuration of the information processing device 2A according to this exemplary embodiment will be described with reference to Fig. 9. Fig. 9 is a block diagram showing the configuration of the information processing device 2A. As shown in Fig. 9, the information processing device 2A includes a control unit 20A, a storage unit 25A, a communication unit 16A, and an input / output unit 17A.

[0116] As shown in FIG. 9 , the control unit 20A includes an acquisition unit 21, a generation unit 22, an inference unit 23, and an output unit 15. Here, the acquisition unit 21, the generation unit 22, and the inference unit 23 have the functions described in the exemplary embodiment 1, as an example. Furthermore, the acquisition unit 21 and the generation unit 22 execute the same processes as the acquisition unit 21 and the generation unit 22 included in the information processing device 1A, as an example. However, the first data D1 acquired by the acquisition unit 21 is data that is the target of inference processing in the information processing device 2A, and may be expressed as data for inference, as an example. The type of the first data may be image data, text data (natural language data), sensing data obtained by a microphone or a sensor, or other data, as described for the information processing device 1A.

[0117] (Generation Unit 22) The generation unit 22 generates second data by adding a prompt PR to the first data D1 acquired by the acquisition unit 21. Here, the prompt is a prompt generated (updated) to improve the flatness of the loss value obtained by referring to the output of the inference model DM to which the second data D2 is input. Furthermore, in this exemplary embodiment, the degrees of freedom of the prompt PR added by the generation unit 22 (e.g., the total number of data points possessed by the prompt) and the degrees of freedom of the prompt generation model PGM (e.g., the total number of parameters possessed by the prompt generation model PGM) are smaller than the degrees of freedom of the inference model DM (e.g., the total number of parameters possessed by the inference model MD) according to this exemplary embodiment. As an example, the prompt PR is a prompt generated (updated) (learned) by the generation (update) process performed by the second generation unit 14 of the information processing device 1A described above. Furthermore, the inference model DM is an inference model generated (updated) (learned) by the generation (update) process performed by the second generation unit 14 of the information processing device 1A described above.

[0118] (Inference unit 23) The inference unit 23 performs inference processing by inputting the second data generated by the generation unit 22 into the inference model DM. As an example, the result of the inference processing by the inference unit 23 is presented to the user by the output unit 15 via the input / output unit 17A.

[0119] As an example, the inference unit 23 outputs F θ (x+p) and θ An inference result is derived according to the value of (x+p).

[0120] Furthermore, if the inference model DM is a model that solves the k-class problem, as described above in <Processing Example 7>, the inference unit 23 may be configured to derive the inference result by referring to (Equation 16).

[0121] According to the above configuration, it is possible to achieve improved flatness while suppressing increases in computational costs (computational resources and computational time) compared to configurations that attempt to improve flatness by generating (updating) parameters of an inference model.

[0122] Furthermore, according to the inventor's findings, improving the flatness of the loss values ​​improves the generalization performance. Therefore, the information processing device 2 configured as described above can execute an inference process with favorably improved generalization performance while suppressing an increase in the calculation cost (calculation resources and calculation time).

[0123] [Example of implementation by software] Some or all of the functions of the information processing devices 1, 1A, 2, 2A (hereinafter also referred to as "each of the above devices") may be implemented by hardware such as an integrated circuit (IC chip), or by software.

[0124] In the latter case, each of the above devices is realized by, for example, a computer that executes instructions of a program, which is software that realizes each function. An example of such a computer (hereinafter referred to as computer C) is shown in Figure 10. Figure 10 is a block diagram showing the hardware configuration of computer C that functions as each of the above devices.

[0125] The computer C includes at least one processor C1 and at least one memory C2. The memory C2 stores a program P for causing the computer C to function as each of the above-mentioned devices. In the computer C, the processor C1 reads and executes the program P from the memory C2, thereby realizing the functions of each of the above-mentioned devices.

[0126] The processor C1 may be, for example, a central processing unit (CPU), a graphics processing unit (GPU), a digital signal processor (DSP), a micro processing unit (MPU), a floating point number processing unit (FPU), a physics processing unit (PPU), a tensor processing unit (TPU), a quantum processor, a microcontroller, or a combination thereof. The memory C2 may be, for example, a flash memory, a hard disk drive (HDD), a solid state drive (SSD), or a combination thereof.

[0127] The computer C may further include a RAM (Random Access Memory) for expanding the program P during execution and for temporarily storing various data. The computer C may also include a communication interface for transmitting and receiving data to and from other devices. The computer C may also include an input / output interface for connecting input / output devices such as a keyboard, a mouse, a display, and a printer.

[0128] The program P can also be recorded on a non-transitory, tangible recording medium M that can be read by the computer C. Such a recording medium M can be, for example, a tape, a disk, a card, a semiconductor memory, or a programmable logic circuit. The computer C can acquire the program P via such a recording medium M. The program P can also be transmitted via a transmission medium. Such a transmission medium can be, for example, a communication network or broadcast waves. The computer C can also acquire the program P via such a transmission medium.

[0129] [Additional Notes] This disclosure includes the technologies described in the following supplementary notes. However, the present invention is not limited to the technologies described in the following supplementary notes, and various modifications are possible within the scope of the claims.

[0130] (Appendix A1) An information processing device comprising: an acquisition means for acquiring first data; a first generation means for generating second data by adding a prompt to the first data; a calculation means for calculating a loss value of an inference model to which the second data is input, by referring to the output of an inference model in which a parameter perturbation has been added to at least one of a plurality of parameters possessed by the inference model; and a second generation means for generating (updating) the prompt and at least one of a prompt generation model that generates the prompt by referring to the loss value of the inference model so as to improve the flatness of the loss value.

[0131] (Appendix A2) The information processing device according to Appendix A1, wherein the second generation means further generates at least one of a plurality of parameters possessed by the inference model by referring to a loss value of the inference model.

[0132] (Appendix A3) The information processing device described in Appendix A1, wherein the calculation means is a model added to the inference model, and further calculates a loss value of the additional model by referring to the output of the additional model to which the second data is input, and the second generation means further generates at least one of one or more parameters possessed by the additional model by referring to at least one of the loss value of the inference model calculated by the calculation means and the loss value of the additional model.

[0133] (Appendix A4) An information processing device described in any one of Appendices A1 to A3, wherein the loss value of the inference model calculated by the calculation means includes: a first loss value determined according to the output of the inference model; and a second loss value determined according to the gradient of the first loss value.

[0134] (Appendix A5) The information processing device described in any one of Appendices A1 to A4, wherein the generation means generates the second data by adding the prompt and a data perturbation to the first data, and the second generation means generates at least one of the prompt, a prompt generation model that generates the prompt, and a plurality of parameters that the inference model has.

[0135] (Appendix A6) An information processing device described in any one of Appendices A1 to A5, wherein the inference model is a model that makes inferences regarding classification problems into multiple classes, the prompt is set for each of the classes, and the second generation means generates at least one of the prompt and a prompt generation model that generates the prompt for each of the classes.

[0136] (Appendix A7) An information processing device comprising: an acquisition means for acquiring first data; a generation means for generating second data by adding a prompt to the first data; and an inference means for performing inference processing by inputting the second data into an inference model, wherein the prompt is generated so as to improve the flatness of the loss value obtained by referring to the output of the inference model to which the second data is input.

[0137] (Appendix A8) An information processing method comprising: acquiring first data; generating second data by adding a prompt to the first data; calculating a loss value of an inference model to which the second data is input, by referring to the output of the inference model in which a parameter perturbation has been added to at least one of a plurality of parameters possessed by the inference model; and generating the prompt and at least one of a prompt generation model that generates the prompt by referring to the loss value of the inference model so as to improve the flatness of the loss value.

[0138] (Appendix A9) An information processing method comprising: acquiring first data; generating second data by adding a prompt to the first data; and performing inference processing by inputting the second data into an inference model, wherein the prompt is generated so as to improve the flatness of the loss value obtained by referring to the output of the inference model to which the second data is input.

[0139] (Appendix A10) A program for causing a computer to function as an information processing device, the program causing the computer to perform the following operations: acquire first data; generate second data by adding a prompt to the first data; calculate a loss value of an inference model to which the second data is input, by referring to the output of an inference model in which a parameter perturbation has been added to at least one of a plurality of parameters possessed by the inference model; and generate the prompt and at least one of a prompt generation model that generates the prompt by referring to the loss value of the inference model so as to improve the flatness of the loss value.

[0140] (Appendix A11) A program for causing a computer to function as an information processing device, the program causing the computer to perform the following operations: acquiring first data; generating second data by adding a prompt to the first data; and performing inference processing by inputting the second data into an inference model, the prompt being a prompt generated to improve flatness of a loss value obtained by referring to an output of the inference model to which the second data is input.

[0141] REFERENCE SIGNS 1, 2, 1A, 2A ... Information processing device 11, 21 ... Acquisition unit 12, 22 ... Generation unit, first generation unit 13 ... Calculation unit 14 ... Second generation unit 23 ... Inference unit

Claims

1. An information processing device comprising: an acquisition means for acquiring first data; a first generation means for generating second data by adding a prompt to the first data; a calculation means for calculating a loss value of an inference model to which the second data is input, by referring to the output of the inference model in which a parameter perturbation has been added to at least one of a plurality of parameters possessed by the inference model; and a second generation means for generating at least one of the prompt and a prompt generation model that generates the prompt by referring to the loss value of the inference model so as to improve the flatness of the loss value.

2. The information processing device described in claim 1, wherein the second generation means further generates at least one of the multiple parameters possessed by the inference model by referring to the loss value of the inference model.

3. The information processing device described in claim 1, wherein the calculation means further calculates a loss value of an additional model that is added to the inference model and to which the second data is input by referring to the output of the additional model, and the second generation means further generates at least one of one or more parameters possessed by the additional model by referring to at least one of the loss value of the inference model calculated by the calculation means and the loss value of the additional model.

4. An information processing device described in any one of claims 1 to 3, wherein the loss value of the inference model calculated by the calculation means includes a first loss value determined according to the output of the inference model and a second loss value determined according to the gradient of the first loss value.

5. An information processing device described in any one of claims 1 to 4, wherein the first generation means generates the second data by adding the prompt and a data perturbation to the first data, and the second generation means generates at least one of the prompt, a prompt generation model that generates the prompt, and a plurality of parameters possessed by the inference model.

6. An information processing device described in any one of claims 1 to 5, wherein the inference model is a model that makes inferences regarding classification problems into multiple classes, the prompt is set for each of the classes, and the second generation means generates at least one of the prompt and a prompt generation model that generates the prompt for each of the classes.

7. An information processing device comprising: an acquisition means for acquiring first data; a generation means for generating second data by adding a prompt to the first data; and an inference means for performing inference processing by inputting the second data into an inference model, wherein the prompt is generated so as to improve the flatness of the loss value obtained by referring to the output of the inference model to which the second data is input.

8. An information processing method comprising: acquiring first data; generating second data by adding a prompt to the first data; calculating a loss value of an inference model to which the second data is input, by referring to the output of the inference model in which a parameter perturbation has been added to at least one of a plurality of parameters possessed by the inference model; and generating the prompt and / or at least one of a prompt generation model that generates the prompt by referring to the loss value of the inference model so as to improve the flatness of the loss value.

9. An information processing method comprising: acquiring first data; generating second data by adding a prompt to the first data; and performing inference processing by inputting the second data into an inference model, wherein the prompt is generated so as to improve the flatness of the loss value obtained by referring to the output of the inference model to which the second data is input.

10. A program for causing a computer to function as the information processing device of claim 1, the program causing a computer to function as the acquisition means, the first generation means, the calculation means, and the second generation means.

11. A program for causing a computer to function as the information processing device according to claim 7, the program causing a computer to function as said acquisition means, said generation means, and said inference means.

Images