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

Abstract

[PROBLEM] To realize low-cost and efficient data operations relating to the optimization of a large-scale language model. [SOLUTION] Provided is an information processing device comprising: a learning unit that learns a large-scale language model composed of 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 conjunction with the execution of an application using the large-scale language model. The sharing control unit performs control for sharing the learning target part subjected to fine tuning by one of the learning unit and the edge with the other of the learning unit and the edge.

Description

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

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

[0002] In recent years, applications utilizing large language models (LLMs) have been developed. For example, Patent Literature 1 discloses a technology for generating summaries using large language models.

[0003] JP 2023-73095 A

[0004] When trying to optimize a large-scale language model for a specific individual or group, issues arise, such as collecting and managing sufficient data for learning and increasing computational costs.

[0005] According to one aspect of the present disclosure, there is provided an information processing device comprising: a learning unit that learns a large-scale language model consisting of a learning target portion and a non-learning target portion; and a sharing control unit that controls the sharing of the learning target portion between an edge, wherein the learning unit fine-tunes the learning target portion using data collected in conjunction with the execution of an application using the large-scale language model; and the sharing control unit controls the sharing of the learning target portion that has been fine-tuned by either the learning unit or the edge with the other.

[0006] According to another aspect of the present disclosure, there is provided an information processing method, including: a processor training a large-scale language model consisting of a training target portion and a non-training target portion; and controlling the sharing of the training target portion between edges; the training further includes fine-tuning the training target portion using data collected in conjunction with the execution of an application using the large-scale language model; and the controlling the sharing further includes controlling the sharing of the training target portion that has been fine-tuned by either one of the edges with the other.

[0007] According to another aspect of the present disclosure, there is provided an information processing system comprising: an information processing terminal; and a server, wherein the information processing terminal comprises: a first learning unit that learns a large-scale language model consisting of a learning object portion and a non-learning object portion; and a first sharing control unit that controls sharing of the learning object portion with the server, wherein the first learning unit fine-tunes the learning object portion using data collected in conjunction with execution of an application using the large-scale language model, and the first sharing control unit controls sharing of the learning object portion fine-tuned by the first learning unit with the server, wherein the server comprises: a second learning unit that learns the large-scale language model; and a second sharing control unit that controls sharing of the learning object portion with the information processing terminal, wherein the second learning unit fine-tunes the learning object portion using data collected in conjunction with execution of the application using the large-scale language model, and the second sharing control unit controls sharing of the learning object portion fine-tuned by the second learning unit with the information processing terminal.

[0008] FIG. 1 is a diagram for describing a schematic configuration example of an information processing system 1 according to an embodiment of the present disclosure. FIG. 2 is a flowchart showing an example of the flow of an information processing method according to the embodiment. FIG. 3 is a block diagram showing an example of the functional configuration of an information processing terminal 10 according to the embodiment. FIG. 4 is a block diagram showing an example of the functional configuration of a server 20 according to the embodiment. FIG. 5 is a diagram for describing a use case to which the information processing system 1 according to the embodiment is applied. FIG. 6 is a diagram for describing a use case to which the information processing system 1 according to the embodiment is applied. FIG. 7 is a diagram showing an example of data stored in a user DB according to the embodiment. FIG. 8 is a diagram showing an example of data stored in an agent DB according to the embodiment. FIG. 9 is a diagram showing a more specific configuration example of a large-scale language model 50 according to the embodiment. FIG. 10 is a diagram for describing a difference between an original large-scale language model 510 and an adapter 520 according to the embodiment. FIG. 11 is a diagram for describing a difference in the execution environment of fine tuning in the information processing terminal 10 and the server 20 according to the embodiment. FIG. 12 is a diagram for describing control of a conversation simulation according to the embodiment. FIG. 13 is a flowchart showing an example of the flow of fine tuning using a conversation simulation and simulation data according to the embodiment. FIG. 14 is a diagram showing a specific example of a conversation simulation according to the embodiment. FIG. 15 is a diagram for describing an example of application of the information processing system 1 according to the embodiment to an agent application mounted in a vehicle. FIG. 16 is a block diagram showing an example of the hardware configuration of an information processing device 90 according to the embodiment.

[0009] Preferred embodiments of the present disclosure will be described in detail below with reference to the accompanying drawings. In this specification and drawings, components having substantially the same functional configurations are designated by the same reference numerals, and redundant description will be omitted.

[0010] In addition, in this specification and drawings, when multiple identical components are to be described separately, 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 between multiple identical components, the letters or the like may be omitted and a description common to all of the multiple identical components may be given.

[0011] The description will be given in the following order: 1. Embodiment 1.1. Overview 1.2. Functional configuration example 1.3. Functional details 1.4. Application example 2. Hardware configuration example 3. Summary

[0012] 1. Embodiments 1.1. Overview As described above, applications utilizing large-scale language models have been developed in recent years. Large-scale language models are versatile and can be applied to a wide variety of tasks.

[0013] However, there are not many large-scale language models that have been trained to be optimized for a given individual, group, etc.

[0014] The reasons for this include the fact that additional learning for optimization requires a large computational cost (computing resources, time, and money), the need to collect sufficient private data for additional learning for optimization, and the need to handle the collected data securely and with consideration for privacy.

[0015] On the other hand, if the above-mentioned challenges can be overcome and an optimized large-scale language model can be provided, it will be possible to realize diverse inferences that are not uniform but are based on individual interests, preferences, past actions, and their history, etc. Furthermore, it will be possible to implement applications that provide diverse responses based on optimized large-scale language models.

[0016] The technical concept of one embodiment of the present disclosure was conceived with the above points in mind, and realizes low-cost and efficient data management for optimizing large-scale language models.

[0017] An overview of an information processing method according to an embodiment of the present disclosure will be described below with reference to FIG.

[0018] FIG. 1 is a diagram illustrating a schematic configuration example of an information processing system 1 according to this embodiment.

[0019] As shown in FIG. 1 , an information processing system 1 according to this embodiment includes a plurality of information processing terminals 10 and a server 20 .

[0020] Each of the multiple information processing terminals 10 is an example of an information processing device that locally executes an application using the large-scale language model 50 and learns the large-scale language model 50. Hereinafter, the information processing terminal 10 may be referred to as an edge.

[0021] One of the features of the large-scale language model 50 according to this embodiment is that it is made up of a non-learning part 51 and a learning part 52 .

[0022] The non-learning target portion 51 according to this embodiment may be a model common to all large-scale language models 50 .

[0023] On the other hand, the training object unit 52 according to this embodiment is subjected to different fine tuning using different data 53 for each large-scale language model 50 .

[0024] The training target portion 52 according to this embodiment may be an adapter that is added to the non-training target portion, which is the original large-scale language model.

[0025] In the example shown in FIG. 1, the information processing terminal 10A has a large-scale language model 50A1, and the information processing terminal 10B has a large-scale language model 50B1.

[0026] The large-scale language model 50A1 possessed by the information processing terminal 10A consists of a non-learning target portion 51 that is common to all large-scale language models 50, and a learning target portion 52A on which fine tuning is performed using data 53A collected in the information processing terminal 10A.

[0027] On the other hand, the large-scale language model 50B1 possessed by the information processing terminal 10B consists of a non-learning target portion 51 that is common to all large-scale language models 50, and a learning target portion 52B on which fine tuning is performed using data 53B collected in the information processing terminal 10B.

[0028] Data 53A includes conversation data, metadata, etc. collected in conjunction with the execution of an application using large-scale language model 50A1. Similarly, data 53B includes conversation data, metadata, etc. collected in conjunction with the execution of an application using large-scale language model 50B1.

[0029] That is, the learning object section 52A has a weight optimized for the user who uses the information processing terminal 10A. Similarly, the learning object section 52B has a weight optimized for the user who uses the information processing terminal 10B.

[0030] Each of the learning target parts 52 according to this embodiment is uploaded to and shared with the server 20. Note that one of the features of the information processing method according to this embodiment is that the non-learning target parts 51 and data 53 are not shared with the server 20.

[0031] Since the learning target portion 52 in this embodiment is sufficiently small in size compared to the non-learning target portion 51, communication costs (size, latency) can be reduced by transmitting and receiving only the learning target portion 52 between the information processing terminal 10 and the server 20.

[0032] Furthermore, the data 53 used for fine-tuning the learning target unit 52 may be sensitive data related to the user of the information processing terminal 10. Therefore, the data 53 can be appropriately protected by not sharing the data 53 with the server 20 (and other information processing terminals 10).

[0033] The server 20 is an example of an information processing device that executes an application using the large-scale language model 50 and performs training of the large-scale language model 50 in the cloud.

[0034] The server 20 combines each of the learning target portions 52 shared from each of the multiple information processing terminals 10 with the non-learning target portions 51 that it stores in advance, and constructs multiple large-scale language models 50 corresponding to the large-scale language models 50 that each of the multiple information processing terminals 10 has.

[0035] 1 , the server 20 combines the learning target portion 52A shared from the information processing terminal 10A with the non-learning target portion 51 stored in advance to construct a large-scale language model 50A2 corresponding to the large-scale language model 50A1 held by the information processing terminal 10A. The large-scale language model 50A2 at the time of construction can be said to be the same model as the large-scale language model 50A1 after fine-tuning has been performed.

[0036] Similarly, the server 20 combines the training target portion 52B shared from the information processing terminal 10B with the non-training target portion 51 stored in advance to construct a large-scale language model 50B2 corresponding to the large-scale language model 50B1 held by the information processing terminal 10B. The large-scale language model 50B2 at the time of construction can be said to be the same model as the large-scale language model 50B1 after fine-tuning has been performed.

[0037] The non-learning target portion 51 and the learning target portion 52 can be combined by simply adding matrices together, which requires relatively low processing costs.

[0038] Next, the server 20 controls the execution of an application that uses the constructed large-scale language model 50 .

[0039] The above application includes, for example, an application that performs a conversation simulation between the constructed large-scale language models 50 .

[0040] In the example shown in FIG. 1, the server 20 controls the execution of a conversation simulation between the large-scale language model 50A2 and the large-scale language model 50B2.

[0041] The server 20 also stores simulation data 54 collected during the execution of the conversation simulation.

[0042] The conversation simulation according to this embodiment can be performed for a long period of time using abundant resources on the cloud, making it possible to collect a large amount of conversation data (simulation data 54).

[0043] Furthermore, according to the conversation simulation of this embodiment, even conversation data that is difficult to collect on the edge side can be easily reproduced.

[0044] Next, the server 20 performs fine tuning of each of the learning target parts 52 using the simulation data 54 collected during the execution of the conversation simulation.

[0045] In the example shown in FIG. 1, the server 20 uses the simulation data 54 to perform fine tuning of each of the learning object portion 52A and the learning object portion 52B.

[0046] Fine tuning using such simulation data 54 makes it possible to further optimize each of the learning target parts 52 .

[0047] Next, the server 20 shares each of the fine-tuned learning target units 52 with the corresponding information processing terminal 10 .

[0048] 1 , the server 20 transmits the fine-tuned learning object portion 52A to the information processing terminal 10A and shares it with the information processing terminal 10A. Similarly, the server 20 transmits the fine-tuned learning object portion 52B to the information processing terminal 10B and shares it with the information processing terminal 10B.

[0049] Each of the information processing terminals 10 combines the non-learning target portion 51 with the learning target portion 52 shared by the server 20 to reconstruct a large-scale language model 50 .

[0050] 1 , information processing terminal 10A combines non-learning target portion 51 with learning target portion 52A shared by server 20 to reconstruct large-scale language model 50A1. Similarly, information processing terminal 10B combines non-learning target portion 51 with learning target portion 52B shared by server 20 to reconstruct large-scale language model 50B1.

[0051] As described above, the non-learning target portion 51 and the learning target portion 52 can be combined by simply adding matrices together, which requires relatively low processing costs.

[0052] Thereafter, each of the information processing terminals 10 executes an application that uses the reconstructed large-scale language model 50 .

[0053] The information processing method according to this embodiment has been outlined above. Next, the flow of the information processing method according to this embodiment will be summarized with reference to FIG.

[0054] FIG. 2 is a flowchart showing an example of the flow of an information processing method according to this embodiment.

[0055] In the example shown in FIG. 2, first, each of the information processing terminals 10 collects data 53 (S101).

[0056] Next, each of the information processing terminals 10 performs fine tuning of the learning target unit 52 using the data 53 collected in step S101 (S102).

[0057] Each information processing terminal 10 transmits the learning object unit 52 that has been fine-tuned in step S102 to the server and shares it (S103).

[0058] Next, the server 20 combines each of the learning target portions 52 shared from each of the information processing terminals 10 in step S103 with the non-learning target portions 51 to construct a large-scale language model 50 (S104).

[0059] Next, the server 20 executes a conversation simulation between the large-scale language models 50 constructed in step S104 (S105).

[0060] The server 20 performs fine tuning of each of the learning target parts 52 using the simulation data 54 collected in conjunction with the conversation simulation in step S105 (S106).

[0061] Next, the server 20 transmits each of the learning target units 52 that have been fine-tuned in step S106 to the corresponding information processing terminal 10 and shares them (S107).

[0062] Each information processing terminal 10 combines the non-learning target portion 51 with the learning target portion 52 shared by the server 20 in step S107, and constructs a large-scale language model 50 (S108).

[0063] Thereafter, each of the information processing terminals 10 executes an application that uses the constructed large-scale language model 50 (S109).

[0064] The flow of the information processing method according to this embodiment has been described above using an example. Note that steps S101 to S109 described above may be executed repeatedly.

[0065] According to the above series of processes, it is possible to realize low-cost and efficient data management for optimizing a large-scale language model.

[0066] <<1.2. Functional Configuration Example>> Next, a functional configuration example of the information processing terminal 10 and the server 20 according to this embodiment will be described.

[0067] First, an example of the functional configuration of the information processing terminal 10 according to this embodiment will be described.

[0068] The information processing terminal 10 according to this embodiment is an example of an information processing device that locally executes an application using a large-scale language model 50 and trains the large-scale language model 50 .

[0069] The information processing terminal 10 according to the present embodiment may be, for example, a smartphone, a tablet, a personal computer (PC), a wearable device, or a car navigation device.

[0070] FIG. 3 is a block diagram showing an example of the functional configuration of the information processing terminal 10 according to this embodiment.

[0071] As shown in FIG. 3, 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.

[0072] (Data Collection Unit 110) The data collection unit 110 according to this embodiment collects data 53 in accordance with the execution of an application using the large-scale language model 50 under the control of the control unit 130.

[0073] For this purpose, the data collection unit 110 according to this embodiment includes various sensors, such as a camera, a microphone, and a biosensor.

[0074] Furthermore, the data collection unit 110 according to this embodiment may collect the data 53 via the network 30, for example, from the Internet or the like.

[0075] (Data Storage Unit 120) The data storage unit 120 according to this embodiment stores the data 53 collected by the data collection unit 110 and the like.

[0076] (Control Unit 130 ) The control unit 130 according to this embodiment controls each component included in the information processing terminal 10 .

[0077] The functions of the control unit 130 according to this embodiment are realized by the cooperation of various processors and memories.

[0078] As shown in FIG. 3 , the control unit 130 according to this embodiment includes a learning unit 140 , an application control unit 150 , and a sharing control unit 160 .

[0079] (Learning Unit 140) The learning unit 140 according to this embodiment executes learning of the large-scale language model 50.

[0080] More specifically, the learning unit 140 according to this embodiment uses the data 53 to perform fine tuning of the learning object unit 52 .

[0081] Furthermore, the learning unit 140 according to this embodiment performs processing such as combining the learning target portion 52 shared from the server 20 with the non-learning target portion 51 held therein, and constructing a large-scale language model 50.

[0082] (Application Control Unit 150 ) The application control unit 150 according to this embodiment controls the execution of an application that uses the large-scale language model 50 .

[0083] The application may be, for example, an agent application.

[0084] The learning unit 140 according to this embodiment is an example of a first learning unit.

[0085] (Sharing Control Unit 160 ) The sharing control unit 160 according to this embodiment controls the sharing of the learning object unit 52 between the information processing terminal 10 and the server 20 .

[0086] For example, the sharing control unit 160 according to this embodiment controls the communication unit 170 so as to transmit the learning object unit 52 on which the learning unit 140 has performed fine tuning to the server 20 .

[0087] Furthermore, for example, the sharing control unit 160 according to this embodiment shares with the learning unit 140 the learning object unit 52 that has been fine-tuned by the server 20 and that the communication unit 170 has received from the server 20 .

[0088] In this way, the sharing control unit 160 according to this embodiment controls the sharing of the learning object unit 52 that has been fine-tuned by either the learning unit 140 or the server 20 with the other.

[0089] The shared control unit 160 according to this embodiment is an example of a first shared control unit.

[0090] (Communication Unit 170) The communication unit 170 according to this embodiment performs information communication with the server 20 and the like via the network 30.

[0091] The functional configuration example of the information processing terminal 10 according to the present embodiment has been described above. Note that the functional configuration described above with reference to Fig. 3 is merely an example, and the functional configuration of the information processing terminal 10 according to the present embodiment is not limited to this example.

[0092] For example, the information processing terminal 10 according to this embodiment may further include a display unit that displays various information in accordance with the control of the control unit 130, an audio output unit that outputs various sounds in accordance with the control of the control unit 130, an operation reception unit that receives operations by the user, and the like.

[0093] The functional configuration of the information processing terminal 10 according to this embodiment can be flexibly modified according to specifications, operation, and the like.

[0094] Next, an example of the functional configuration of the server 20 according to this embodiment will be described.

[0095] The server 20 according to this embodiment is an example of an information processing device that executes an application using a large-scale language model 50 and trains the large-scale language model 50 in the cloud.

[0096] FIG. 4 is a block diagram showing an example of the functional configuration of the server 20 according to this embodiment.

[0097] As shown in FIG. 4, the server 20 according to this embodiment includes a data storage unit 210, a control unit 220, and a communication unit 270.

[0098] (Data Storage Unit 210) The data storage unit 210 according to this embodiment stores the simulation data 54 and the like collected in association with the conversation simulation.

[0099] (Control Unit 220) The control unit 220 according to this embodiment controls each component included in the server 20.

[0100] The functions of the control unit 220 according to this embodiment are realized by the cooperation of various processors and memories.

[0101] As shown in FIG. 4 , the control unit 220 according to this embodiment includes a learning unit 230 , a simulation control unit 240 , an extended data collection unit 250 , and a shared control unit 260 .

[0102] (Learning Unit 230) The learning unit 230 according to this embodiment executes learning of the large-scale language model 50.

[0103] More specifically, the learning unit 230 according to this embodiment performs fine tuning of the learning target unit 52 using data collected in conjunction with the execution of an application using the large-scale language model 50 .

[0104] Furthermore, the learning unit 230 according to this embodiment performs processing such as combining the learning target portion 52 shared from the information processing terminal 10 with the non-learning target portion 51 held therein, and constructing a large-scale language model 50.

[0105] The learning unit 230 according to this embodiment is an example of a second learning unit.

[0106] (Simulation Control Unit 240) The simulation control unit 240 according to this embodiment is an example of an execution control unit that controls the execution of an application that uses the large-scale language model 50.

[0107] As described above, applications using the large-scale language model 50 include applications that perform conversation simulations between large-scale language models 50 .

[0108] That is, the simulation control unit 240 according to this embodiment may control the execution of a conversation simulation between large-scale language models 50 .

[0109] In this case, the learning section 230 according to this embodiment performs fine tuning of the learning object section 52 using simulation data 54 collected in conjunction with the execution of the conversation simulation.

[0110] (Extended Data Collection Unit 250) The extended data collection unit 250 according to this embodiment controls the collection of extended data via the communication unit 270 based on the content of the conversation simulation.

[0111] The learning section 230 according to this embodiment may perform fine tuning of the learning object section 52 further based on the extended data.

[0112] (Sharing Control Unit 260 ) The sharing control unit 260 according to this embodiment controls the sharing of the learning object unit 52 between each of the multiple information processing terminals 10 and the learning unit 230 .

[0113] For example, the sharing control unit 260 according to this embodiment controls the communication unit 270 so as to transmit each of the learning target units 52 on which the learning unit 230 has performed fine tuning to the corresponding information processing terminal 10 .

[0114] Also, for example, the sharing control unit 260 according to this embodiment shares with the learning unit 230 the learning target units 52 that have been fine-tuned by each of the information processing terminals 10 and that are received by the communication unit 270 from each of the information processing terminals 10 .

[0115] In this way, the sharing control unit 260 according to this embodiment controls the sharing of the learning object unit 52 that has been fine-tuned by either the learning unit 230 or the information processing terminal 10 with the other.

[0116] The shared control unit 260 according to this embodiment is an example of a second shared control unit.

[0117] (Communication Unit 270) The communication unit 270 according to this embodiment performs information communication with the information processing terminal 10 and the like via the network 30.

[0118] The functional configuration of the server 20 according to the present embodiment has been described above. Note that the functional configuration described with reference to Fig. 4 is merely an example, and the functional configuration of the server 20 according to the present embodiment is not limited to this example.

[0119] For example, the server 20 according to this embodiment may further include a display unit that displays various information in accordance with the control of the control unit 220, an audio output unit that outputs various sounds in accordance with the control of the control unit 220, an operation reception unit that receives operations by the user, and the like.

[0120] The functional configuration of the server 20 according to this embodiment can be flexibly modified according to specifications, operation, and the like.

[0121] <<1.3. Detailed Functions>> Next, the functions of the information processing system 1 according to this embodiment will be described in detail.

[0122] First, a use case to which the information processing system 1 according to this embodiment is applied will be described, and then the details of the functions will be described along with the use case.

[0123] 5 and 6 are diagrams for explaining use cases to which the information processing system 1 according to this embodiment is applied.

[0124] In this use case, it is assumed that the information processing terminal 10 is an information processing device mounted on a vehicle, such as a car navigation system, and an agent application using a large-scale language model 50 is utilized by a user riding in the vehicle.

[0125] 5 and 6, Alice and Bob are given as examples of the users. Eve in the same use case is an example of an agent in an agent application using the large-scale language model 50.

[0126] For example, in Phase 1, the information processing terminal 10 can detect the subject of the conversation from Alice's movements.

[0127] The information processing terminal 10 acquires an image of the object using a camera, and identifies the object as "Sky Tree" using image captioning technology.

[0128] In this way, the information processing terminal 10 according to this embodiment can complement information not included in the conversation content and provide information that is in line with the conversation content.

[0129] Furthermore, the information processing terminal 10 acquires the conversation content using speech-to-text technology throughout all phases. Furthermore, the information processing terminal 10 may sense the user's emotions, such as surprise or excitement, from the tone, manner of speaking, speed, etc., of the speech throughout all phases.

[0130] Furthermore, the information processing terminal 10 can acquire non-confidential open information, such as road congestion status in Phase 2 and weather and date and time information in Phase 3, from the Internet or the like.

[0131] According to these, for example, in Phase 2, the information processing terminal 10 can consider the content of the open information in addition to the content of the utterance and the emotion, and can propose a useful means for achieving the goal.

[0132] Furthermore, for example, in Phase 3, the information processing terminal 10 can develop a positive conversation by using the expressive power of the large-scale language model 50 while taking into account the content of the conversation so far.

[0133] Furthermore, for example, in Phases 4 and 5, the information processing terminal 10 can acquire new information about Alice by expanding the conversation and asking more in-depth questions. By diverting such new information about the user as a learning topic, it is expected that the optimization will be further accelerated.

[0134] The use cases to which the information processing system 1 according to this embodiment is applied have been described above.

[0135] Next, the collection and storage of the data 53 according to this embodiment will be described in detail.

[0136] 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 or the server 20. This makes it possible to appropriately protect sensitive data such as personal information.

[0137] The data storage unit 120 of each information processing terminal 10 may include a user DB that stores data relating to users among the data 53, and an agent DB that stores data relating to agents among the data 53.

[0138] Fig. 7 is a diagram showing an example of data stored in a user DB according to this embodiment, and Fig. 8 is a diagram showing an example of data stored in an agent DB according to this embodiment.

[0139] The user DB and agent DB commonly store the ID, previous ID, and date and time.

[0140] ID is the identifier of the corresponding data group (record). Previous ID is the identifier of the record immediately before the corresponding record in chronological order. Date Time is the date and time when the corresponding record was acquired.

[0141] By saving the above information, it is possible to understand the order and flow of the conversation, as well as the date and time.

[0142] Next, data specific to the user DB will be described.

[0143] The user DB may store spoken text, specker, sentiment, and target detection.

[0144] Spoken text is text data obtained using Local STT (Speech-To-Text) technology.

[0145] Specker is a speaker (user) obtained using local speaker recognition technology.

[0146] Sentiment is the result of analyzing the emotion of an utterance obtained from various sensor data. Sentiment may be used to calculate the speaker's reaction score to the agent's utterance.

[0147] Target detection is metadata related to the subject of the conversation. In this example, the metadata includes the location where the conversation took place, the direction of the Skytree visible from that location, and an image of the Skytree.

[0148] On the other hand, the agent DB may store generated text, image captioning, and RAG.

[0149] Generated text is text data generated and spoken by the Agent using Local TTS (Text-To-Speech) technology.

[0150] Image captioning is the caption information (text data) inferred by Local AI for the captured image.

[0151] The RAG is reference information obtained from the Internet related to the topic. In this example, the RAG contains general information about the Skytree obtained from a website.

[0152] The data 53 according to this embodiment has been described above with specific examples. The data 53 according to this embodiment is used as learning data used for fine tuning of the learning object unit 52 by the learning unit 140 in step S102 shown in FIG.

[0153] Furthermore, the data 53 according to this embodiment may be used as reference information when performing inference (e.g., inference of speech content) using the large-scale language model 50 in executing an application using the large-scale language model 50 in step S109 shown in FIG. 2.

[0154] Note that the data 53 stored in the data storage unit 120 may be partially or entirely deletable at the user's request to ensure the user's right to be forgotten. Deletion of the data 53 may be achieved, for example, by deleting related records based on the results of a similarity search.

[0155] Next, fine-tuning of the large-scale language model 50 according to this embodiment will be described in detail.

[0156] In general, it is not uncommon for a large-scale language model to have more than 7 billion parameters, and updating all of the parameters through additional learning requires enormous hardware resources.

[0157] For example, when fine-tuning all parameters of GPT-3, which has 175 billion parameters, is performed, 1.2 TB of memory is required, and the storage required to save the model is 350 GB / epoch.

[0158] Therefore, in the information processing method according to this embodiment, the target of fine tuning is limited to the learning target portion 52, which is a part of the large-scale language model 50, thereby reducing the required hardware resources.

[0159] As described above, the large-scale language model 50 according to this embodiment is made up of a non-learning part 51 and a learning part 52 .

[0160] FIG. 9 is a diagram showing a more specific example of the configuration of the large-scale language model 50 according to this embodiment.

[0161] As shown in FIG. 9, the large-scale language model 50 according to this embodiment may be composed of an original large-scale language model 510 and an adapter 520, which is a layer added in parallel to the original large-scale language model 510.

[0162] The original large-scale language model 510 is an example of the non-learning target portion 51. The adapter 520 is an example of the learning target portion 52.

[0163] FIG. 10 is a diagram for explaining the difference between the original large-scale language model 510 and the adapter 520.

[0164] As shown in FIG. 10, one of the features of the adapter 520 is that it is small in size, while the original large-scale language model 510 is large in size.

[0165] Furthermore, in the original large-scale language model 510, the initial values ​​of the weights are unique for each model, whereas in the adapter 520, the initial values ​​of the weights are 0 or random.

[0166] As described above, the original large-scale language model 510 is not subject to training (fine tuning), but the adapter 520 is subject to training and the weights are updated.

[0167] Therefore, during inference, the original large-scale language model 510 uses weights specific to each model, and the adapter 520 uses weights updated by learning.

[0168] Fine tuning of the adapter 520 according to this embodiment may be achieved using a technique such as LoRA (Low-Rank Adaptation), for example.

[0169] For example, if fine-tuning is performed using LoRA to update some of the parameters of GPT-3, which has 175 billion parameters, the required memory is reduced to 350 GB, and the storage required to save the model is reduced to 35 MB / epoch.

[0170] The above describes specific examples of the non-learning target portion 51 and the learning target portion 52 according to this embodiment. Next, detailed descriptions will be given of fine tuning of the learning target portion 52 by the information processing terminal 10 according to this embodiment and fine tuning of the learning target portion 52 by the server 20 according to this embodiment.

[0171] FIG. 11 is a diagram for explaining the difference in the execution environment of fine tuning between the information processing terminal 10 and the server 20. In FIG.

[0172] As shown in FIG. 11, the information processing terminal 10 has fewer hardware resources than the server 20 .

[0173] Furthermore, as described above, the server 20 can use the simulation data collected as the conversation simulation is executed as learning data, and therefore the size of the learning data is expected to be larger than that of the information processing terminal 10.

[0174] Furthermore, in the information processing terminal 10, the learning time is shorter than that of the server 20 in order to ensure time for use by the user.

[0175] Furthermore, since the information processing terminal 10 can use data 53 related to actual conversations by users, rather than simulations, as training data, it is expected that the quality of the training data set will be higher than that of the server 20.

[0176] Taking the above differences into consideration, the learning unit 140 of the information processing terminal 10 may perform fine tuning when the following conditions are satisfied as a trigger.

[0177] A predetermined time has passed since the previous fine tuning. Sufficient time for using the information processing terminal 10 can be secured. Sufficient learning data (data 53) has been secured. The information processing terminal 10 is being charged. It is a time period when the information processing terminal 10 is not used frequently, such as late at night.

[0178] Furthermore, the learning unit 140 may perform small-scale fine tuning by using only data 53 that has not previously been used as learning data, or by setting small hyper-parameters related to learning, such as batch size.

[0179] On the other hand, the learning unit 230 of the server 20 may perform fine tuning when the learning target unit 52 is shared by the information processing terminal 10 as a trigger.

[0180] Furthermore, the learning unit 230 may perform large-scale, long-term fine tuning using simulation data 54 collected as conversation simulations are performed, as far as resources permit.

[0181] As described above, according to the information processing method of this embodiment, it is possible to realize low-cost and efficient fine tuning according to the hardware characteristics of the information processing terminal 10 and the server 20, respectively.

[0182] Next, the control of the conversation simulation according to this embodiment will be described in detail.

[0183] As described above, the simulation control unit 240 of the server 20 according to this embodiment controls the execution of an application that uses the large-scale language model 50 .

[0184] FIG. 12 is a diagram for explaining the control of the conversation simulation according to this embodiment.

[0185] First, the learning object unit 52 is shared by the server 20 from a plurality of information processing terminals 10 .

[0186] FIG. 12 illustrates a case where a plurality of learning objects 52 including a learning object 52A, a learning object 52J, a learning object 52M, and a learning object 52X are shared by the server 20.

[0187] The data 53 used for fine tuning for the plurality of learning target units 52 differs for each information processing terminal 10. Therefore, there are basically no plurality of learning target units 52 with the exact same weights.

[0188] Next, the learning unit 230 combines each of the plurality of learning target portions 52 with the non-learning target portion 51 to construct a plurality of large-scale language models 50 .

[0189] In the example shown in FIG. 12, the learning unit 230 constructs a plurality of large-scale language models 50 including a large-scale language model 50A2, a large-scale language model 50J2, a large-scale language model 50M2, and a large-scale language model 50X2.

[0190] Next, the simulation control unit 240 performs matching of the large-scale language model 50 that executes the conversation simulation.

[0191] At this time, the simulation control unit 240 may perform matching based on the similarity of the large-scale language models 50 in order to further deepen the configuration of each of the large-scale language models 50 .

[0192] The similarity may be, for example, the similarity of the weights of the learning target portion 52, the similarity of the probability distribution of the inference results by the large-scale language model 50, or the like.

[0193] In the example shown in FIG. 12, the simulation control unit 240 matches the large-scale language model 50A2 with the large-scale language model 50M2, and matches the large-scale language model 50J2 with the large-scale language model 50X2 based on the similarity.

[0194] The matching index is not limited to similarity, and various indices suitable for the use case can be adopted.

[0195] After the matching is completed, the simulation control unit 240 executes the conversation simulation. During the execution of the conversation simulation, simulation data 54 is collected.

[0196] 12, simulation data 54AM is collected as a conversation simulation between large-scale language models 50A2 and 50M2 is executed. The simulation data 54AM is used for fine-tuning the learning target portions 52A and 52M.

[0197] 12, simulation data 54JX is collected as a conversation simulation between the large-scale language models 50J2 and 50X2 is executed. The simulation data 54JX is used for fine-tuning the learning target portions 52J and 52X.

[0198] The conversation simulation according to this embodiment can be performed for a long time and at high speed using abundant resources on the cloud, making it possible to collect a large amount of conversation data (simulation data 54).

[0199] Furthermore, according to the conversation simulation of this embodiment, even conversation data that is difficult to collect on the edge side can be easily reproduced.

[0200] Furthermore, the extended data collection unit 250 according to this embodiment collects extended data based on the content of the conversation simulation.

[0201] For example, the extended data collection unit 250 may search and collect words that appear in the conversation simulation on the Internet.

[0202] Furthermore, for example, the extended data collection section 250 may search and collect other words that are highly similar to the words that appeared in the conversation simulation on the Internet.

[0203] The learning section 230 according to this embodiment performs fine tuning of the learning object section 52 using the simulation data 54 and the extended data collected as described above.

[0204] The control of the conversation simulation according to this embodiment has been described above. Next, the flow of the conversation simulation according to this embodiment and fine tuning using simulation data will be summarized with reference to FIG.

[0205] FIG. 13 is a flowchart showing an example of the flow of conversation simulation and fine tuning using simulation data according to this embodiment.

[0206] In the example shown in FIG. 13, first, each information processing terminal 10 shares its learning target unit 52 with the server 20 (S201).

[0207] The learning unit 230 of the server 20 combines each of the learning target parts 52 shared in step S201 with the non-learning target parts 51 to construct a plurality of large-scale language models 50 (S202).

[0208] Next, the simulation control unit 240 matches a large-scale language model 50 for conducting a conversation simulation from among the plurality of large-scale language models 50 constructed in S202 (S203).

[0209] The simulation control unit 240 controls the execution of a conversation simulation between the large-scale language models 50 matched in step S203, and collects simulation data 54 (S204).

[0210] Next, the extended data collection unit 250 collects extended data based on the simulation data 54 collected in step S204 (S205).

[0211] The learning unit 230 fine-tunes the learning target portion 52 of the large-scale language model 50 that performed the conversation simulation in step S204 using the simulation data 54 collected in step S204 and the extended data collected in step S205 (S206).

[0212] Next, a specific example of the conversation simulation according to this embodiment will be given below: Fig. 14 is a diagram showing a specific example of the conversation simulation according to this embodiment.

[0213] 14 indicates the large scale language model 50J2 in FIG. 12, and LLM X in FIG. 14 indicates the large scale language model 50X2 in FIG.

[0214] The simulation control unit 240 according to this embodiment may control the large-scale language model 50 so that questions that link events and emotions are asked in the conversation simulation.

[0215] In the example shown in FIG. 14, the simulation control unit 240 controls the large-scale language model 50J2 in Phase 1 to ask a question asking whether or not the subject has ever been moved by something.

[0216] In addition, the simulation control unit 240 according to this embodiment may control the large-scale language model 50 so that, when a further response is made to an answer to a question in a conversation simulation, a response is made that includes emotions and additional information.

[0217] In the example shown in FIG. 14, the simulation control unit 240 controls the large-scale language model 50J2 in Phase 2 to provide a response that shows empathy and includes additional information about the view of the Skytree at dusk and in the daytime.

[0218] In addition, the simulation control unit 240 according to this embodiment may control the large-scale language model so that, when additional information is obtained in a conversation simulation, a response is made indicating whether the person is positive or negative about the additional information.

[0219] In the example shown in FIG. 14, in Phase 2, the simulation control unit 240 controls the large-scale language model 50X2 to respond affirmatively to the additional information about the Sky Tree.

[0220] Furthermore, the simulation control unit 240 according to this embodiment may control the large-scale language model 50 so that the received question is also posed to the other party in the conversation simulation.

[0221] In the example shown in FIG. 14, the simulation control unit 240 controls the large-scale language model 50X2 in Phase 2 so that the question received in Phase 1 is also sent to the large-scale language model 50J2.

[0222] According to the control described above, it is possible to collect a large amount of data more efficiently than in a real conversation.

[0223] As described above, the extended data collection unit 250 according to this embodiment may also collect extended data based on the content of the conversation simulation.

[0224] For example, the extended data collection unit 250 may search the Internet for and collect words such as "Sky Tree" and "Atami" that appear in the conversation simulation.

[0225] Furthermore, for example, the extended data collection unit 250 may search and collect on the Internet words such as "Tokyo Tower" which is highly similar to the word "Skytree" that appeared in the conversation simulation, and "Shizuoka" which is highly similar to the word "Atami" that appeared in the conversation simulation.

[0226] By collecting the simulation data 54 and the extended data as described above, it is possible to greatly expand the learning data and realize large-scale learning for personal optimization.

[0227] <<1.4. Application Examples>> Next, application examples of the information processing system 1 according to this embodiment will be described.

[0228] As described in the above use case, the information processing system 1 according to this embodiment is applicable to an agent application installed in a vehicle.

[0229] FIG. 15 is a diagram for explaining an example in which the information processing system 1 according to this embodiment is applied to an agent application installed in a vehicle.

[0230] In this application example, the information processing terminal 10A may be, for example, a car navigation device equipped with an agent function and mounted on a vehicle.

[0231] When the vehicle departs, the information processing terminal 10A downloads the latest learning object part 52A from the server 20.

[0232] Furthermore, while driving, the information processing terminal 10 uses the large-scale language model 50A1 to respond (inference) to prompts from the user in the vehicle.

[0233] The information processing terminal 10 can respond based on the individual's tastes and preferences using the personally optimized large-scale language model 50A1.

[0234] Furthermore, by using the accumulated data 53A as the context of the prompt, it is expected that the accuracy of the response will be improved.

[0235] Furthermore, the information processing terminal 10 stores new data 53A during driving. Specific examples of the data 53 are as described with reference to FIGS.

[0236] After returning home, the information processing terminal 10 performs fine tuning of the learning object portion 52A using the accumulated data 53A, and uploads the fine-tuned learning object portion 52A to the server 20.

[0237] Thereafter, the information processing terminal 10 may be powered off until the next time it is used.

[0238] On the other hand, the server 20 executes a series of processes including the conversation simulation and fine tuning described with reference to FIGS.

[0239] The learning object part 52A that has been fine-tuned by the server 20 is downloaded to the information processing terminal 10 the next time the vehicle is used.

[0240] The above describes an example in which the information processing system 1 according to this embodiment is applied to an agent application installed in a vehicle.

[0241] Next, an example in which the information processing system 1 according to this embodiment is applied to fan engagement will be described with reference to FIG.

[0242] In this application example, data 53 may include posts posted by a fan F1 of an artist A1 on an application such as SNS, NFTs (Non-Fungible Tokens) held in the application, and the like.

[0243] In addition, images taken by a camera 112 installed at the event venue and audio recorded by a microphone 114 installed at the event venue may also be collected as data 53.

[0244] The data 53 collected as described above may be provided to artist A1 as a profile of fan F1.

[0245] In addition, the information processing terminal 10 can suggest to the artist A1 what the fan F1 likes, what words he / she would like to hear, what actions he / she would like him / her to take, etc., by inference using a large-scale language model 50 that has been fine-tuned using the data 53.

[0246] Furthermore, information regarding interactions between artist A1 who has received the proposal as described above and fan F1 at an event venue or the like may be collected by camera 112 and microphone 114 and stored as data 53 .

[0247] By collecting the data 53 as described above, it is possible to realize inferences that are optimized for the target fans.

[0248] 2. Hardware Configuration Example Next, a hardware configuration example common to the information processing terminal 10 and the server 20 according to an embodiment of the present disclosure will be described. Fig. 17 is a block diagram showing a hardware configuration example of an information processing device 90 according to an embodiment of the present disclosure. The information processing device 90 may be a device having a hardware configuration equivalent to that of each of the above devices.

[0249] 17 , the information processing device 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 880, a drive 881, a connection port 882, and a communication device 883. Note that the hardware configuration shown here is an example, and some of the components may be omitted. Furthermore, the information processing device 90 may include further components in addition to the components shown here.

[0250] (Processor 871) The processor 871 functions, for example, as an arithmetic processing device or control device, and controls the overall operation of each component or part of it based on various programs recorded in the ROM 872, RAM 873, storage 880, or removable storage medium 901.

[0251] (ROM 872, RAM 873) The ROM 872 is a means for storing programs to be read into the processor 871, data to be used for calculations, etc. The RAM 873 temporarily or permanently stores, for example, the programs to be read into the processor 871 and various parameters that change as appropriate when the programs are executed.

[0252] (Host bus 874, bridge 875, external bus 876, interface 877) The processor 871, ROM 872, and RAM 873 are connected to one another via, for example, a host bus 874 that is capable of high-speed data transmission. On the other hand, the host bus 874 is connected to, for example, an external bus 876 that has a relatively low data transmission speed via a bridge 875. Furthermore, the external bus 876 is connected to various components via an interface 877.

[0253] (Input Device 878) For example, a mouse, keyboard, touch panel, button, switch, lever, etc. are used as the input device 878. Furthermore, a remote controller (hereinafter referred to as a remote control) capable of transmitting control signals using infrared rays or other radio waves may also be used as the input device 878. The input device 878 also includes an audio input device such as a microphone.

[0254] (Output Device 879) The output device 879 is a device capable of visually or audibly notifying the user of acquired information, such as a display device such as a CRT (Cathode Ray Tube), LCD, or organic EL, an audio output device such as a speaker or headphones, a printer, a mobile phone, or a facsimile. The output device 879 according to the present disclosure also includes various vibration devices capable of outputting tactile stimulation.

[0255] (Storage 880) The storage 880 is a device for storing various types of data. For example, a magnetic storage device such as a hard disk drive (HDD), a semiconductor storage device, an optical storage device, or a magneto-optical storage device may be used as the storage 880.

[0256] (Drive 881) The drive 881 is a device that reads information recorded on a removable storage medium 901 such as a magnetic disk, an optical disk, a magneto-optical disk, or a semiconductor memory, or writes information to the removable storage medium 901.

[0257] (Removable storage medium 901) The removable storage medium 901 is, for example, a DVD medium, a Blu-ray (registered trademark) medium, an HD DVD medium, various semiconductor storage media, etc. Of course, the removable storage medium 901 may also be, for example, an IC card equipped with a contactless IC chip, an electronic device, etc.

[0258] (Connection Port 882) The connection port 882 is a port for connecting an external device 902, such as a USB (Universal Serial Bus) port, an IEEE 1394 port, a SCSI (Small Computer System Interface), an RS-232C port, or an optical audio terminal.

[0259] (Externally Connected Device 902) The externally connected device 902 is, for example, a printer, a portable music player, a digital camera, a digital video camera, or an IC recorder.

[0260] (Communication device 883) The communication device 883 is a communication device for connecting to a network, such as a communication card for wired or wireless LAN, Bluetooth (registered trademark), or WUSB (Wireless USB), a router for optical communication, a router for ADSL (Asymmetric Digital Subscriber Line), or a modem for various types of communication.

[0261] 3. Summary As described above, the server 20 according to an embodiment of the present disclosure includes a learning unit 230 that learns a large-scale language model 50 including a learning target portion 52 and a non-learning target portion 51, and a sharing control unit 260 that controls sharing of the learning target portion 52 with the information processing terminal 10. The learning unit 230 performs fine tuning of the learning target portion 52 using data 53 collected in conjunction with execution of an application that uses the large-scale language model 50. The sharing control unit 260 controls sharing of the learning target portion 52 that has been fine-tuned by either the learning unit 230 or the information processing terminal 10 with the other.

[0262] According to the above configuration, it is possible to realize low-cost and efficient data management for optimizing a large-scale language model.

[0263] Although the preferred embodiments of the present disclosure have been described in detail above with reference to the accompanying drawings, the technical scope of the present disclosure is not limited to such examples. It is clear that a person skilled in the art of the present disclosure can conceive of various modified or altered examples within the scope of the technical idea described in the claims, and it is understood that these also naturally fall within the technical scope of the present disclosure.

[0264] Furthermore, the steps of the processes described in this disclosure do not necessarily have to be processed in chronological order according to the order shown in the flowcharts or sequence diagrams. For example, the steps of the processes of each device may be processed in an order different from the order shown, or may be processed in parallel.

[0265] Furthermore, the series of processes performed by each device described in this disclosure may be realized using software, hardware, or a combination of software and hardware. The programs constituting the software may be provided, for example, inside or outside each device and stored in advance in a non-transitory computer-readable medium. Each program is then loaded into RAM when executed by a computer and executed by various processors. Examples of the storage medium include a magnetic disk, an optical disk, a magneto-optical disk, and a flash memory. Furthermore, the computer programs may be distributed, for example, via a network, without using a storage medium.

[0266] Furthermore, the effects described herein are merely descriptive or exemplary and are not limiting. In other words, the technology according to the present disclosure may achieve other effects that are apparent to those skilled in the art from the description of this specification, in addition to or in place of the above-described effects.

[0267] Note that the following configurations also fall within the technical scope of the present disclosure. (1) An information processing device comprising: a learning unit that learns a large-scale language model consisting of a learning target unit and a non-learning target unit; and a sharing control unit that controls sharing of the learning target unit with an edge, wherein the learning unit fine-tunes the learning target unit using data collected in connection with execution of an application using the large-scale language model, and the sharing control unit controls sharing of the learning target unit fine-tuned by either the learning unit or the edge with the other. (2) The sharing control unit shares with the learning unit the learning target unit fine-tuned by the edge using data received from the edge and collected in connection with execution of an application using the large-scale language model by the edge, and the learning unit combines the shared learning target unit and the non-learning target unit to construct the large-scale language model. (3) The information processing device according to any one of (1) or (2), further comprising: an execution control unit that controls execution of an application using the large-scale language model. (4) The information processing device according to (3), wherein the execution control unit controls execution of the conversation simulation between the large-scale language models, and the learning unit performs fine-tuning of the training target unit using simulation data collected in conjunction with execution of the conversation simulation. (5) The information processing device according to (4), wherein the sharing control unit shares the training target unit received from each of the plurality of edges with the learning unit, and the learning unit constructs the plurality of large-scale language models corresponding to each of the plurality of edges by combining each of the shared training target units with the non-training target unit. (6) The information processing device according to (5), wherein the execution control unit performs matching of the large-scale language model for executing the conversation simulation from the plurality of large-scale language models corresponding to each of the edges. (7) The information processing device according to (6), wherein the execution control unit performs the matching based on the similarity of the large-scale language models.(8) The information processing device according to (7), wherein the similarity includes at least one of a similarity of weights of the learning target portion or a similarity of inference results of the large-scale language model. (9) The information processing device according to any one of (4) to (8), wherein the execution control unit controls the large-scale language model so that a question linking an event and an emotion is asked in the conversation simulation. (10) The information processing device according to any one of (4) to (9), wherein, when a further response is made to an answer to a question in the conversation simulation, the execution control unit controls the large-scale language model so that a response including an emotion and additional information is made. (11) The information processing device according to any one of (4) to (10), wherein, when additional information is obtained in the conversation simulation, the execution control unit controls the large-scale language model so that a response indicating whether the additional information is positive or negative is made to the additional information. (12) The information processing device according to any one of (4) to (11), wherein the execution control unit controls the large-scale language model to ask the received question to the other party in the conversation simulation. (13) The information processing device according to any one of (4) to (12), further comprising: an extended data collection unit that collects extended data based on the content of the conversation simulation, wherein the learning unit further uses the extended data to fine-tune the learning target unit. (14) The information processing device according to any one of (1) to (13), wherein the learning target unit is an adapter added to the non-learning target unit, which is an original large-scale language model.(15) An information processing method, comprising: a processor performing training of a large-scale language model consisting of a training target portion and a non-training target portion; and controlling the sharing of the training target portion between edges; wherein the performing of the training further comprises performing fine-tuning of the training target portion using data collected in conjunction with the execution of an application using the large-scale language model; and controlling the sharing further comprises controlling the sharing of the training target portion that has been fine-tuned by one of the training or the edges with the other. (16) An information processing system comprising: an information processing terminal; and a server; wherein the information processing terminal comprises: a first learning unit that learns a large-scale language model consisting of a learning object portion and a non-learning object portion; and a first sharing control unit that controls sharing of the learning object portion with the server; wherein the first learning unit fine-tunes the learning object portion using data collected in conjunction with execution of an application using the large-scale language model; and the first sharing control unit controls sharing of the learning object portion fine-tuned by the first learning unit with the server; and wherein the server comprises: a second learning unit that learns the large-scale language model; and a second sharing control unit that controls sharing of the learning object portion with the information processing terminal; wherein the second learning unit fine-tunes the learning object portion using data collected in conjunction with execution of an application using the large-scale language model; and the second sharing control unit controls sharing of the learning object portion fine-tuned by the second learning unit with the information processing terminal.

[0268] REFERENCE SIGNS LIST 1 Information processing system 10 Information processing terminal 130 Control unit 140 Learning unit 150 Application control unit 160 Shared control unit 20 Server 220 Control unit 230 Learning unit 240 Simulation control unit 250 Extended data collection unit 260 Shared control unit 50 Large-scale language model 51 Non-learning target unit 52 Learning target unit 53 Data 54 Simulation data

Claims

1. An information processing device comprising: a learning unit that learns a large-scale language model consisting of a learning target portion and a non-learning target portion; and a sharing control unit that controls the sharing of the learning target portion between an edge and the learning unit, wherein the learning unit fine-tunes the learning target portion using data collected in conjunction with the execution of an application using the large-scale language model, and the sharing control unit controls the sharing of the learning target portion that has been fine-tuned by either the learning unit or the edge with the other.

2. The information processing device described in claim 1, wherein the sharing control unit shares with the learning unit the learning target portion that has been fine-tuned by the edge using data received from the edge and collected in connection with the execution of an application using the large-scale language model by the edge, and the learning unit combines the shared learning target portion and the non-learning target portion to construct the large-scale language model.

3. The information processing device according to claim 1, further comprising: an execution control unit that controls execution of an application that uses the large-scale language model.

4. The information processing device according to claim 3, wherein the execution control unit controls the execution of a conversation simulation between the large-scale language models, and the learning unit performs fine tuning of the learning target unit using simulation data collected in conjunction with the execution of the conversation simulation.

5. The information processing device described in claim 4, wherein the sharing control unit shares the learning target parts received from each of the plurality of edges with the learning unit, and the learning unit combines each of the shared plurality of learning target parts with the non-learning target parts to construct a plurality of the large-scale language models corresponding to each of the plurality of edges.

6. The information processing device according to claim 5, wherein the execution control unit performs matching of the large-scale language model for executing the conversation simulation from a plurality of the large-scale language models corresponding to each of the edges.

7. The information processing device according to claim 6, wherein the execution control unit performs the matching based on the similarity of the large-scale language model.

8. The information processing device according to claim 7, wherein the similarity includes at least one of a similarity of weights of the learning object parts and a similarity of inference results of the large-scale language model.

9. The information processing device according to claim 4, wherein the execution control unit controls the large-scale language model so that questions that link events and emotions are asked in the conversation simulation.

10. The information processing device according to claim 4, wherein the execution control unit controls the large-scale language model so that, when a further response is made to an answer to a question in the conversation simulation, a response is made that includes emotion and additional information.

11. The information processing device according to claim 4, wherein the execution control unit controls the large-scale language model so that, when additional information is obtained in the conversation simulation, a response indicating whether the additional information is positive or negative is given.

12. The information processing device according to claim 4, wherein the execution control unit controls the large-scale language model so that the received question is also posed to the other party in the conversation simulation.

13. The information processing device according to claim 4, further comprising an extended data collection unit that collects extended data based on the content of the conversation simulation, wherein the learning unit further uses the extended data to perform fine tuning of the learning target portion.

14. The information processing device according to claim 1, wherein the training target portion is an adapter added to the non-training target portion, which is an original large-scale language model.

15. An information processing method, comprising: a processor training a large-scale language model consisting of a training target portion and a non-training target portion; and controlling the sharing of the training target portion between edges; wherein the training further comprises fine-tuning the training target portion using data collected in conjunction with the execution of an application using the large-scale language model; and the controlling the sharing further comprises controlling the sharing of the training target portion that has been fine-tuned by one of the edges with the other.

16. An information processing system comprising: an information processing terminal; and a server, wherein the information processing terminal comprises: a first learning unit that learns a large-scale language model consisting of a learning object portion and a non-learning object portion; and a first sharing control unit that controls sharing of the learning object portion with the server, wherein the first learning unit fine-tunes the learning object portion using data collected in conjunction with execution of an application using the large-scale language model, and the first sharing control unit controls sharing of the learning object portion fine-tuned by the first learning unit with the server, wherein the server comprises: a second learning unit that learns the large-scale language model; and a second sharing control unit that controls sharing of the learning object portion with the information processing terminal, wherein the second learning unit fine-tunes the learning object portion using data collected in conjunction with execution of an application using the large-scale language model, and the second sharing control unit controls sharing of the learning object portion fine-tuned by the second learning unit with the information processing terminal.

Images