Information processing apparatus, information processing method, and information processing system

Abstract

[Problem] To realize low-cost and efficient data operation related to optimization of a large-scale language model. [Solution] An information processing apparatus is provided, including: a learning unit that learns a large-scale language model including a learning target part and a non-learning target part; and a sharing control unit that controls sharing of the learning target part with an edge. The learning unit performs fine-tuning of the learning target part by using data collected in association with execution of an application that uses the large-scale language model. The sharing control unit performs control for sharing the learning target part that is fine-tuned by one of the learning unit and the edge with the other of the learning unit and the edge.

Description

Technical Field

[0001] This disclosure relates to information processing apparatus, information processing methods, and information processing systems. Background Technology

[0002] In recent years, applications utilizing large language models (LLMs) have been developed. For example, Patent Document 1 discloses a technique for generating summary sentences by utilizing large language models.

[0003] Citation List

[0004] Patent documents

[0005] Patent Document 1: Japanese Patent Application Publication No. 2023-73095 Summary of the Invention

[0006] The problem to be solved by the present invention

[0007] When optimizing large language models for specific individuals or groups, issues arise such as the need to collect and manage sufficient data for learning, as well as increased computational costs.

[0008] Solution to the problem

[0009] According to one aspect of this disclosure, an information processing apparatus is provided, comprising: a learning unit that performs learning of a large language model including a learning target portion and a non-learning target portion; and a sharing control unit that controls the sharing of the learning target portion with an edge, wherein the learning unit performs fine-tuning of the learning target portion using data collected during the execution of an application using the large language model, and the sharing control unit performs control to share the learning target portion fine-tuned by one of the learning unit or the edge with the other of the learning unit or the edge.

[0010] Furthermore, according to another aspect of this disclosure, an information processing method is provided, comprising: a processor performing the following operations: performing learning on a large language model including a learning target part and a non-learning target part; and controlling the sharing of the learning target part with the edge, wherein performing learning further comprises performing fine-tuning of the learning target part using data collected during the execution of an application using the large language model, and controlling the sharing further comprises performing control to share the learning target part fine-tuned by one of the performing learning or the edge with the other of the performing learning or the edge.

[0011] Furthermore, according to another aspect of this disclosure, an information processing system is provided, comprising an information processing terminal and a server. The information processing terminal includes: a first learning unit that performs learning of a large language model including a learning target portion and a non-learning target portion; and a first sharing control unit that controls the sharing of the learning target portion with the server. The first learning unit uses data collected during the execution of an application using the large language model to perform fine-tuning of the learning target portion, and the first sharing control unit performs control to share the fine-tuned learning target portion with the server. The server includes: a second learning unit that performs learning of the large language model; and a second sharing control unit that controls the sharing of the learning target portion with the information processing terminal. The second learning unit uses data collected during the execution of an application using the large language model to perform fine-tuning of the learning target portion, and the second sharing control unit performs control to share the fine-tuned learning target portion with the information processing terminal. Attached Figure Description

[0012] Figure 1 This is a diagram illustrating a schematic example of the configuration of an information processing system 1 according to one embodiment of the present disclosure.

[0013] Figure 2 This is a flowchart illustrating an example of the flow of an information processing method according to an embodiment.

[0014] Figure 3 This is a block diagram illustrating an example of the functional configuration of the information processing terminal 10 according to an embodiment.

[0015] Figure 4 This is a block diagram illustrating an example of the functional configuration of server 20 according to an embodiment.

[0016] Figure 5 This is a diagram used to illustrate use cases of information processing system 1 according to an implementation method.

[0017] Figure 6 This is a diagram used to illustrate use cases of information processing system 1 according to an implementation method.

[0018] Figure 7 This is a diagram illustrating an example of data stored in a user's database according to an implementation method.

[0019] Figure 8 This is a diagram illustrating an example of data stored in a proxy DB according to an implementation method.

[0020] Figure 9This is a diagram illustrating a more specific example of the configuration of the large language model 50 according to the implementation method.

[0021] Figure 10 This is a diagram used to illustrate the differences between the original large language model 510 and the adapter 520 according to the implementation method.

[0022] Figure 11 This is a diagram used to illustrate the differences in the fine-tuned execution environment between the information processing terminal 10 and the server 20 according to the implementation method.

[0023] Figure 12 This is a diagram used to illustrate the control of the dialogue simulation according to the implementation method.

[0024] Figure 13 This is a flowchart illustrating an example of a dialogue simulation and fine-tuning process using simulated data according to an implementation method.

[0025] Figure 14 This is a diagram illustrating a specific example of a dialogue simulation according to an implementation method.

[0026] Figure 15 This is a diagram illustrating an example of applying the information processing system 1 according to the embodiment to an agent application installed on a vehicle.

[0027] Figure 16 This is a diagram illustrating an example of using the information processing system 1 according to the implementation method in a fan meeting.

[0028] Figure 17 This is a block diagram illustrating an example of the hardware configuration of the information processing apparatus 90 according to an embodiment. Detailed Implementation

[0029] Preferred embodiments of this disclosure will now be described in detail with reference to the accompanying drawings. Note that in this specification and drawings, repeated descriptions of components having substantially the same functional configuration will be omitted by assigning the same reference numerals to them.

[0030] Furthermore, in this specification and accompanying drawings, when multiple identical types of components are described in a manner that distinguishes them from each other, letters or the like may be added to the end of the reference numerals. On the other hand, when it is not necessary to distinguish multiple identical types of components, the aforementioned letters or the like may be omitted, and a description of all common components of the same type may be given.

[0031] Note that the descriptions will be given in the following order.

[0032] 1. Implementation Method

[0033] 1.1. Overview

[0034] 1.2. Example of Function Configuration

[0035] 1.3. Functional Details

[0036] 1.4. Application Examples

[0037] 2. Hardware Configuration Examples

[0038] 3. Summary

[0039] <1. Implementation Method>

[0040] <<1.1. Summary>>

[0041] As mentioned above, applications utilizing large language models have been developed in recent years.

[0042] Due to the versatility of large language models, they can be applied to a wide variety of tasks.

[0043] However, there are not many large language models that have been learned to optimize for specific individuals, groups, etc.

[0044] As factors in this regard, the following issues can be listed: for example, additional learning for optimization requires significant computational costs (computing resources, time, and expenses), additional learning for optimization requires the collection of sufficient private data, and the collection of data requires consideration of privacy and security.

[0045] On the other hand, given that the aforementioned problems can be overcome and an optimized large language model can be provided, inconsistent and various types of reasoning based on individual interests, preferences, past behavior, and history can be performed. Furthermore, applications that perform various responses based on the optimized large language model can be implemented.

[0046] The technical concept of one embodiment of this disclosure is conceived by focusing on the above-mentioned points and is intended to achieve low-cost and efficient data processing for optimizing large language models.

[0047] The following will refer to Figure 1 An overview of an information processing method according to one embodiment of this disclosure is provided.

[0048] Figure 1 This is a diagram illustrating a schematic example of the configuration of the information processing system 1 according to this embodiment.

[0049] like Figure 1 As shown, the information processing system 1 according to this embodiment includes multiple information processing terminals 10 and a server 20.

[0050] Each of the multiple information processing terminals 10 is an example of an information processing device that locally executes an application using the large language model 50 and performs learning of the large language model 50.

[0051] In the following text, the information processing terminal 10 may be referred to as the edge.

[0052] One of the features of the large language model 50 according to this embodiment is that the large language model 50 includes a non-learning target part 51 and a learning target part 52.

[0053] According to this embodiment, the non-learning target portion 51 can be a common model for all large language models 50.

[0054] On the other hand, different fine-tuning is performed on the learning objective portion 52 according to this embodiment, using different data 53 for each large language model 50.

[0055] According to this embodiment, the learning objective portion 52 can be an adapter added to the non-learning objective portion of the original large language model.

[0056] exist Figure 1 In the example shown, information processing terminal 10A has a large language model 50A1, and information processing terminal 10B has a large language model 50B1.

[0057] The large language model 50A1 included in the information processing terminal 10A includes a non-learning objective portion 51 common to all large language models 50, and a learning objective portion 52A that performs fine-tuning on it using data 53A collected in the information processing terminal 10A.

[0058] On the other hand, the large language model 50B1 included in the information processing terminal 10B includes a non-learning objective portion 51 common to all large language models 50, and a learning objective portion 52B that performs fine-tuning on it using data 53B collected in the information processing terminal 10B.

[0059] Data 53A includes dialogue data, metadata, etc., collected in connection with the execution of an application using the large language model 50A1. Similarly, data 53B includes dialogue data, metadata, etc., collected in connection with the execution of an application using the large language model 50B1.

[0060] In other words, the learning objective portion 52A has weights optimized for users of the information processing terminal 10A. Similarly, the learning objective portion 52B has weights optimized for users of the information processing terminal 10B.

[0061] Each of the learning target portions 52 according to this embodiment is uploaded to server 20 and shared with server 20. Note that one of the characteristics of the information processing method according to this embodiment is that non-learning target portions 51 and data 53 are not shared with server 20.

[0062] Since the learning target portion 52 according to this embodiment is small enough in size compared to the non-learning target portion 51, communication costs (size and latency) can be reduced by sending and receiving only the learning target portion 52 between the information processing terminal 10 and the server 20.

[0063] Furthermore, the data 53 to be used for fine-tuning the learning target section 52 may be sensitive data related to users of the information processing terminal 10. Therefore, by not sharing data 53 with the server 20 (and other information processing terminals 10), data 53 can be appropriately protected.

[0064] Server 20 is an example of an information processing device that executes applications using the large language model 50 in the cloud and performs learning of the large language model 50.

[0065] The server 20 combines each of the learning target portions 52 shared by each of the multiple information processing terminals 10 with the pre-stored non-learning target portions 51, and constructs multiple large language models 50 corresponding to the large language models 50 included in each of the multiple information processing terminals 10.

[0066] exist Figure 1 In the example shown, server 20 combines the learning target portion 52A shared from information processing terminal 10A with the pre-stored non-learning target portion 51 to construct a large language model 50A2 corresponding to the large language model 50A1 included in information processing terminal 10A. The constructed large language model 50A2 can be said to be the same model as the large language model 50A1 after fine-tuning.

[0067] Similarly, server 20 combines the learning target portion 52B shared from information processing terminal 10B with the pre-stored non-learning target portion 51, and constructs a large language model 50B2 corresponding to the large language model 50B1 included in information processing terminal 10B. The constructed large language model 50B2 can be said to be the same model as the large language model 50B1 after fine-tuning.

[0068] Note that the combination of the non-learning objective part 51 and the learning objective part 52 can be easily achieved through matrix addition, which involves relatively low processing cost.

[0069] Next, server 20 controls the execution of applications using the constructed large language model 50.

[0070] The aforementioned applications include, for example, the simulation of dialogue between large language models 50 that perform the construction.

[0071] exist Figure 1 In the example shown, server 20 controls the execution of the dialogue simulation between large language model 50A2 and large language model 50B2.

[0072] In addition, server 20 stores simulation data 54 collected during the execution of the aforementioned dialogue simulation.

[0073] Since the abundant resources in the cloud can be used to perform the dialogue simulation according to this embodiment for a long time, a large amount of dialogue data (simulation data 54) can be collected.

[0074] Furthermore, dialogue data that is difficult to collect at the edge can be easily reproduced through dialogue simulation according to this embodiment.

[0075] Next, server 20 uses the simulation data 54 collected during the execution of the dialogue simulation to perform fine-tuning of each of the learning objective sections 52.

[0076] exist Figure 1 In the example shown, server 20 uses simulated data 54 to perform fine-tuning of each of learning objective parts 52A and 52B.

[0077] By fine-tuning using such simulated data 54, each of the learning objective parts 52 can be further optimized.

[0078] Next, server 20 shares each of the fine-tuned learning objective sections 52 with the corresponding information processing terminal 10.

[0079] exist Figure 1 In the example shown, server 20 sends the fine-tuned learning objective portion 52A to information processing terminal 10A and shares learning objective portion 52A. Similarly, server 20 sends the fine-tuned learning objective portion 52B to information processing terminal 10B and shares learning objective portion 52B.

[0080] Each of the information processing terminals 10 combines the non-learning target portion 51 with the learning target portion 52 shared from the server 20 and reconstructs the large language model 50.

[0081] exist Figure 1In the example shown, information processing terminal 10A combines the non-learning target portion 51 with the learning target portion 52A shared from server 20 and reconstructs the large language model 50A1. Similarly, information processing terminal 10B combines the non-learning target portion 51 with the learning target portion 52B shared from server 20 and reconstructs the large language model 50B1.

[0082] As mentioned above, the combination of the non-learning target part 51 and the learning target part 52 can be easily achieved through matrix addition, which involves relatively low processing cost.

[0083] Subsequently, each execution in the information processing terminal 10 uses the reconstructed large language model 50 for application.

[0084] The above describes an overview of the information processing method according to this embodiment. Next, refer to... Figure 2 The flow of the information processing method according to this embodiment will be described in sequence.

[0085] Figure 2 This is a flowchart illustrating an example of the flow of the information processing method according to this embodiment.

[0086] exist Figure 2 In the example shown, firstly, each of the information processing terminals 10 collects data 53 (S101).

[0087] Next, each of the information processing terminals 10 uses the data 53 collected in step S101 to perform fine-tuning of the learning target portion 52 (S102).

[0088] Each of the information processing terminals 10 sends the learning target portion 52, which has been fine-tuned in step S102, to the server and shares the learning target portion 52 (S103).

[0089] Next, the server 20 will combine each of the learning target parts 52 shared from each information processing terminal 10 in step S103 with the non-learning target parts 51 to construct a large language model 50 (S104).

[0090] Next, server 20 performs a dialogue simulation between the large language models 50 constructed in step S104 (S105).

[0091] Server 20 uses the simulation data 54 collected during the dialogue simulation in step S105 to perform fine-tuning of each of the learning target parts 52 (S106).

[0092] Next, server 20 sends each of the learning target parts 52 that have been fine-tuned in step S106 to the corresponding information processing terminal 10 and shares each of the learning target parts 52 (S107).

[0093] Each of the information processing terminals 10 combines the non-learning target part 51 with the learning target part 52 shared from the server 20 in step S107, and constructs a large language model 50 (S108).

[0094] Subsequently, each of the information processing terminals 10 executes the application using the constructed large language model 50 (S109).

[0095] The above describes an example of the flow of the information processing method according to this embodiment. Note that steps S101 to S109 described above can be repeated.

[0096] Through the series of processes described above, low-cost and efficient data processing can be achieved for optimizing large language models.

[0097] <<1.2. Example of Function Configuration>>

[0098] Next, an example of the functional configuration of the information processing terminal 10 and the server 20 according to this embodiment will be described.

[0099] First, an example of the functional configuration of the information processing terminal 10 according to this embodiment will be described.

[0100] According to this embodiment, the information processing terminal 10 is an example of an information processing device that locally executes an application using the large language model 50 and performs learning of the large language model 50.

[0101] The information processing terminal 10 according to this embodiment can be, for example, a smartphone, tablet computer, personal computer (PC), wearable device, car navigation device, etc.

[0102] Figure 3 This is a block diagram illustrating an example of the functional configuration of the information processing terminal 10 according to this embodiment.

[0103] like Figure 3 As shown, the information processing terminal 10 according to this embodiment includes a data collection unit 110, a data storage unit 120, a control unit 130, and a communication unit 170.

[0104] (Data collection unit 110)

[0105] According to this embodiment, the data collection unit 110 collects data 53 during the execution of an application using the large language model 50 under the control of the control unit 130.

[0106] For this purpose, the data collection unit 110 according to this embodiment includes various sensors. Examples of sensors include camera devices, microphones, biometric sensors, etc.

[0107] Furthermore, the data collection unit 110 according to this embodiment can collect data 53 from, for example, the Internet via a network 30.

[0108] (Data storage unit 120)

[0109] According to this embodiment, the data storage unit 120 stores data 53, etc., collected by the data collection unit 110.

[0110] (Control Unit 130)

[0111] According to this embodiment, the control unit 130 controls each of the components included in the information processing terminal 10.

[0112] The functions of the control unit 130 according to this embodiment are achieved through cooperation between various processors and memories.

[0113] like Figure 3 As shown, the control unit 130 according to this embodiment includes a learning unit 140, an application control unit 150, and a sharing control unit 160.

[0114] (Learning Unit 140)

[0115] The learning unit 140 of this embodiment performs the learning of the large language model 50.

[0116] More specifically, the learning unit 140 according to this embodiment uses data 53 to perform fine-tuning of the learning target portion 52.

[0117] Furthermore, the learning unit 140 according to this embodiment performs processes such as combining the learning target portion 52 shared from the server 20 with the non-learning target portion 51 to construct a large language model 50.

[0118] (Application Control Unit 150)

[0119] According to this embodiment, the application control unit 150 controls the execution of the application using the large language model 50.

[0120] The aforementioned applications include, for example, proxy applications.

[0121] The learning unit 140 according to this embodiment is an example of a first learning unit.

[0122] (Shared control unit 160)

[0123] According to this embodiment, the shared control unit 160 controls the sharing of the learning target portion 52 between the information processing terminal 10 and the server 20.

[0124] For example, the shared control unit 160 according to this embodiment controls the communication unit 170 so that the communication unit 170 sends the learning target portion 52, which has been fine-tuned by the learning unit 140, to the server 20.

[0125] Furthermore, for example, the shared control unit 160 according to this embodiment shares with the learning unit 140 the learning target portion 52 received from the server 20 by the communication unit 170 and fine-tuned by the server 20.

[0126] As described above, the shared control unit 160 according to this embodiment performs control to share the learning target portion 52, which has been fine-tuned by one of the learning unit 140 and the server 20, with the other of the learning unit 140 and the server 20.

[0127] The shared control unit 160 according to this embodiment is an example of a first shared control unit.

[0128] (Communication Unit 170)

[0129] According to this embodiment, the communication unit 170 performs information communication with the server 20, etc., via the network 30.

[0130] The above describes an example of the functional configuration of the information processing terminal 10 according to this embodiment.

[0131] Note, refer to Figure 3 The functional configurations described above are merely examples, and the functional configurations of the information processing terminal 10 according to this embodiment are not limited to such examples.

[0132] For example, the information processing terminal 10 according to this embodiment may also include a display unit that displays various types of information under the control of the control unit 130, an audio output unit that outputs various types of sounds under the control of the control unit 130, and an operation receiving unit that receives user operations.

[0133] The functional configuration of the information processing terminal 10 according to this embodiment can be flexibly modified according to specifications, operation, etc.

[0134] Next, an example of the functional configuration of the server 20 according to this embodiment will be described.

[0135] The server 20 according to this embodiment is an example of an information processing device that executes an application using the large language model 50 in the cloud and performs learning of the large language model 50.

[0136] Figure 4 This is a block diagram illustrating an example of the functional configuration of server 20 according to this embodiment.

[0137] like Figure 4 As shown, the server 20 according to this embodiment includes a data storage unit 210, a control unit 220, and a communication unit 270.

[0138] (Data storage unit 210)

[0139] According to this embodiment, the data storage unit 210 stores simulation data 54 and the like collected during dialogue simulation.

[0140] (Control Unit 220)

[0141] According to this embodiment, the control unit 220 controls each of the components included in the server 20.

[0142] The functions of the control unit 220 according to this embodiment are achieved through cooperation between various processors and memory.

[0143] like Figure 4 As shown, the control unit 220 according to this embodiment includes a learning unit 230, an analog control unit 240, an extended data collection unit 250, and a shared control unit 260.

[0144] (Learning Unit 230)

[0145] The learning unit 230 of this embodiment performs the learning of the large language model 50.

[0146] More specifically, the learning unit 230 according to this embodiment uses data collected during the execution of the application using the large language model 50 to perform fine-tuning of the learning objective portion 52.

[0147] Furthermore, the learning unit 230 according to this embodiment performs processes such as combining the learning target portion 52 shared from the information processing terminal 10 with the non-learning target portion 51 to construct a large language model 50.

[0148] The learning unit 230 according to this embodiment is an example of a second learning unit.

[0149] (Analog Control Unit 240)

[0150] The analog control unit 240 according to this embodiment is an example of an execution control unit that controls the execution of an application using the large language model 50.

[0151] As mentioned above, applications using the large language model 50 include applications that perform dialogue simulations between the large language models 50.

[0152] In other words, the simulation control unit 240 according to this embodiment can control the execution of dialogue simulation between large language models 50.

[0153] In this case, the learning unit 230 according to this embodiment uses the simulation data 54 collected during the execution of the dialogue simulation to perform fine-tuning of the learning target part 52.

[0154] (Extended data collection unit 250)

[0155] According to this embodiment, the extended data collection unit 250 controls the collection of extended data via the communication unit 270 based on the content of the dialogue simulation.

[0156] According to this embodiment, the learning unit 230 can also perform fine-tuning of the learning target portion 52 based on the extended data.

[0157] (Shared control unit 260)

[0158] According to this embodiment, the shared control unit 260 controls the sharing of the learning target portion 52 between each of the plurality of information processing terminals 10 and the learning unit 230.

[0159] For example, according to this embodiment, the shared control unit 260 controls the communication unit 270 so that the communication unit 270 sends each of the learning target portions 52, which have been fine-tuned by the learning unit 230, to the corresponding information processing terminal 10.

[0160] Furthermore, for example, the shared control unit 260 and the learning unit 230 according to this embodiment share the learning target portion 52, which is received by the communication unit 270 from each information processing terminal 10 and has been fine-tuned by each information processing terminal 10.

[0161] As described above, the shared control unit 260 according to this embodiment performs control to share the learning target portion 52, which has been fine-tuned by one of the learning unit 230 and the information processing terminal 10, with the other of the learning unit 230 and the information processing terminal 10.

[0162] The shared control unit 260 according to this embodiment is an example of a second shared control unit.

[0163] (Communication Unit 270)

[0164] According to this embodiment, the communication unit 270 performs information communication with the information processing terminal 10 and the like via the network 30.

[0165] An example of the functional configuration of server 20 according to this embodiment has been described above. Note that, refer to... Figure 4 The functional configurations described above are merely examples, and the functional configurations of the server 20 according to this embodiment are not limited to such examples.

[0166] For example, the server 20 according to this embodiment may also include a display unit that displays various types of information under the control of the control unit 220, an audio output unit that outputs various types of sounds under the control of the control unit 220, and an operation receiving unit that receives user operations.

[0167] The functional configuration of the server 20 according to this embodiment can be flexibly modified according to specifications, operation, etc.

[0168] <<1.3. Functional Details>>

[0169] Next, the functions of the information processing system 1 according to this embodiment will be described in detail.

[0170] First, a use case of using the information processing system 1 according to this embodiment will be described, and the details of the function will be described after the use case.

[0171] Figure 5 and Figure 6 This is a diagram used to illustrate use cases of the information processing system 1 according to this embodiment.

[0172] In this use case, it is assumed that the information processing terminal 10 is an information processing device (e.g., a car navigation system) installed in a vehicle, and the user riding in the vehicle utilizes an agent application using the large language model 50.

[0173] exist Figure 5 and Figure 6 The use case shown provides Alice and Bob as examples of the users mentioned above. Additionally, Eve in the use case is an example of a proxy in a proxy application using the Large Language Model 50.

[0174] For example, in stage 1, the information processing terminal 10 can detect the topic of the conversation based on Alice's behavior.

[0175] The information processing terminal 10 uses a camera device to acquire an image of the subject and uses image description technology to identify the subject as "Skytree".

[0176] As described above, the information processing terminal 10 according to this embodiment can supplement information not included in the speech content and can provide information based on the dialogue content.

[0177] Furthermore, throughout all stages, the information processing terminal 10 uses speech-to-text technology to acquire the dialogue content. Additionally, throughout all stages, the information processing terminal 10 can sense the user's emotions (e.g., surprise and excitement) from the tone, manner of speaking, and speed of the speech.

[0178] In addition, the information processing terminal 10 can obtain non-confidential public information from the Internet, such as road congestion in phase 2 and weather and date and time information in phase 3.

[0179] Using this information, for example, in stage 2, the information processing terminal 10 considers not only the content of the discourse and emotions, but also the content of the open information, and can propose useful means to achieve the purpose.

[0180] Furthermore, for example, in stage 3, the information processing terminal 10 can develop active dialogue by using the expressive capabilities of the large language model 50 while considering the current dialogue content.

[0181] Furthermore, for example, in stages 4 and 5, the information processing terminal 10 can spontaneously ask questions to expand and deepen the conversation and acquire new information about Alice. This new information about the user is reused as a learning topic, which is expected to further accelerate the optimization process.

[0182] The above describes a use case of using the information processing system 1 according to this embodiment.

[0183] Next, the collection and storage of data 53 according to this embodiment will be described in detail.

[0184] As described above, the data 53 according to this embodiment is collected for each information processing terminal 10 and is not shared with other information processing terminals 10 and server 20. Therefore, sensitive data such as personal information can be appropriately protected.

[0185] Each of the information processing terminals 10 may include a user DB and an agent DB in its data storage unit 120. The user DB stores data about the user in multiple data entries 53, and the agent DB stores data about the agent in multiple data entries 53.

[0186] Figure 7 This is a diagram illustrating an example of data stored in the user's database according to this embodiment. Furthermore, Figure 8 This is a diagram illustrating an example of data stored in the proxy DB according to this embodiment.

[0187] The ID, previous ID, and date and time are stored in both the user database and the agent database.

[0188] The ID is the identifier for the corresponding data group (record). The previous ID is the identifier of the record that precedes the corresponding record in chronological order. Additionally, the date and time are the date and time when the corresponding record was retrieved.

[0189] With the above information stored, the order and flow of the conversation, as well as its date and time, can be determined.

[0190] Next, we will describe the data specific to the user's database.

[0191] Spoken text, speaker, emotion, and object detection can be stored in the user's database.

[0192] Speech text is text data obtained through local speech-to-text (STT) technology.

[0193] The speaker is the speaker (user) obtained through local speaker recognition technology.

[0194] Emotion is the result of emotion analysis during speech, obtained from various types of sensor data. Emotion can be used to calculate a speaker's response score to an agent's utterances.

[0195] Furthermore, object detection is metadata about the topic of the conversation. In this example, the metadata includes information about the location where the conversation is taking place, as well as the direction and image of the Skytree visible from that location.

[0196] On the other hand, the generated text, image descriptions, and RAGs can be stored in the proxy DB.

[0197] The generated text is text data generated and spoken by an agent using local text-to-speech (TTS) technology.

[0198] Image descriptions are information related to the descriptions (textual data) of the acquired images inferred by local AI.

[0199] In addition, the RAG is reference information obtained from the Internet related to the content included in the conversation. In this example, the RAG includes general information about the Tokyo Skytree obtained from a website.

[0200] The above describes a specific example of data 53 according to this embodiment. Data 53 according to this embodiment is used as learning data, which is to be used for… Figure 2 The fine-tuning of the learning target portion 52 by the learning unit 140 in step S102 shown.

[0201] In addition, Figure 2In the execution of the application using the large language model 50 in step S109 shown, the data 53 according to this embodiment can be used as reference information when reasoning (e.g., reasoning about discourse content) is performed by the large language model 50.

[0202] Note that the data 53 stored in the data storage unit 120 can be partially or entirely deleted upon user request, ensuring the user's right to be forgotten. Deletion of data 53 can be achieved, for example, by deleting related records based on similarity search results.

[0203] Next, the fine-tuning of the large language model 50 according to this embodiment will be described in detail.

[0204] It is not uncommon for large language models to have more than 7 billion parameters, and updating all parameters through additional learning requires enormous hardware resources.

[0205] For example, performing fine-tuning on a GPT-3 model with 175 billion parameters to update all parameters requires 1.2 TB of memory, and saving the model requires 350 GB of storage per epoch.

[0206] Therefore, by using the information processing method according to this embodiment, the required hardware resources are reduced by setting the fine-tuning target only as the learning target part 52, which is part of the large language model 50.

[0207] As described above, the large language model 50 according to this embodiment includes a non-learning target portion 51 and a learning target portion 52.

[0208] Figure 9 This is a diagram illustrating a more specific example of the configuration of the large language model 50 according to this embodiment.

[0209] like Figure 9 As shown, the large language model 50 according to this embodiment may include an original large language model 510 and an adapter 520, wherein the adapter 520 is a layer added in parallel to the original large language model 510.

[0210] The original large language model 510 is an example of the non-learning objective part 51. Furthermore, the adapter 520 is an example of the learning objective part 52.

[0211] Figure 10 This is a diagram used to illustrate the differences between the original large language model 510 and the adapter 520.

[0212] like Figure 10 As shown, one of the features is that the original large language model 510 is large in size, while the adapter 520 is small in size.

[0213] Furthermore, in the original large language model 510, the initial values ​​of the weights are unique for each model, while in the adapter 520, the initial values ​​of the weights are zero or random.

[0214] Furthermore, as mentioned above, the original large language model 510 is not the target of learning (fine-tuning), while the adapter 520 is the target of learning, and its weights are updated.

[0215] Therefore, during inference, weights unique to each model are used in the original large language model 510, and weights updated through learning are used in the adapter 520.

[0216] Fine-tuning of the adapter 520 according to this embodiment can be performed using techniques such as low-rank adaptive (LoRA).

[0217] For example, when using LoRA to perform fine-tuning for updating some of the parameters in GPT-3, which has 175 billion parameters, the required memory is reduced to 350 GB, and the storage space required to save the model is reduced to 35 MB per round.

[0218] The above describes specific examples of the non-learning target portion 51 and the learning target portion 52 according to this embodiment. Next, the fine-tuning of the learning target portion 52 by the information processing terminal 10 according to this embodiment and the fine-tuning of the learning target portion 52 by the server 20 according to this embodiment will be described in detail.

[0219] Figure 11 This is a diagram used to illustrate the differences in the fine-tuning execution environment between the information processing terminal 10 and the server 20.

[0220] like Figure 11 As shown, compared with server 20, information processing terminal 10 has fewer hardware resources.

[0221] Furthermore, as mentioned above, in server 20, the simulation data collected during the execution of the dialogue simulation can be used as learning data, and therefore, it is assumed that the size of the learning data is larger than the size of the learning data in information processing terminal 10.

[0222] Furthermore, in the information processing terminal 10, the learning time is shorter than that in the server 20 in order to ensure the user's usage time, etc.

[0223] Furthermore, in the information processing terminal 10, data 53 related to actual conversations with non-simulated users can be used as learning data, and therefore, it is assumed that the quality of the learning dataset is higher than that of the learning dataset in the server 20.

[0224] Considering the above differences, the learning unit 140 of the information processing terminal 10 can perform fine-tuning by triggering the fulfillment of the conditions listed below.

[0225] - The scheduled time has passed since the last minor adjustment.

[0226] - This ensures sufficient usage time for the information processing terminal 10.

[0227] - Ensure sufficient learning data (Data 53).

[0228] - Information processing terminal 10 is charging.

[0229] - During periods when the information processing terminal 10 is used less frequently, such as late at night.

[0230] In addition, the learning unit 140 can perform small-scale fine-tuning, for example, by using only the data 53 that is not used as learning data and setting learning-related hyperparameters such as batch size to small values.

[0231] On the other hand, the learning unit 230 of the server 20 can perform fine-tuning triggered by sharing the learning target portion 52 from the information processing terminal 10.

[0232] Furthermore, provided resources permit, the learning unit 230 can use the simulation data 54 collected during the execution of the dialogue simulation to perform large-scale and long-duration fine-tuning.

[0233] As described above, the information processing method according to this embodiment can perform low-cost and efficient fine-tuning depending on the hardware characteristics of both the information processing terminal 10 and the server 20.

[0234] Next, the control of the dialogue simulation according to this embodiment will be described in detail.

[0235] As described above, the simulation control unit 240 of the server 20 according to this embodiment controls the execution of the application using the large language model 50.

[0236] Figure 12 This is a diagram used to illustrate the control of the dialogue simulation according to this embodiment.

[0237] First, the learning target portion 52 from multiple information processing terminals 10 is shared with the server 20.

[0238] Figure 12 This illustrates a scenario where multiple learning objective parts 52, including learning objective part 52A, learning objective part 52J, learning objective part 52M, and learning objective part 52X, are shared with server 20.

[0239] In the multiple learning objective sections 52, the data 53 used for fine-tuning is different for each information processing terminal 10. Therefore, there are essentially no multiple learning objective sections 52 with exactly the same weights.

[0240] Next, learning unit 230 combines each learning objective part in multiple learning objective parts 52 with non-learning objective parts 51, and constructs multiple large language models 50.

[0241] exist Figure 12 In the example shown, the learning unit 230 constructs multiple large language models 50, including large language model 50A2, large language model 50J2, large language model 50M2 and large language model 50X2.

[0242] Next, the simulation control unit 240 performs matching of the large language model 50 to perform dialogue simulation.

[0243] At this point, the analog control unit 240 can perform matching based on the similarity between the large language models 50 to further refine the corresponding configuration of the large language models 50.

[0244] The aforementioned similarity can be, for example, the similarity between the weights of the learning target part 52, or the similarity between the probability distributions of the inference results performed by the large language model 50.

[0245] exist Figure 12 In the example shown, the simulation control unit 240 matches large language model 50A2 with large language model 50M2 based on similarity, and matches large language model 50J2 with large language model 50X2.

[0246] Note that the matching score is not limited to similarity and can use a variety of scores suitable for each use case.

[0247] The simulation control unit 240 performs a dialogue simulation after matching is completed. Furthermore, simulation data 54 is collected during the execution of the dialogue simulation.

[0248] exist Figure 12 In the example shown, simulation data 54AM is collected during the execution of a dialogue simulation between large language model 50A2 and large language model 50M2. Simulation data 54AM is used for fine-tuning of learning target part 52A and learning target part 52M.

[0249] In addition, Figure 12 In the example shown, simulation data 54JX is collected during the execution of a dialogue simulation between large language model 50J2 and large language model 50X2. The simulation data 54JX is used for fine-tuning of the learning target part 52J and the learning target part 52X.

[0250] Since the dialogue simulation according to this embodiment can be performed for a long time and at high speed using the abundant resources in the cloud, a large amount of dialogue data (simulation data 54) can be collected.

[0251] Furthermore, using the dialogue simulation according to this embodiment, even dialogue data that is difficult to collect at the edge can be easily reproduced.

[0252] Furthermore, the extended data collection unit 250 of this embodiment collects extended data based on dialogue simulation content.

[0253] For example, the extended data collection unit 250 can search for and collect words that appear in the dialogue simulation on the Internet.

[0254] Furthermore, the extended data collection unit 250 can, for example, search and collect other words on the Internet that have a high degree of similarity to the words that appear in the dialogue simulation.

[0255] According to this embodiment, the learning unit 230 uses the simulation data 54 and extended data collected as described above to perform fine-tuning of the learning target portion 52.

[0256] The control of the dialogue simulation according to this embodiment has been described above. Next, refer to... Figure 13 The process of dialogue simulation and fine-tuning using simulation data according to this embodiment will be described in turn.

[0257] Figure 13 This is a flowchart illustrating an example of the dialogue simulation and fine-tuning process using simulation data according to this embodiment.

[0258] exist Figure 13 In the example shown, firstly, each of the information processing terminals 10 shares the learning target portion 52 with the server 20 (S201).

[0259] The learning unit 230 of server 20 combines each of the learning target parts 52 shared in step S201 with the non-learning target parts 51 to construct multiple large language models 50 (S202).

[0260] Next, the simulation control unit 240 performs matching of the large language models 50 constructed in S202 to perform dialogue simulation (S203).

[0261] The simulation control unit 240 controls the execution of the dialogue simulation between the matched large language models 50 in step S203 and collects simulation data 54 (S204).

[0262] Next, the extended data collection unit 250 collects extended data based on the simulation data 54 collected in step S204 (S205).

[0263] Learning unit 230 uses the simulation data 54 collected in step S204 and the extended data collected in step S205 to fine-tune the learning target part 52 of the large language model 50, which performed dialogue simulation in step S204 (S206).

[0264] Next, a specific example of dialogue simulation according to this embodiment will be described. Figure 14 This is a diagram illustrating a specific example of a dialogue simulation according to this embodiment.

[0265] Figure 14 LLM J in Figure 12 The large language model 50J2 in the middle, and Figure 14 LLM X in Figure 12 The large language model in the middle is 50X2.

[0266] The simulation control unit 240 according to this embodiment can control the large language model 50, enabling questions that associate events and emotions to be asked in dialogue simulation.

[0267] exist Figure 14 In the example shown, the analog control unit 240 controls the large language model 50J2 such that in phase 1, the large language model 50J2 asks a question about whether there are things that can evoke emotional responses.

[0268] Furthermore, when responding further to the answer to a question in a dialogue simulation, the simulation control unit 240 according to this embodiment can control the large language model 50 to make a response that includes emotion and additional information.

[0269] exist Figure 14 In the example shown, the analog control unit 240 controls the large language model 50J2 such that in phase 2, the large language model 50J2 makes an instruction to empathize and includes additional information related to the view of the Skytree at dusk and during the day.

[0270] Furthermore, when additional information is obtained during dialogue simulation, the simulation control unit 240 according to this embodiment can control the large language model to make an instruction to the large language model whether the response to the additional information is affirmative or negative.

[0271] exist Figure 14 In the example shown, the analog control unit 240 controls the large language model 50X2 so that in phase 2, the large language model 50X2 indicates that the large language model 50X2 responds affirmatively to the additional information related to Tokyo Skytree.

[0272] Furthermore, in the dialogue simulation, the simulation control unit 240 according to this embodiment can control the large language model 50, so that the large language model 50 also asks the other party for the received question.

[0273] exist Figure 14 In the example shown, the analog control unit 240 controls the large language model 50X2 so that in phase 2, the large language model 50X2 also queries the large language model 50J2 for the question received in phase 1.

[0274] With the controls described above, large amounts of data can be collected more efficiently compared to real-world conversations.

[0275] Furthermore, as described above, the extended data collection unit 250 according to this embodiment can collect extended data based on the content of the dialogue simulation.

[0276] For example, the extended data collection unit 250 can search for and collect words such as "Skytree" and "Atami" that appear in the dialogue simulation on the Internet.

[0277] Furthermore, for example, the extended data collection unit 250 can search and collect words such as "Tokyo Tower" which has a high similarity to the word "Skytree" that appears in the dialogue simulation, and "Shizuoka" which has a high similarity to the word "Atami" that appears in the dialogue simulation.

[0278] By collecting the simulated data 54 as described above and by collecting extended data, a significant expansion of the learning data can be achieved, and large-scale learning optimized for individuals can be performed.

[0279] <<1.4. Examples of Application>>

[0280] Next, an example of the application of the information processing system 1 according to this embodiment will be described.

[0281] As described in the above use case, the information processing system 1 according to this embodiment can be applied to an agent application installed in a vehicle.

[0282] Figure 15 This is a diagram illustrating an example of applying the information processing system 1 according to this embodiment to an agent application installed on a vehicle.

[0283] In this application example, the information processing terminal 10A may be, for example, a car navigation device installed in a vehicle and having agent functions.

[0284] When the vehicle departs, the information processing terminal 10A downloads the latest learning objective section 52A from the server 20.

[0285] Furthermore, during driving, the information processing terminal 10 uses the large language model 50A1 to respond to prompts from the user in the vehicle and perform reasoning.

[0286] The information processing terminal 10 can use a separately optimized large language model 50A1 to make responses that reflect personal tastes and preferences.

[0287] Furthermore, using the accumulated data 53A as context for the prompts improves the accuracy of the expected response.

[0288] In addition, the information processing terminal 10 accumulates new data 53A during driving. A specific example of data 53 is shown above. Figure 7 and Figure 8 As mentioned above.

[0289] After returning home, the information processing terminal 10 uses the accumulated data 53A to perform fine-tuning on the learning objective portion 52A, and uploads the fine-tuned learning objective portion 52A to the server 20.

[0290] After this, the information processing terminal 10 can be powered off until the next time it is used.

[0291] On the other hand, server 20 executes the above-mentioned references. Figure 12 and Figure 13 The description includes a series of processes for dialogue simulation and fine-tuning.

[0292] The fine-tuned learning objective portion 52A, executed by server 20, is downloaded to information processing terminal 10 the next time the vehicle is used.

[0293] The above describes an example of applying the information processing system 1 according to this embodiment to an agent application installed in a vehicle.

[0294] Next, refer to Figure 16 The following describes an example of using the information processing system 1 according to this embodiment in a fan meeting.

[0295] In this application example, posts made by artist A1's fan F1 on applications such as SNS, and non-fungible tokens (NFTs) held in the application can be collected as data 53.

[0296] In addition, images captured by camera devices 112 installed at the event venue and sounds recorded by microphones 114 installed at the event venue can also be collected as data 53.

[0297] The data 53 collected as described above can be provided to artist A1 as a profile of fan F1.

[0298] In addition, the information processing terminal 10 can use data 53 to perform reasoning on the finely tuned large language model 50, and suggest things that fan F1 likes, words that fan F1 wants to hear, and actions that fan F1 wants artist A1 to take.

[0299] In addition, information related to the interaction between artist A1 and fan F1 at event venues and other locations, which has accepted the above recommendations, can be collected by camera device 112 and microphone 114 and accumulated as data 53.

[0300] By collecting data 53 as described above, inference optimized for the target audience can be performed.

[0301] <2. Hardware Configuration Example>

[0302] Next, an example of a common hardware configuration of the information processing terminal 10 and the server 20 according to one embodiment of this disclosure will be described. Figure 17 This is a block diagram illustrating an example of the hardware configuration of an information processing apparatus 90 according to one embodiment of the present disclosure. The information processing apparatus 90 may be an apparatus having a hardware configuration equivalent to the hardware configurations of the various apparatuses described above.

[0303] like Figure 17 As shown, the information processing apparatus 90 includes, for example, a processor 871, a ROM 872, a RAM 873, a host bus 874, a bridge 875, an external bus 876, an interface 877, an input device 878, an output device 879, a storage device 880, a driver 881, a connection port 882, and a communication device 883. Note that the hardware configuration described herein is exemplary, and some components may be omitted. Furthermore, the information processing apparatus 90 may also include components other than those described herein.

[0304] (Processor 871)

[0305] The processor 871 is used as, for example, an arithmetic processing device or a control device, and controls all or some of the operations of the various components based on various programs recorded in ROM 872, RAM 873, storage device 880 or removable storage medium 901.

[0306] (ROM 872 and RAM 873)

[0307] ROM 872 is a device for storing programs read by processor 871, data used for calculation, etc. RAM 873 stores, for example, programs read by processor 871, various parameters that change appropriately when the program is executed, etc., temporarily or permanently.

[0308] (Host bus 874, bridge 875, external bus 876 and interface 877)

[0309] The processor 871, ROM 872, and RAM 873 are interconnected via, for example, a host bus 874, through which high-speed data transfer can be performed. On the other hand, the host bus 874 is connected, for example, via a bridge 875 to an external bus 876, which has a relatively low data transfer speed. Furthermore, the external bus 876 is connected to various components via an interface 877.

[0310] (Input device 878)

[0311] For example, a mouse, keyboard, touchpad, button, switch, joystick, etc., can be used as input device 878. Alternatively, a remote control (hereinafter referred to as a remote control) capable of transmitting control signals using infrared or other radio waves can also be used as input device 878. Furthermore, input device 878 includes audio input devices such as microphones.

[0312] (Output device 879)

[0313] Output device 879 is, for example, a device capable of visually or audibly notifying a user of the information acquired, such as a cathode ray tube (CRT), LCD, or OLED display device, an audio output device such as a speaker or headphones, a printer, a mobile phone, or a fax machine. Furthermore, output device 879 according to this disclosure includes various vibrating devices capable of outputting tactile stimuli.

[0314] (Storage device 880)

[0315] Storage device 880 is a device used to store various types of data. For example, magnetic storage devices such as hard disk drives (HDDs), semiconductor storage devices, optical storage devices, magneto-optical storage devices, etc., are used as storage device 880.

[0316] (Driver 881)

[0317] The drive 881 is, for example, a device for reading information recorded in or writing information to a removable storage medium 901 such as a disk, optical disk, magneto-optical disk, or semiconductor memory.

[0318] (Removable storage medium 901)

[0319] The removable storage medium 901 can be, for example, DVD media, Blu-ray (registered trademark) media, HD DVD media, or any of various semiconductor storage media. Of course, the removable storage medium 901 can also be, for example, an IC card including a contactless IC chip, an electronic device, etc.

[0320] (Connect to port 882)

[0321] Connection port 882 is, for example, a port for connecting external connection devices 902, such as a Universal Serial Bus (USB) port, an IEEE 1394 port, a Small Computer System Interface (SCSI) port, an RS-232C port, or an optical audio terminal.

[0322] (External connection device 902)

[0323] External connection device 902 is, for example, a printer, portable music player, digital camera, digital video camera, IC recorder, etc.

[0324] (Communication Equipment 883)

[0325] Communication device 883 is a communication device for connecting to a network, and is, for example, a communication card for wired or wireless LAN, Bluetooth (registered trademark) or WUSB (wireless USB), a router for optical communication, a router for asymmetric digital subscriber line (ADSL), a modem for various types of communication, etc.

[0326] <3. Summary>

[0327] As described above, according to one embodiment of this disclosure, server 20 includes a learning unit 230 and a sharing control unit 260. Learning unit 230 performs learning of a large language model 50, including a learning target portion 52 and a non-learning target portion 51. Sharing control unit 260 controls the sharing of the learning target portion 52 with information processing terminal 10. Learning unit 230 uses data 53 collected during the execution of an application using the large language model 50 to fine-tune the learning target portion 52. Sharing control unit 260 performs control to share the learning target portion 52, which has been fine-tuned by one of learning unit 230 and information processing terminal 10, with the other.

[0328] Using the above configuration, low-cost and efficient data processing can be achieved for optimizing large language models.

[0329] Although preferred embodiments of the present disclosure have been described in detail with reference to the accompanying drawings, the technical scope of the present disclosure is not limited to such examples. It will be apparent to those skilled in the art that various changes or modifications can be conceived within the scope of the technical concept disclosed in the claims, and such changes or modifications are naturally understood to also fall within the technical scope of the present disclosure.

[0330] Furthermore, the individual steps related to the processing described in this disclosure do not necessarily need to be performed in chronological order as shown in the flowchart or sequence diagram. For example, the individual steps related to the processing to be performed by various devices may be performed in a different order than that described, or they may be processed in parallel.

[0331] Furthermore, the series of processes to be performed by the various devices described in this disclosure can be executed using software, hardware, or any combination of software and hardware. Programs constituting the software are, for example, located internally or externally to each device and pre-stored in a non-transitory computer-readable storage medium. Additionally, the various programs are loaded into RAM, for example, when executed by a computer, and executed by various processors. The aforementioned storage media are, for example, magnetic disks, optical disks, magneto-optical disks, flash memory, etc. Furthermore, the aforementioned computer programs can be delivered, for example, via a network without using a storage medium.

[0332] Furthermore, the effects described in this specification are illustrative or exemplary only, and not restrictive. In other words, the technology according to this disclosure can achieve, or replace, other effects that will be apparent to those skilled in the art based on the description in this specification.

[0333] Note that configurations such as the following also fall within the technical scope of this disclosure. (1)

[0335] An information processing apparatus, comprising:

[0336] Learning units, said learning units perform learning of a large language model including learning objective parts and non-learning objective parts; and

[0337] A shared control unit controls the sharing of the learning target portion with the edge.

[0338] in,

[0339] The learning unit uses data collected during the execution of the application using the large language model to perform fine-tuning of the learning objective portion, and

[0340] The shared control unit performs control to share the learning target portion, which has been fine-tuned by one of the learning units or the edge, with the other of the learning unit or the edge. (2)

[0342] According to the information processing apparatus described in (1), wherein,

[0343] The shared control unit shares the learning target portion with the learning unit. This learning target portion is received from the edge and fine-tuned by the edge using data collected during the execution of the application using the large language model by the edge.

[0344] The learning unit constructs the large language model by combining the shared learning objective portion with the non-learning objective portion. (3)

[0346] According to the information processing device described in (1) or (2),

[0347] Also includes:

[0348] An execution control unit controls the execution of applications using the large language model. (4)

[0350] According to the information processing device described in (3), wherein,

[0351] The execution control unit controls the execution of the dialogue simulation between the large language models, and

[0352] The learning unit uses simulation data collected during the execution of the dialogue simulation to fine-tune the learning objective portion. (5)

[0354] According to the information processing device described in (4), wherein,

[0355] The shared control unit shares the learning target portion received from each of the plurality of edges with the learning unit, and

[0356] The learning unit constructs multiple large language models corresponding to each of the multiple edges by combining each of the multiple shared learning target parts with the non-learning target parts. (6)

[0358] According to the information processing apparatus described in (5), wherein,

[0359] The execution control unit performs matching of the large language models used to perform the dialogue simulation among multiple large language models corresponding to the respective edges. (7)

[0361] According to the information processing apparatus described in (6), wherein,

[0362] The execution control unit performs the matching based on the similarity between the large language models. (8)

[0364] According to the information processing apparatus described in (7), wherein,

[0365] The similarity includes at least one of the similarity between the weights of the learning target components or the similarity between the results of inference performed by the large language model. (9)

[0367] The information processing apparatus according to any one of (4) to (8), wherein,

[0368] The execution control unit controls the large language model to ask questions about the association between events and emotions in the dialogue simulation. (10)

[0370] The information processing apparatus according to any one of (4) to (9), wherein,

[0371] When the dialogue simulation provides further responses to the answers to questions, the execution control unit controls the large language model to make responses that include emotion and additional information. (11)

[0373] The information processing apparatus according to any one of (4) to (10), wherein,

[0374] When additional information is obtained in the dialogue simulation, the execution control unit controls the large language model to make an indication of whether the large language model responds to the additional information affirmatively or negatively. (12)

[0376] The information processing apparatus according to any one of (4) to (11), wherein,

[0377] The execution control unit controls the large language model to ask the other party a received question in the dialogue simulation. (13)

[0379] The information processing apparatus according to any one of (4) to (12),

[0380] Also includes:

[0381] An extended data collection unit collects extended data based on the content of the dialogue simulation, wherein...

[0382] The learning unit further uses the extended data to perform fine-tuning of the learning objective portion. (14)

[0384] The information processing apparatus according to any one of (1) to (13), wherein,

[0385] The learning objective part is an adapter added to the non-learning objective part, which is the original large language model. (15)

[0387] An information processing method, comprising:

[0388] The processor performs the following operations:

[0389] Perform learning on a large language model that includes both learning objective and non-learning objective parts; and

[0390] The sharing of the learning objective portion of control and edge.

[0391] in,

[0392] The execution learning also includes using data collected during the execution of the application using the large language model to fine-tune the learning objective portion, and

[0393] The control sharing also includes execution control to share the learning target portion, which has been fine-tuned by one of the execution learning or the edge, with the other of the execution learning or the edge. (16)

[0395] An information processing system, comprising:

[0396] Information processing terminals and servers

[0397] in,

[0398] The information processing terminal includes:

[0399] The first learning unit performs learning of a large language model, including both learning objective and non-learning objective parts; and

[0400] A first shared control unit controls the sharing of the learning target portion with the server.

[0401] The first learning unit uses data collected during the execution of the application using the large language model to perform fine-tuning of the learning objective portion, and

[0402] The first shared control unit performs control to share the learning target portion, which has been fine-tuned by the first learning unit, with the server.

[0403] The server includes:

[0404] The second learning unit performs the learning of the large language model; and

[0405] The second shared control unit controls the sharing of the learning target portion with the information processing terminal.

[0406] The second learning unit uses data collected during the execution of the application using the large language model to perform fine-tuning of the learning objective portion, and

[0407] The second shared control unit performs control to share the learning target portion, which has been fine-tuned by the second learning unit, with the information processing terminal.

[0408] List of reference numerals

[0409] 1. Information Processing System

[0410] 10 Information Processing Terminals

[0411] 130 Control Unit

[0412] 140 Learning Units

[0413] 150 Application Control Unit

[0414] 160 Shared Control Unit

[0415] 20 servers

[0416] 220 Control Unit

[0417] 230 Learning Units

[0418] 240 Analog Control Unit

[0419] 250 Extended Data Collection Units

[0420] 260 Shared Control Unit

[0421] 50 major language models

[0422] 51. Non-learning objective section

[0423] 52 Learning Objectives Section

[0424] 53 Data

[0425] 54 Simulation Data

Claims

1. An information processing apparatus, comprising: A learning unit that performs learning of a large language model, including both learning objective parts and non-learning objective parts; as well as A shared control unit controls the sharing of the learning target portion with the edge. in, The learning unit uses data collected during the execution of the application using the large language model to perform fine-tuning of the learning objective portion, and The shared control unit performs control to share the learning target portion, which has been fine-tuned by one of the learning units or the edge, with the other of the learning unit or the edge.

2. The information processing apparatus according to claim 1, wherein, The shared control unit shares the learning target portion with the learning unit. This learning target portion is received from the edge and fine-tuned by the edge using data collected during the execution of the application using the large language model by the edge. The learning unit constructs the large language model by combining the shared learning objective portion with the non-learning objective portion.

3. The information processing apparatus according to claim 1, further comprising: An execution control unit controls the execution of applications using the large language model.

4. The information processing apparatus according to claim 3, wherein, The execution control unit controls the execution of the dialogue simulation between the large language models, and The learning unit uses simulation data collected during the execution of the dialogue simulation to fine-tune the learning objective portion.

5. The information processing apparatus according to claim 4, wherein, The shared control unit shares the learning target portion received from each of the plurality of edges with the learning unit, and The learning unit constructs multiple large language models corresponding to each of the multiple edges by combining each of the multiple shared learning target parts with the non-learning target parts.

6. The information processing apparatus according to claim 5, wherein, The execution control unit performs matching of the large language models used to perform the dialogue simulation among multiple large language models corresponding to the respective edges.

7. The information processing apparatus according to claim 6, wherein, The execution control unit performs the matching based on the similarity between the large language models.

8. The information processing apparatus according to claim 7, wherein, The similarity includes at least one of the similarity between the weights of the learning target components or the similarity between the results of inference performed by the large language model.

9. The information processing apparatus according to claim 4, wherein, The execution control unit controls the large language model to ask questions about the association between events and emotions in the dialogue simulation.

10. The information processing apparatus according to claim 4, wherein, When the dialogue simulation provides further responses to the answers to questions, the execution control unit controls the large language model to make responses that include emotion and additional information.

11. The information processing apparatus according to claim 4, wherein, When additional information is obtained in the dialogue simulation, the execution control unit controls the large language model to make an indication of whether the large language model responds to the additional information affirmatively or negatively.

12. The information processing apparatus according to claim 4, wherein, The execution control unit controls the large language model to ask the other party a received question in the dialogue simulation.

13. The information processing apparatus according to claim 4, further comprising: An extended data collection unit collects extended data based on the content of the dialogue simulation, wherein... The learning unit further uses the extended data to perform fine-tuning of the learning objective portion.

14. The information processing apparatus according to claim 1, wherein, The learning objective part is an adapter added to the non-learning objective part, which is the original large language model.

15. An information processing method, comprising: The processor performs the following operations: Perform learning on a large language model that includes both learning objective parts and non-learning objective parts; as well as The sharing of the learning objective portion of control and edge. in, The execution learning also includes using data collected during the execution of the application using the large language model to fine-tune the learning objective portion, and The control sharing also includes execution control to share the learning target portion, which has been fine-tuned by one of the execution learning or the edge, with the other of the execution learning or the edge.

16. An information processing system, comprising: Information processing terminals and servers in, The information processing terminal includes: The first learning unit performs learning of a large language model, including both learning objective and non-learning objective parts; and A first shared control unit controls the sharing of the learning target portion with the server. The first learning unit uses data collected during the execution of the application using the large language model to perform fine-tuning of the learning objective portion. The first shared control unit performs control to share the learning target portion, which has been fine-tuned by the first learning unit, with the server. The server includes: The second learning unit performs the learning of the large language model; and The second shared control unit controls the sharing of the learning target portion with the information processing terminal. The second learning unit uses data collected during the execution of the application using the large language model to perform fine-tuning of the learning objective portion, and The second shared control unit performs control to share the learning target portion, which has been fine-tuned by the second learning unit, with the information processing terminal.

Images