Information processing system, information processing method, and program

Abstract

The present technology relates to an information processing system, an information processing method, and a program that enable a user to immediately receive a proposal of information pertaining to an optimal architecture only by inputting a prompt relating to the architecture of a neural network. An acquisition unit acquires a prompt relating to the architecture of a neural network. A processing unit inputs the prompt into a fine-tuned language model to output recommendation information pertaining to an architecture of an available model. The present technology can be applied to, for example, a model provision device or the like that outputs recommendation information on the basis of a prompt which has been input by a model developer by characters or voice and transmits, to a model development device of the model developer, an available model corresponding to the recommendation information.

Description

Information processing system, information processing method, and program

[0001] The present technology relates to an information processing system, an information processing method, and a program, and in particular to an information processing system, an information processing method, and a program that enable a user to instantly receive suggestions regarding optimal architecture simply by inputting a prompt regarding the architecture of a neural network.

[0002] In the machine learning model (AI (Artificial Intelligence) model) development platform, a user can design a machine learning model or use a machine learning model that is registered by default.

[0003] However, it is difficult for users with little knowledge or experience in developing machine learning models to design machine learning models. Meanwhile, default machine learning models cannot meet the user's detailed requirements. Therefore, it is difficult for such users to use the machine learning model that is optimal for the required task, the data used, and the execution device. Therefore, there is a need for the development of an AI assistant that allows users to receive suggestions for optimal machine learning models through simple user interaction.

[0004] Therefore, there is a method for selecting and proposing a DNN (Deep Neural Network) appropriate for a task specified by a user from among the DNNs registered in Hugging Face in chat (see, for example, Non-Patent Document 1). However, with this method, the DNNs proposed are limited to those registered in Hugging Face. Furthermore, since the user cannot specify hardware constraints such as model size or FLOPs (Floating-point Operations), it is difficult to propose a DNN appropriate for the hardware constraints.

[0005] There is also a method for optimizing DNN architecture using GPT-4 (see, for example, Non-Patent Document 2). In this method, GPT-4 performs Neural Architecture Search (NAS) by acquiring a DNN architecture to propose based on input DNN conditions, and repeatedly proposes DNNs based on that architecture until the desired performance is achieved. However, because GPT-4 is not trained on a dataset specialized for proposing DNN architectures, the performance of the DNNs it can propose is limited. Because the user must repeatedly submit queries until a DNN with the desired performance is proposed, it takes a long time for the optimal DNN to be proposed to the user.

[0006] On the other hand, there are also methods that use language models to modify the architecture of DNNs (see, for example, LLMATIC and EvoPrompting, Non-Patent Documents 3 and 4). In this method, the base architecture of a DNN expressed in a program is modified using prompts, and the architecture is optimized by performing training and evaluation on the DNN with the modified architecture. However, because architecture optimization is performed by modifying the base architecture, the flexibility and creativity of the optimized DNN are limited. Since there is no information on the accuracy of the DNN with the modified architecture, it is necessary to optimize the architecture while checking the accuracy using evolution search, etc., and it takes a long time before the optimal architecture is proposed to the user.

[0007] Shen, Yongliang, et al., "HuggingGPT: Solving AI Tasks with ChatGPT and its Friends in HuggingFace", arXiv preprint arXiv:2303.17580, 2023Zheng, Mingkai, et al., "Can GPT-4 Perform Neural Architecture Search?", arXiv preprint arXiv:2304.10970, 2023.Nasir, Muhammad U, et al., "LLMatic: Neural Architecture Search via Large Language Models and Quality-Diversity Optimization", arXiv preprint arXiv:2306.01102, 2023Chen, Angelica, David M. Dohan, and David R. So, "EvoPrompting: Language Models for Code-Level Neural Architecture Search". arXiv preprint arXiv:2302.14838, 2023

[0008] There is no method in place that allows a user to simply input a prompt regarding the architecture of a neural network such as a DNN and instantly receive suggestions regarding the optimal architecture.

[0009] The present technology was developed in light of these circumstances, and enables users to instantly receive suggestions regarding the optimal architecture simply by inputting a prompt regarding the neural network architecture.

[0010] An information processing system or program according to one aspect of the present technology is an information processing system or a program for causing a computer to function as an information processing system, which includes an acquisition unit that acquires prompts related to the architecture of a neural network, and a processing unit that outputs recommendation information related to the architecture of the neural network by inputting the prompts into a fine-tuned language model.

[0011] An information processing method according to one aspect of the present technology is an information processing method including an information processing system acquiring prompts related to a neural network architecture and inputting the prompts into a fine-tuned language model to output recommendation information related to the neural network architecture.

[0012] In one aspect of the present technology, a prompt regarding a neural network architecture is obtained, and the prompt is input to a fine-tuned language model to output recommendation information regarding a neural network architecture.

[0013] 1 is a diagram illustrating an example of the configuration of a model providing device. FIG. 1 is a diagram illustrating a first example of an interactive screen. FIG. 2 is a diagram illustrating a second example of an interactive screen. FIG. 3 is a diagram illustrating a third example of an interactive screen. FIG. 4 is a diagram illustrating a fourth example of an interactive screen. A flowchart explaining recommendation processing. A block diagram showing an example of the configuration of a model generating device. FIG. 1 is a diagram illustrating an example of the structure of model zoo data. FIG. 2 is a diagram illustrating a first example of model zoo data integration. FIG. 3 is a diagram illustrating a second example of model zoo data integration. FIG. 4 is a diagram illustrating a first example of a dataset. FIG. 5 is a diagram illustrating a second example of a dataset. FIG. 6 is a diagram illustrating a sixth example of a dataset. FIG. 7 is a diagram illustrating a seventh example of a dataset. FIG. 8 is a diagram illustrating an example of performance data. FIG. 9 is a diagram illustrating an eighth example of a dataset. FIG. 10 is a flowchart explaining fine-tuning processing. FIG. 11 is a diagram illustrating an example of the configuration of a model selling device. FIG. 12 is a diagram illustrating an example of a purchase screen. A block diagram showing an example of the configuration of computer hardware.

[0014] Hereinafter, modes for carrying out the present technology (hereinafter referred to as embodiments) will be described. The description will be made in the following order: 1. First embodiment (model providing device) 2. Second embodiment (model selling device) 3. Computer

[0015] 1. First Embodiment Configuration Example of Model Providing Device FIG. 1 is a diagram showing a configuration example of a model providing device that is a first embodiment of an information processing system to which the present technology is applied.

[0016] 1 is composed of an assist unit 11, a providing unit 12, an information generating unit 13, and a provided model storage unit 14. The model providing device 10 generates recommendation information regarding the architecture of a neural network such as a DNN based on character input from a model developer, and transmits the neural network corresponding to the recommendation information to the model developer's development device (not shown).

[0017] Specifically, the assist unit 11 is configured by a language model storage unit 21 , an acquisition unit 22 , a processing unit 23 , and a display unit 24 .

[0018] The language model storage unit 21 stores language models such as fine-tuned large language models (LLMs).

[0019] The acquisition unit 22 acquires a prompt related to the architecture of the neural network, entered by text from the user. The prompt includes, for example, information about the device that executes the neural network, the required tasks, and the required performance. The acquisition unit 22 also acquires satisfaction information entered by text from the user, indicating that the user is satisfied with the recommendation information being displayed on the display unit 24. The acquisition unit 22 supplies the prompt and satisfaction information to the processing unit 23 and the display unit 24.

[0020] The processing unit 23 acquires the language model stored in the language model storage unit 21. The processing unit 23 inputs the prompt supplied from the acquisition unit 22 into the language model to obtain recommendation information as output. The recommendation information is architecture information, which is information about the architecture of the recommended available model, and information notifying the performance and execution device of the available model. An available model is a neural network that the model providing device 10 can provide as a development asset, i.e., a neural network that can be selected by a model developer. The architecture information of the available model includes the name of the available model, the base model, the number of layers, a dataset for pre-learning, pre-learning weights, a learning method, etc. The processing unit 23 supplies the recommendation information to the display unit 24. In accordance with the satisfaction information supplied from the acquisition unit 22, the processing unit 23 supplies the providing unit 12 with architecture information corresponding to the satisfaction information.

[0021] The display unit 24 displays an interactive screen showing a dialogue between the user and the AI ​​assistant, using the prompt or satisfaction information supplied from the acquisition unit 22 as the user's speech and the recommendation information supplied from the processing unit 23 as the AI ​​assistant's speech.

[0022] The providing unit 12 reads out, as a provided model, a provideable model corresponding to the architecture information supplied from the processing unit 23 from the provided model storage unit 14. The providing unit 12 transmits the provided model to a model development device (not shown).

[0023] The information generation unit 13 evaluates the trained provided model that is trained by a model development device (not shown) and transmitted to it. Based on the evaluation, the information generation unit 13 generates model zoo data (model-related data) for the trained provided model and supplies it together with the trained provided model to the provided model storage unit 14. The model zoo data is data that describes the outline and specifications of the machine learning model.

[0024] The provision model storage unit 14 stores the provision available models. The provision model storage unit 14 stores the model zoo data supplied from the information generation unit 13, and stores the trained provision models as provision available models.

[0025] <First Example of Interactive Screen> FIG. 2 is a diagram showing a first example of an interactive screen.

[0026] In the example of FIG. 2 , the user inputs "Please suggest an image recognition model" as a prompt for a task requiring image recognition. Therefore, the prompt is first displayed as a user's utterance on the interactive screen 41 of FIG. 2 . Next, recommendation information obtained by a fine-tuned language model using this prompt as input is displayed on the interactive screen 41 as an utterance from the AI ​​assistant. In the example of FIG. 2 , this recommendation information is "The ViT-based XX model has a latency of XX and an accuracy of XX. Do you want to deploy it?", which notifies the name and base model of the recommended available model, as well as the latency and accuracy as the performance of the available model.

[0027] If the user is not satisfied with the latency in this utterance, he or she inputs "Please propose a faster model" as a prompt that requires a lower latency for accuracy. This prompt is then displayed as the user's utterance on the interactive screen 41.

[0028] Next, the AI ​​assistant displays the recommendation information obtained by the fine-tuned language model using this prompt as input. In the example of Figure 2, this recommendation information is "The CNN-based XX model is XX% faster and has XX accuracy," which notifies the name and base model of the newly recommended available model, as well as the latency improvement rate and accuracy of the available model. The latency improvement rate is the improvement rate of the currently recommended available model compared to the latency of the previously recommended available model.

[0029] If the user is satisfied with the latency improvement rate in this utterance but is not satisfied with the recommended available model, the user inputs a prompt to request a change in the recommended available model, with the required accuracy being the same as the latency of the currently recommended available model. In the example of Figure 2, this prompt is "Please also propose another model with the same latency." This prompt is displayed as the user's utterance on the interactive screen 41.

[0030] In this way, the user repeats inputting the prompt until the AI ​​assistant displays recommended information about available models that meet the desired performance.

[0031] <Second Example of Interactive Screen> FIG. 3 is a diagram showing a second example of an interactive screen.

[0032] In the example of FIG. 3 , the user inputs a prompt similar to the first prompt in the example of FIG. 2 , and similar to the interactive screen 41, this prompt is first displayed as the user's speech on the interactive screen 42 of FIG. 3 . Next, recommendation information obtained by a fine-tuned language model using this prompt as input is displayed on the interactive screen 42 as the speech of the AI ​​assistant. In the example of FIG. 3 , this recommendation information is information notifying the name and base model of the recommended available model, two types of execution devices, and the execution latency of each execution device. Specifically, the recommendation information is, "The ViT-based XX model is IMXa with latency XX, and when upgraded to a new edge device, it is XX. Which one would you like to use?"

[0033] If the user wants to run a model available on IMXa but is not satisfied with the latency in this utterance, the user inputs "Please propose a model with latency of XX or less on IMXa" as a prompt to set the required latency during execution by IMXa to a predetermined value or less. This prompt is then displayed on the interactive screen 42 as the user's utterance.

[0034] In this way, the user repeats inputting prompts until the AI ​​assistant displays recommended information about available models that meet the desired performance when executed on the desired execution device.

[0035] <Third Example of Interactive Screen> FIG. 4 is a diagram showing a third example of an interactive screen.

[0036] In the example of Fig. 4, the user inputs a prompt similar to the first prompt in the example of Fig. 2, and this prompt is first displayed as the user's utterance on the interactive screen 43 of Fig. 4, similar to the interactive screen 41. Next, the interactive screen 43 displays recommendation information obtained by a fine-tuned language model using this prompt as input as the AI ​​assistant's utterance.

[0037] In the example of Figure 4, this recommendation information is information that notifies the name of the recommended available model, the base model, two types of number of layers, and the latency and accuracy improvement rates resulting from differences in the number of layers. Specifically, the recommendation information is "The CNN-based XX model has a latency of XX and an accuracy of XX" and "Reducing XX layers will result in a XX decrease in accuracy and an XX improvement in latency. Would you like to reduce them?"

[0038] If the user is satisfied with the latency of the available model with fewer layers recommended in this utterance but is not satisfied with the accuracy, the user inputs a prompt requesting a higher accuracy for the available model with fewer layers. In the example of Figure 4, this prompt is "Yes, I want more accuracy." This prompt is displayed as the user's utterance on the interactive screen 43.

[0039] In this way, the user repeats inputting the prompt until the AI ​​assistant displays recommended information about available models that meet the desired performance.

[0040] <Fourth Example of Interactive Screen> FIG. 5 is a diagram showing a fourth example of an interactive screen.

[0041] In the example of FIG. 5 , the user inputs a prompt similar to the first prompt in the example of FIG. 2 , and similar to the interactive screen 41, this prompt is first displayed as the user's speech on the interactive screen 44 of FIG. 5 . Next, recommendation information obtained by a fine-tuned language model using this prompt as input is displayed on the interactive screen 44 as the speech of the AI ​​assistant. In the example of FIG. 5 , this recommendation information is information notifying the name of the recommended available model, the base model, two types of pre-training datasets, and the rate of improvement in accuracy due to differences in the pre-training datasets. Specifically, the recommendation information is "CNN-based XX model has accuracy of XX" and "Using a model pre-trained with XX dataset will improve accuracy by XX. Would you like to use it?"

[0042] If the user is satisfied with the accuracy of the available model pre-trained with the latter pre-training dataset in this utterance, the user inputs satisfaction information such as "Yes." As a result, this satisfaction information is displayed on the interactive screen 44 as the user's utterance.

[0043] Note that the examples in Figures 2 to 5 are merely examples, and the user prompt may also be "a low-latency image classification model," "a model with mIoU (mean intersection over union) of 80 or more that has been pre-trained using Cityscapes," "a language model of 5 GB or less," etc.

[0044] <Explanation of Recommendation Processing> Fig. 6 is a flowchart illustrating the recommendation processing by the model providing device 10 of Fig. 1. This recommendation processing is started, for example, when the user inputs an instruction to start the recommendation processing.

[0045] 6, the acquisition unit 22 determines whether a prompt regarding the neural network architecture has been input by the user. If it is determined in step S11 that a prompt has been input, the acquisition unit 22 acquires the prompt and supplies it to the processing unit 23 and the display unit 24. The process then proceeds to step S12.

[0046] In step S11, the display unit 24 displays the prompt supplied from the acquisition unit 22 on the dialogue screen as a user's utterance.

[0047] In step S12, the processing unit 23 generates recommendation information as output by inputting the prompt supplied from the acquisition unit 22 into the language model stored in the language model storage unit 21. The processing unit 23 supplies the recommendation information to the display unit 24.

[0048] In step S13, the display unit 24 displays the recommendation information generated by the processing of step S12 on the dialogue screen as an utterance of the AI ​​assistant. At this time, the user determines whether or not they are satisfied with the available model corresponding to this recommendation information. For example, if the available model meets the required performance and required tasks on the desired execution device, the user determines that they are satisfied with the available model. If the user determines that they are satisfied with the available model corresponding to the recommendation information, they input satisfaction information. If they determine that they are not satisfied, they input the prompt again.

[0049] In step S14, the acquisition unit 22 determines whether or not satisfaction information has been input by the user. If it is determined in step S14 that satisfaction information has been input, the acquisition unit 22 supplies the satisfaction information to the processing unit 23 and the display unit 24.

[0050] In step S15, the display unit 24 displays the satisfaction information supplied from the acquisition unit 22 on the dialogue screen as a user's utterance. In step S16, the processing unit 23 supplies the providing unit 12 with architecture information corresponding to the satisfaction information supplied from the acquisition unit 22.

[0051] In step S17, the providing unit 12 reads out the available model corresponding to the architecture information supplied in step S16 from the provided model storage unit 14 as a provided model, and transmits it to the model development device, whereupon the recommendation process ends.

[0052] On the other hand, if it is determined in step S14 that the satisfaction information has not been input, the process returns to step S10, and the subsequent processes are repeated.

[0053] <Configuration Example of Model Generation Device> FIG. 7 is a block diagram showing a configuration example of a model generation device that generates a language model to be stored in the language model storage unit 21 of FIG.

[0054] The model generating device 50 in FIG. 7 is made up of a collection unit 51, a DB (Data Base) 52, a data set construction unit 53, and a fine tuning unit 54.

[0055] The collection unit 51 collects model zoo data of available models as collected data from model zoo data on websites such as OpenMMLab and Hugging Face. The collection unit 51 collects performance data of available models as collected data from performance data indicating the performance of multiple machine learning models that differ in at least one piece of architecture information, which is described on web pages such as ablation studies in academic papers and internal technical reports. The collection unit 51 also collects model zoo data stored in the provided model storage unit 14 as collected data.

[0056] The collection unit 51 updates the DB 52 by supplying the collected data to the DB 52. At this time, the collection unit 51 integrates, as necessary, model zoo data collected from a certain collection source with model zoo data collected from other collection sources that are registered in the DB 52. The DB 52 registers the collected data supplied from the collection unit 51.

[0057] The data set construction unit 53 constructs a data set for fine tuning based on the collected data registered in the DB 52 and supplies it to the fine tuning unit 54 .

[0058] The fine tuning unit 54 fine-tunes (instruction tunes) a language model such as an LLM using the dataset supplied from the dataset construction unit 53, and generates a fine-tuned language model. This fine-tuned language model is stored in the language model storage unit 21 and functions as a generator of recommendation information.

[0059] The model generation device 50 configured as described above continues to collect collected data at predetermined times after generating a fine-tuned language model, and updates the fine-tuned language model, thereby updating the fine-tuned language model stored in the language model storage unit 21.

[0060] <Example of Model Zoo Data Structure> FIG. 8 is a diagram showing an example of the structure of model zoo data.

[0061] 8 is composed of a machine learning model overview 71, usage instructions 72, and specification table information 73. The overview 71 is metadata such as a description of the machine learning model. The usage instructions 72 is information such as programs required to use the machine learning model.

[0062] The specification table information 73 includes a specification table that expresses specifications such as architecture information and accuracy of the machine learning model. This specification table is generated, for example, for each pre-training dataset, and the name of the dataset ("Cityscapes" in the example of FIG. 8) is added as a title. In the example of FIG. 8, the specification table has the following items: "Method," "Backbone," "Crop Size," "Lr schd," "Mem (GB)," "Inf time (fps)," "Device," "mIoU," "mIoU (ms+flip)," "config," and "download."

[0063] The column corresponding to the item "Method" contains the name of the machine learning model. The column corresponding to the item "Backbone" contains information representing the base model of the machine learning model. The column corresponding to the item "Crop Size" contains the size of the image data used in the machine learning model. The column corresponding to the item "Lr schd" contains information about the learning rate of the machine learning model. The column corresponding to the item "Mem (GB)" contains the memory capacity required to run the machine learning model.

[0064] The column corresponding to the item "Inf time (fps)" contains the inference time by the machine learning model. The column corresponding to the item "Device" contains information indicating the execution device. The columns corresponding to the items "mloU" and "mIoU (ms+flip)" contain the mIoU as the accuracy of the machine learning model. The column corresponding to "config" contains a link to the configuration file of the machine learning model. The column corresponding to "download" contains a link to a web page where the machine learning model can be downloaded or a web page with the machine learning model log.

[0065] <First Example of Integration of Model Zoo Data> FIG. 9 is a diagram showing a first example of integration of model zoo data by the collection unit 51 of FIG.

[0066] 9, the degree of similarity (similarity) between the model zoo data including the specification table of A in Fig. 9 registered in DB 52 and the model zoo data including the specification table of B in Fig. 9 collected by collection unit 51 is higher than a threshold. In this case, collection unit 51 infers that the specification table of A in Fig. 9 and the specification table of B in Fig. 9 are specification tables of the same type of available model and integrates them, thereby expanding the specification table of A in Fig. 9 and generating the specification table of C in Fig. 9.

[0067] Specifically, the specification table in A of Fig. 9 has items "Backbone," "Mem (GB)," and "mIoU," and the specifications of two available models corresponding to each item are written in two columns. The table in B of Fig. 9 has items "Backbone," "Latency (ms)," and "mIoU," and the specifications of two available models corresponding to each item are written in two columns.

[0068] Therefore, the collection unit 51 adds the item "Latency (ms)" of the specification table of B of FIG. 9 that does not overlap with the items in the specification table of A of FIG. 9 to the items "Backbone," "Mem (GB)," and "mIoU" in the specification table of A of FIG. 9. Then, the collection unit 51 adds two columns and enters the specifications of the two available models that are entered in the specification table of B of FIG. 9 in the columns of the items "Backbone," "Latency (ms)," and "mIoU" in each of the added rows. As a result, the specification table of A of FIG. 9 is expanded into the specification table of C of FIG. 9 in which the specifications of the four available models that correspond to the items "Backbone," "Mem (GB)," "mIoU," and "Latency (ms)" are entered in the four columns, respectively.

[0069] As described above, the collection unit 51 integrates specification tables of model zoo data with high similarity. This prevents specification tables from being collected from different sources and being separately registered in the DB 52, even though they have high similarity. As a result, the dataset construction unit 53 can use the integrated tables to generate a highly accurate dataset for fine tuning.

[0070] <Second Example of Integration of Model Zoo Data> FIG. 10 is a diagram showing a second example of integration of model zoo data by the collection unit 51 of FIG.

[0071] 10, similarly to the example of Fig. 9, the model zoo data including the specification table of Fig. 9A and the model zoo data including the specification table of Fig. 9B, which are registered in DB 52, are integrated. At this time, in the example of Fig. 10, the collection unit 51 calculates a reliability score that indicates the reliability of the specification tables of Fig. 9A and Fig. 9B. This reliability score is determined based on the similarity of the corresponding model zoo data with the model zoo data collected from the provided model storage unit 14, the reliability of the repository from which the data was collected, the rank of the academic society corresponding to the source of the data, etc.

[0072] In the example of Fig. 10, the reliability score of the specification table A in Fig. 9 is 1, and the reliability score of the specification table B in Fig. 9 is 0.5. In this case, the collection unit 51 adds an item "Score" to the specification table C in Fig. 9, which is obtained by integrating the specification table A in Fig. 9 and the specification table B in Fig. 9. Then, the collection unit 51 enters the reliability scores of the specification tables before integration corresponding to the four columns of the item "Score." As a result, the specification table shown in Fig. 10 is generated.

[0073] As described above, the collection unit 51 associates the specifications of each available model in the integrated specification table with the reliability score of the specification table before the integration in which the specification was described. Therefore, the dataset construction unit 53 can make adjustments such that, for example, information described in the integrated specification table with a higher corresponding reliability score is used to generate a dataset for fine tuning. As a result, the dataset construction unit 53 can generate a highly accurate dataset for fine tuning.

[0074] <First Example of Data Set> FIG. 11 is a diagram showing a first example of a dataset generated by the dataset construction unit 53 of FIG.

[0075] In the example of FIG. 11 , the dataset construction unit 53 uses the model zoo data of FIG. 8 to generate the instructions, inputs, and responses shown in FIG. 11 as a dataset. The instructions in FIG. 11 are a task for generating columns from metadata. The input in FIG. 11 is the outline 71 in FIG. 8. The response in FIG. 11 is, for example, part of the data in the third row from the top of the specification table included in the specification table information 73: "{"Method": "Mask2Former", "Backbone": "Swin-T", "Model size": 6.36 GB, "mIoU": 81.71, "config": config...}".

[0076] <Second Example of Data Set> FIG. 12 is a diagram showing a second example of the data set generated by the data set construction unit 53 of FIG.

[0077] In the example of Fig. 12, the dataset construction unit 53 uses the model zoo data of Fig. 8 to generate the instructions, inputs, and responses shown in Fig. 12 as a dataset. The instructions in Fig. 12 are tasks for generating metadata from tables. The inputs in Fig. 12 are specification tables included in specification table information 73, and the responses are overviews 71.

[0078] As described above, the datasets in Figures 11 and 12 are not simply for learning the output of the specifications of a deliverable model, but are for learning by associating the specifications and overview of the deliverable model. Similar to the Table-GPT dataset, the datasets in Figures 11 and 12 include instructions focused on tables. Table-GPT is described, for example, in Li, Peng, et al., "Table-gpt: Table-tuned gpt for diverse table tasks," arXiv preprint arXiv:2310.09263, 2023.

[0079] <Third Example of Data Set> FIG. 13 is a diagram showing a third example of a dataset generated by the dataset construction unit 53 of FIG.

[0080] In the example of Fig. 13, the dataset construction unit 53 generates the instructions, inputs, and responses shown in A of Fig. 13 as a dataset using the model zoo data of Fig. 8. The instruction in A of Fig. 13 is a task of predicting and filling masked data, which is masked data in a column in the specification table information 73.

[0081] The input of A in Fig. 13 is a column in which data contained in specification table information 73 is masked, and an overview 71. Specifically, this column is, for example, a column in which data is masked, corresponding to the title of the specification table of Fig. 8 and the items "Backbone," "mIoU," and "config" in the first line from the top of the specification table, as shown in B in Fig. 13. In this case, the response of A in Fig. 13 is data corresponding to the title of the specification table of Fig. 8 and the items "Backbone," "mIoU," and "config" in the first line from the top of the specification table.

[0082] 13 is a dataset for learning to fill in some columns of the specification table information 73. As a result, a language model fine-tuned using this dataset can understand the specification table information 73.

[0083] <Fourth Example of Data Set> FIG. 14 is a diagram showing a fourth example of a dataset generated by the dataset construction unit 53 of FIG.

[0084] In the example of Fig. 14 , the dataset construction unit 53 generates the instructions, inputs, and responses shown in A of Fig. 14 as a dataset using the model zoo data of Fig. 8 . The instruction in A of Fig. 14 is a task to predict mIoU as the specifications of the machine learning model specified by the input. The input in A of Fig. 14 is a column that specifies a available model corresponding to the mIoU of the prediction target by masking the mIoU of the prediction target in the specification table information 73, or a statement that specifies a available model corresponding to the mIoU of the prediction target.

[0085] Specifically, the column generated as input is, for example, a column in which data corresponding to the item "mIoU" in the first row from the top of the specification table of the specification table information 73 shown in B of FIG. 14 is masked. The sentence generated as input is, for example, "Mask2Former with Backbone R-XX-DXX trained on Cityscapes," which specifies the available model corresponding to the first row from the top of the specification table of the specification table information 73. In this case, the response in A of FIG. 14 is 80.44, which is written in the column corresponding to the item "mIoU" in the first row from the top of the specification table of the specification table information 73.

[0086] 14 is a dataset for learning the specifications of available models registered in the specification table, so that a language model fine-tuned using this dataset can understand the specification table.

[0087] <Fifth Example of Data Set> FIG. 15 is a diagram showing a fifth example of a dataset generated by the dataset construction unit 53 of FIG.

[0088] In the example of FIG. 15 , the dataset construction unit 53 uses the model zoo data of FIG. 8 to generate the instructions, inputs, and responses shown in A of FIG. 15 as a dataset. The instruction in A of FIG. 15 is a task to add a new column to the input table. The input in A of FIG. 15 is part of the specification table information 73. Specifically, the input is, for example, a table consisting of only the first row from the top of the specification table in the specification table information 73, as shown in B of FIG. 15 . In this case, the response in A of FIG. 15 is, for example, all columns in the third row from the top of the specification table in the specification table information 73.

[0089] 15 is a dataset for learning to add columns based on a specification table, so that a language model fine-tuned using this dataset can output recommendation information based on the newly collected specification table.

[0090] 11 to 15 are generated with a focus on the structure of the model zoo data, so the fine-tuning unit 54 can use this data set to perform fine-tuning suited to the structure of the model zoo data.

[0091] <Sixth Example of Data Set> FIG. 16 is a diagram showing a sixth example of a dataset generated by the dataset construction unit 53 of FIG.

[0092] In the example of Fig. 16, the dataset construction unit 53 uses model zoo data to generate the instructions, inputs, and responses shown in Fig. 16 as a dataset. The instructions in Fig. 16 are a task related to output of recommendation information, and are a task to recommend architecture information of Mask2Former that uses only the ReLU (Rectified Linear Unit) function.

[0093] The input in Fig. 16 is a specification table. The response in Fig. 16 is, for example, data in the columns of items "Backbone" and "config" as architecture information of Mask2Former that uses only the ReLU function, and data in the column of item "mIoU" as accuracy, which are registered in the specification table.

[0094] <Seventh Example of Data Set> FIG. 17 is a diagram showing a seventh example of a dataset generated by the dataset construction unit 53 of FIG.

[0095] In the example of FIG. 17 , the dataset construction unit 53 uses model zoo data to generate the instructions, inputs, and responses shown in FIG. 17 as a dataset. The instructions in FIG. 17 are tasks related to outputting recommendation information, and are tasks for predicting the accuracy of Mask2Former using only the ReLU function. The input in FIG. 17 is a specification table. The response in FIG. 17 is, for example, data in the column of the item "mIoU" registered in the specification table, which represents the accuracy of Mask2Former using only the ReLU function.

[0096] As described above, the instructions in the datasets of FIGS. 16 and 17 are instructions related to accuracy prediction. Specifically, in the specification tables of model zoo data, numerical data such as accuracy and latency are more important than general table data such as Table-GPT. Therefore, the dataset construction unit 53 generates the datasets of FIGS. 16 and 17 that include instructions related to accuracy prediction, which is one type of numerical data. This allows the fine-tuning unit 54 to use this dataset to perform fine-tuning appropriate for the specification table.

[0097] The instructions in the datasets of Figures 16 and 17 also limit the functions used by the available model corresponding to the recommendation information. Specifically, if the execution device is an edge device or the like, the functions that the execution device can support will differ depending on the hardware that the execution device has. Therefore, the dataset construction unit 53 generates the datasets of Figures 16 and 17 assuming that the functions used by the neural network in prompts of the fine-tuned language model will be limited. This allows the fine-tuning unit 54 to use this dataset to perform fine-tuning appropriate for the execution device.

[0098] If the specification table does not contain a specification for a supplyable model that uses only functions defined by instructions, the collection unit 51 generates a specification for that supplyable model. Specifically, the collection unit 51 replaces the functions used by the supplyable model registered in the specification table with functions defined by instructions, calculates accuracy, and generates a specification that includes that accuracy. The dataset construction unit 53 then uses the specification table for that specification as input.

[0099] <Example of Performance Data> FIG. 18 is a diagram showing an example of performance data.

[0100] 18 shows the performance data in terms of accuracy and GFLOPs (Giga Floating-point Operations Per Second), which is an index of computational cost, for each of multiple machine learning models that differ in at least one piece of architecture information. Specifically, this accuracy is the accuracy (In-1k acc.) when ImageNet-1K is used as the training dataset.

[0101] According to the performance data in Figure 18, the accuracy of ResNet (Residual Network) is 66.13 and GELOPs is 3.09. When depthwise convolution is used in ResNet, the accuracy is 68.28 ± 0.08 and GELOPs is 1.35. On the other hand, when the function used in ResNet is changed from the ReLU function to the GELU (Gaussian Error Linear Units) function, the accuracy is 70.62 ± 0.14 and GELOPs is 3.15.

[0102] <Eighth Example of Data Set> FIG. 19 is a diagram showing an eighth example of a dataset generated by the dataset construction unit 53 of FIG.

[0103] In the example of Fig. 19, the dataset construction unit 53 generates the instructions, inputs, and responses shown in Fig. 19 as a dataset using the performance data of Fig. 18. The instructions in Fig. 19 are a task for proposing a method for improving the processing speed of a supplyable model specified by the input and predicting the processing speed of the supplyable model when that method is used. The input in Fig. 19 is information for specifying a Res-Net corresponding to the performance data of Fig. 18.

[0104] The response in Fig. 19 generates a method for modifying the architecture information as a method for improving the processing speed of the Res-Net specified in the input, and the GFLOPs corresponding to the Res-Net processing speed when that method is used, based on the performance data in Fig. 18. Specifically, the method for modifying the architecture information employs depthwise convolution, and the GFLOPs is 1.35.

[0105] The instructions for the dataset may be a task to propose a method for improving other performance such as accuracy and predict the performance when that method is used.

[0106] 19 is a dataset for learning how to propose architecture information modification methods that improve the performance of a producible model based on performance data. As a result, a language model fine-tuned using this dataset can propose architecture information modification methods that improve the performance of a producible model.

[0107] <Ninth Example of Data Set> FIG. 20 is a diagram showing a ninth example of a dataset generated by the dataset construction unit 53 of FIG.

[0108] In the example of Fig. 20, the dataset construction unit 53 generates the instructions, inputs, and responses shown in Fig. 20 as a dataset using the performance data of Fig. 18. The instructions in Fig. 20 are a task for predicting a change in performance when a GELU function is used instead of a ReLU function in a provisionable model specified by the input. The input in Fig. 20 is information specifying a Res-Net corresponding to the performance data of Fig. 18.

[0109] As the response in Fig. 20, information representing changes in accuracy and GFLOPs as performance is generated based on the performance data in Fig. 18. Specifically, information representing an improvement from 66.13 to 70.62 is generated as information representing a change in accuracy, and information representing an increase from 3.09 to 3.15 is generated as information representing a change in GFLOPs.

[0110] The instructions in the dataset may be a task to predict changes in performance that accompany changes in other architectural information, such as whether or not depthwise convolution is employed.

[0111] As described above, the dataset in Figure 20 is a dataset for learning to predict performance changes associated with changes in the architecture information of a production model based on performance data. As a result, a language model fine-tuned using this dataset can predict performance changes associated with changes in the architecture information of a production model.

[0112] If the performance data does not include data representing the performance of a provisionable model that uses functions compatible with the target device, which is a device assumed to be the execution device, the collection unit 51 generates such data. Specifically, the collection unit 51 replaces the functions used by the provisionable model corresponding to the performance data with functions compatible with the target device, investigates the performance, and generates data representing the performance. The dataset construction unit 53 then generates a response based on the performance data including that data.

[0113] <Explanation of Fine Tuning Process> FIG. 21 is a flowchart illustrating the fine tuning process performed by the model generating device 50 of FIG.

[0114] In step S31 of Fig. 21, the collection unit 51 collects, as collected data, model zoo data of available models on the website, performance data of the available models, and model zoo data stored in the provided model storage unit 14 of Fig. 1. In step S32, the collection unit 51 supplies the collected data collected in step S31 to the DB 52, thereby updating the DB 52.

[0115] In step S33, the collection unit 51 determines whether or not the DB 52 updated in step S32 includes collected data corresponding to a function that can be supported by the target device.

[0116] If it is determined in step S33 that collected data corresponding to functions compatible with the target device is not included, the process proceeds to step S34. In step S34, the dataset construction unit 53 generates collected data corresponding to functions compatible with the target device and supplies it to the DB 52, thereby updating the DB 52. The process then proceeds to step S35.

[0117] On the other hand, if it is determined in step S33 that collected data corresponding to a function that can be supported by the target device is included, the process proceeds to step S35.

[0118] In step S35, the data set construction unit 53 constructs a data set for fine tuning based on the collected data registered in the DB 52, and supplies the data set to the fine tuning unit .

[0119] In step S36, the fine-tuning unit 54 fine-tunes the language model using the dataset constructed in step S35. Specifically, the fine-tuning unit 54 uses the input of the dataset as input, the response as output, and solves the task as the instructions of the dataset to generate a fine-tuned language model. Then, the fine-tuning process ends.

[0120] After the fine-tuning process is completed, steps S31 to S35 are performed at a predetermined timing, and the fine-tuned learning model is fine-tuned again and updated using the fine-tuning dataset obtained as a result. This updating is performed continuously.

[0121] When an available model is added, the model generation device 50 performs continual learning. This allows the model generation device 50 to re-learn the fine-tuned language model with minimal additional learning and without degrading performance, without having to re-learn the fine-tuned language model. When an available model is deleted, the model generation device 50 performs machine unlearning. This allows the model generation device 50 to additionally learn the fine-tuned language model so as to forget and not recommend the deleted available model.

[0122] As described above, in the model providing device 10, the acquisition unit 22 acquires a prompt related to the neural network architecture. The processing unit 23 then inputs the prompt into the fine-tuned language model, thereby outputting recommendation information. Therefore, the user can instantly receive a proposal for optimal architecture information simply by inputting a prompt related to the neural network architecture.

[0123] The model generation device 50 generates a fine-tuned language model based on a database of model zoo data and performance data for available models. Therefore, a language model suitable for recommending architecture information for available models can be generated. As a result, a user can obtain recommendation information corresponding to various requirements, such as the neural network execution device, required tasks, and required performance, by inputting these requirements as prompts. Therefore, processing costs can be reduced compared to performing NAS until the desired performance is achieved.

[0124] Because model zoo data and performance data are obtained from the website, highly accurate fine-tuning can be performed based on the model zoo data and performance data of all available models published on the website. As a result, recommendation information for available models desired by the user can be output in a short time. In other words, the available models can be optimized in a short time.

[0125] Since the model zoo data is also acquired from the provided model storage unit 14, fine tuning can be performed based on the model zoo data of the trained provided model as well. This makes it possible to perform fine tuning based on the model zoo data of a neural network that uses an edge device or the like as an execution device, which is not widely disclosed on websites, for example.

[0126] Collected data collected from multiple collection sources is registered in DB 52, and the dataset construction unit 53 constructs a dataset for fine tuning based on the collected data. Therefore, a fine-tuning dataset containing instructions of the same format is constructed even for collected data collected from different collection sources. This allows the fine-tuned language model to collectively handle available models for which collected data exists from multiple collection sources. As a result, the flexibility of recommended available models is high.

[0127] 2. Second Embodiment Configuration Example of Model Vending Device FIG. 22 is a diagram showing a configuration example of a model selling device that is a second embodiment of an information processing system to which the present technology is applied.

[0128] In the model selling device 110 in Fig. 22, parts corresponding to those in the model providing device 10 in Fig. 1 are given the same reference numerals. Therefore, the description of those parts will be omitted as appropriate, and the following description will focus on parts that differ from the model providing device 10. The model selling device 110 generates recommendation information based on character input from a model purchaser, and sells neural networks and the like corresponding to that recommendation information to the model purchaser.

[0129] Specifically, the model selling device 110 comprises an assisting unit 111 and a billing processing unit 112. The assisting unit 111 differs from the assisting unit 11 in that the model developer acts as the model purchaser, and in that the assisting unit 111 comprises a language model storage unit 121, a processing unit 123, and a display unit 124 instead of the language model storage unit 21, the processing unit 23, and the display unit 24. The rest of the configuration is the same as that of the assisting unit 11.

[0130] The language model storage unit 121 stores language models such as fine-tuned LLMs. The method for generating these fine-tuned language models is basically the same as in the first embodiment. However, the collected data is collected data corresponding to sales models, which are neural networks that can be sold by the model selling device 110, sales devices, which are execution devices that can be sold, and sales datasets, which are datasets for additional learning that can be sold. The recommendation information is information that notifies the architecture information of the recommended sales model, the sales device, the sales dataset, etc. The architecture information of the sales model includes the name, base model, number of layers, datasets for pre-learning and pre-learning weights, learning method, etc.

[0131] The processing unit 123 obtains recommendation information as an output by inputting the prompt supplied from the acquisition unit 22 into the language model stored in the language model storage unit 121. The processing unit 123 supplies the recommendation information to the display unit 124. In response to the satisfaction information supplied from the acquisition unit 22, the processing unit 123 supplies the recommendation information corresponding to the satisfaction information to the billing processing unit 112.

[0132] Similar to the display unit 24, the display unit 124 displays an interactive screen based on the prompt or satisfaction information supplied from the acquisition unit 22 and the recommendation information supplied from the processing unit 23. The display unit 124 also displays a purchase screen for purchasing a sales model, a sales device, and a sales data set corresponding to the satisfaction information, based on the satisfaction information.

[0133] The billing processing unit 112 acquires input from the model purchaser on the purchase screen displayed on the display unit 124. Based on the input, the billing processing unit 112 performs billing processing on the model purchaser for at least one of the sales model, sales device, and sales data set corresponding to the recommendation information supplied from the processing unit 23. The billing processing unit 112 performs processing to provide the model purchaser with at least one of the sales model, sales device, and sales data set for which billing processing has been completed.

[0134] Since the model selling device 110 sells sales models corresponding to the recommendation information, the reliability of the specifications of the sales models is important. Therefore, it is desirable to generate a fine-tuned language model based on a DB in which reliability scores are registered when integrating model zoo data, as described in FIG. 10 , for example.

[0135] <Example of Purchase Screen> FIG. 23 is a diagram showing an example of a purchase screen displayed on the display unit 124 of FIG.

[0136] In the example of FIG. 23 , the user inputs the requested task "object detection," the data used "own factory data," the requested accuracy "xx or higher," and the requested memory capacity "xx or less" as a prompt. In this case, the processing unit 123 inputs this prompt into the fine-tuned language model to obtain recommendation information. In the example of FIG. 23 , this recommendation information notifies IMXa as the recommended sales device, notifies CNN as the base model of the recommended sales model, and notifies the object detection dataset as the recommended sales dataset.

[0137] When the user inputs satisfaction information indicating that they are satisfied with the recommendation information, a purchase screen 130 shown in Fig. 23 is displayed on the display unit 124. A message urging the user to purchase the sales device, sales model, and sales data set corresponding to the satisfaction information is displayed at the top of the purchase screen 130. A device purchase button 131, a model purchase button 132, and a data set purchase button 133 are provided at the bottom of the purchase screen 130.

[0138] The device purchase button 131 is operated by the model purchaser when purchasing a sales device corresponding to the satisfaction information. The model purchase button 132 is operated by the model purchaser when purchasing a sales model corresponding to the satisfaction information. The data set purchase button 133 is operated by the model purchaser when purchasing a sales data set corresponding to the satisfaction information.

[0139] As described above, in the model selling device 110, the acquisition unit 22 acquires a prompt related to the neural network architecture. The processing unit 123 then inputs the prompt into the fine-tuned language model, thereby outputting recommendation information. Therefore, the user can instantly receive a proposal for optimal architecture information simply by inputting a prompt related to the neural network architecture. As a result, the user can easily purchase the desired selling model.

[0140] In the first and second embodiments, the conversation between the user and the AI ​​assistant may be conducted by voice.

[0141] 3. Computer The above-described series of processes can be executed by hardware or software. When the series of processes is executed by software, the programs that make up the software are installed on a computer. Here, the term "computer" includes computers built into dedicated hardware, and general-purpose personal computers, for example, that can execute various functions by installing various programs.

[0142] FIG. 24 is a block diagram showing an example of the hardware configuration of a computer that executes the above-described series of processes by a program.

[0143] In the computer, a CPU (Central Processing Unit) 901, a ROM (Read Only Memory) 902, and a RAM (Random Access Memory) 903 are interconnected by a bus 904.

[0144] An input / output interface 905 is also connected to the bus 904. An input unit 906, an output unit 907, a storage unit 908, a communication unit 909, and a drive 910 are connected to the input / output interface 905.

[0145] The input unit 906 includes a keyboard, a mouse, a microphone, etc. The output unit 907 includes a display, a speaker, etc. The output unit 907 corresponds to, for example, the display unit 24 in FIG. 1. The storage unit 908 includes a hard disk, a non-volatile memory, etc. The storage unit 908 corresponds to, for example, the language model storage unit 21 in FIG. 1, the provided model storage unit 14, the DB 52 in FIG. 7, etc. The communication unit 909 includes a network interface, etc. The drive 910 drives removable media 911 such as a magnetic disk, an optical disk, a magneto-optical disk, or a semiconductor memory.

[0146] In the computer configured as above, the CPU 901 performs the above-described series of processes by, for example, loading a program stored in the storage unit 908 into the RAM 903 via the input / output interface 905 and bus 904 and executing the program. That is, the CPU 901 performs the processes of the acquisition unit 22, processing unit 23, provision unit 12, and information generation unit 13, and also performs the processes of the collection unit 51, dataset construction unit 53, and fine-tuning unit 54 in FIG. 7 .

[0147] The program executed by the computer (CPU 901) can be provided by being recorded on a removable medium 911 such as a package medium, for example. The program can also be provided via a wired or wireless transmission medium such as a local area network, the Internet, or digital satellite broadcasting.

[0148] In a computer, the program can be installed in the storage unit 908 via the input / output interface 905 by inserting the removable medium 911 into the drive 910. The program can also be received by the communication unit 909 via a wired or wireless transmission medium and installed in the storage unit 908. Alternatively, the program can be installed in the ROM 902 or the storage unit 908 in advance.

[0149] The program executed by the computer may be a program that processes in chronological order according to the order described in this specification, or may be a program that processes in parallel or at the required timing, such as when called.

[0150] In this specification, a system refers to a collection of multiple components (devices, modules (components), etc.), regardless of whether all of the components are contained in the same housing. Therefore, multiple devices housed in separate housings and connected via a network, and a single device housed in a single housing with multiple modules, are both systems.

[0151] The embodiments of the present technology are not limited to the above-described embodiments, and various modifications are possible without departing from the spirit of the present technology.

[0152] For example, it is possible to adopt a configuration in which all or part of the above-described embodiments are combined.

[0153] For example, the present technology can be configured as a cloud computing system in which a single function is shared and processed collaboratively by a plurality of devices via a network.

[0154] Furthermore, each step described in the above flowchart can be executed by one device, or can be shared and executed by a plurality of devices.

[0155] Furthermore, when one step includes multiple processes, the multiple processes included in that one step can be executed by one device or can be shared and executed by multiple devices.

[0156] The effects described in this specification are merely examples and are not limiting, and other effects may also be present.

[0157] The present technology may have the following configurations. (1) An information processing system comprising: an acquisition unit that acquires a prompt related to a neural network architecture; and a processing unit that outputs recommendation information related to the neural network architecture by inputting the prompt into a fine-tuned language model. (2) The information processing system described in (1), wherein the fine-tuned language model is a model fine-tuned based on a database of model-related data that describes an overview and specifications of a machine learning model. (3) The information processing system described in (2), wherein the fine-tuned language model is a model fine-tuned by solving a task of predicting mask data in the model-related data. (4) The information processing system described in (2) or (3), wherein the fine-tuned language model is a model fine-tuned by solving a task of predicting the specification in the model-related data. (5) The information processing system described in any of (2) to (4), wherein the fine-tuned language model is a model fine-tuned by solving a task of adding the specification in the model-related data. (6) The information processing system according to any one of (2) to (5), wherein the fine-tuned language model is a model fine-tuned by solving a task of recommending architecture information regarding the architecture of the machine learning model using only predetermined functions in the model-related data. (7) The information processing system according to any one of (2) to (6), wherein the fine-tuned language model is a model fine-tuned by solving a task of predicting the accuracy of the machine learning model using only predetermined functions in the model-related data. (8) The information processing system according to any one of (2) to (7), wherein the specifications are expressed using a table. (9) The information processing system according to any one of (2) to (8), wherein the database is generated by integrating the model-related data collected from a plurality of collection sources.(10) The information processing system according to (9), wherein the database is generated by integrating the model-related data whose similarity is higher than a threshold. (11) The information processing system according to (10), wherein the database is generated by integrating the model-related data whose similarity is higher than the threshold based on the respective reliability of the model-related data whose similarity is higher than the threshold. (12) The information processing system according to (1), wherein the fine-tuned language model is a model fine-tuned based on a database of performance data representing the performance of multiple machine learning models having different architecture information. (13) The information processing system according to (12), wherein the fine-tuned language model is a model fine-tuned by solving a task that suggests a method for improving a predetermined performance of the machine learning model based on the performance data. (14) The information processing system according to (12) or (13), wherein the fine-tuned language model is a model fine-tuned by solving a task that predicts a change in performance due to a change in the architecture information of the machine learning model based on the performance data. (15) The information processing system according to (1), further comprising: a providing unit that provides the neural network corresponding to the recommendation information output by the processing unit. (16) The information processing system according to (15), further comprising: a provided model acquisition unit that acquires the neural network corresponding to the recommendation information output by the processing unit, wherein the providing unit provides the neural network acquired by the provided model acquisition unit. (17) The information processing system according to (1), further comprising: an information generation unit that generates model-related data provided by the providing unit and describing an outline and specifications of the trained neural network, wherein the fine-tuned language model is a model fine-tuned based on the database of the model-related data generated by the information generation unit.(18) The information processing system according to (1), further comprising: a language model acquisition unit that acquires the fine-tuned language model. (19) An information processing method, comprising: an information processing system acquiring a prompt related to a neural network architecture, and outputting recommendation information related to the neural network architecture by inputting the prompt into a fine-tuned language model. (20) A program for causing a computer to function as an information processing system comprising: an acquisition unit that acquires a prompt related to a neural network architecture, and a processing unit that inputs the prompt into a fine-tuned language model to output recommendation information related to the neural network architecture.

[0158] 10 Model providing device, 22 Acquisition unit, 23 Processing unit, 110 Model selling device, 123 Processing unit

Claims

1. An information processing system comprising: an acquisition unit that acquires prompts related to neural network architecture; and a processing unit that outputs recommendation information related to neural network architecture by inputting the prompts into a fine-tuned language model.

2. The information processing system according to claim 1, wherein the fine-tuned language model is a model fine-tuned based on a database of model-related data that describes an overview and specifications of a machine learning model.

3. The information processing system according to claim 2, wherein the fine-tuned language model is a model that is fine-tuned by solving a task of predicting mask data in the model-related data.

4. The information processing system according to claim 2, wherein the fine-tuned language model is a model that is fine-tuned by solving a task of predicting the specifications in the model-related data.

5. The information processing system according to claim 2, wherein the fine-tuned language model is a model that has been fine-tuned by solving a task of adding the specification in the model-related data.

6. The information processing system of claim 2, wherein the fine-tuned language model is a model that has been fine-tuned by solving a task of recommending architectural information regarding the architecture of the machine learning model that uses only predetermined functions in the model-related data.

7. The information processing system of claim 2, wherein the fine-tuned language model is a model fine-tuned by solving a task of predicting the accuracy of the machine learning model using only predetermined functions in the model-related data.

8. The information processing system according to claim 2, wherein the specifications are expressed using a table.

9. The information processing system according to claim 2, wherein the database is generated by integrating the model-related data collected from a plurality of sources.

10. The information processing system according to claim 9, wherein the database is generated by integrating the model-related data whose similarity is higher than a threshold value.

11. The information processing system according to claim 10, wherein the database is generated by integrating the model-related data based on the reliability of each of the model-related data whose similarity is higher than the threshold value.

12. The information processing system according to claim 1, wherein the fine-tuned language model is a model fine-tuned based on a database of performance data representing the performance of multiple machine learning models with different architecture information.

13. The information processing system of claim 12, wherein the fine-tuned language model is a model fine-tuned by solving a task that suggests a method for improving a predetermined performance of the machine learning model based on the performance data.

14. The information processing system of claim 12, wherein the fine-tuned language model is a model that has been fine-tuned by solving a task of predicting the change in performance that accompanies a change in the architectural information of the machine learning model based on the performance data.

15. The information processing system according to claim 1, further comprising: a providing unit that provides the neural network corresponding to the recommendation information output by the processing unit.

16. The information processing system according to claim 15, further comprising a provision model acquisition unit that acquires the neural network corresponding to the recommendation information output by the processing unit, wherein the provision unit provides the neural network acquired by the provision model acquisition unit.

17. The information processing system of claim 1, further comprising an information generation unit that generates model-related data provided by the provision unit and describing an overview and specifications of the trained neural network, wherein the fine-tuned language model is a model that has been fine-tuned based on the database of model-related data generated by the information generation unit.

18. The information processing system according to claim 1, further comprising a language model acquisition unit that acquires the fine-tuned language model.

19. An information processing method comprising: an information processing system; obtaining prompts related to neural network architectures; and outputting recommendation information related to neural network architectures by inputting the prompts into a fine-tuned language model.

20. A program for causing a computer to function as an information processing system comprising: an acquisition unit that acquires prompts related to neural network architecture; and a processing unit that outputs recommendation information related to neural network architecture by inputting the prompts into a fine-tuned language model.

Images