Model creation apparatus, model creation method, and model creation system
By using a model creation device to fine-tune or transfer learning from a baseline model, the problem of inappropriate model selection after learning is solved, and more accurate model creation is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-09-29
- Publication Date
- 2026-04-03
AI Technical Summary
In existing technologies, the selected learning model may be inappropriate, resulting in inaccurate results.
By creating a model creation device, a target model suitable for the target environment is determined, and the target model is created by fine-tuning or transfer learning using a benchmark model.
Creating a suitable model more easily can provide more accurate results.
Smart Images
Figure CN115936138B_ABST
Abstract
Description
Technical Field
[0001] This disclosure relates to a model creation apparatus, a model creation method, and a model creation system. Background Technology
[0002] A learning completion model providing system is known to select one or more learning completion models from a plurality of learning completion models stored in a database that match the user's purpose of use, based on usage requests obtained from a user-side device (see, for example, Patent Document 1, paragraph 0037). Patent Document 1 also discloses a technique for fine-tuning the selected learning completion model before providing it to the user-side device.
[0003] Prior art literature
[0004] Patent documents
[0005] Patent Document 1: International Publication No. 2018 / 142766 Summary of the Invention
[0006] The problem that the invention aims to solve
[0007] In Patent Document 1, it may be possible to provide the learning completion model required by the user-side device in a short time. However, in Patent Document 1, the learning completion model selected from the database is limited to those that conform to the purpose of the user-side device. That is, if the learning completion model required by the user-side device is, for example, a learning completion model that outputs the user's age when a user's facial image is input, the selected learning completion model is limited to those learning completion models stored in the database that output the user's age when a user's facial image is input. As a result, there is a possibility that the selected learning completion model is inappropriate, and thus there is a possibility that the learning completion model provided to the user-side device cannot provide accurate results.
[0008] Methods for solving problems
[0009] According to this disclosure, the following structure is provided.
[0010] [Structure 1]
[0011] A model creation apparatus is used to create a target model, which is a model adapted to a target environment and configured to output a target output when a target input is received.
[0012] The model creation device includes:
[0013] The candidate model determination unit is configured to determine a plurality of candidate models, which are models that are respectively suitable for a plurality of candidate environments different from the target environment, and are configured to output a reference output when a reference input is input. The reference input and the reference output satisfy one or both of the following conditions: the reference input is different from the target input and the reference output is different from the target output.
[0014] The reference environment determination unit is configured to determine a reference environment from the plurality of candidate environments based on reference output data obtained by inputting teaching data of the reference input associated with the target environment into the plurality of candidate models, and teaching data of the reference output corresponding to the teaching data of the reference input.
[0015] The benchmark model determination unit is configured to determine a model that is suitable for the benchmark environment and is configured to output the target output when the target input is input as the benchmark model;
[0016] The target model creation unit is configured to create the target model based on the baseline model.
[0017] [Structure 2]
[0018] As described in Structure 1, the model creation apparatus, wherein,
[0019] The reference input is different from the target input, and the reference output is different from the target output.
[0020] [Structure 3]
[0021] As described in structure 1 or 2, the model creation apparatus, wherein,
[0022] The reference environment determination unit is configured to determine the candidate model with the highest correlation between the reference output data obtained by inputting the reference input teaching data into the plurality of candidate models and the reference output teaching data, and to determine the candidate environment suitable for the candidate model with the highest correlation as the reference environment.
[0023] [Structure 4]
[0024] The model creation apparatus described in any of structures 1 to 3, wherein,
[0025] The target model creation unit is configured to set the reference model as the target model.
[0026] [Structure 5]
[0027] The model creation apparatus described in any of structures 1 to 3, wherein,
[0028] The target model creation unit is configured to create the target model by fine-tuning the baseline model.
[0029] [Structure 6]
[0030] The model creation apparatus described in any of structures 1 to 3, wherein,
[0031] The target model creation unit is configured to create the target model by performing transfer learning on the baseline model.
[0032] [Structure 7]
[0033] The model creation apparatus described in any of structures 1 to 6, wherein,
[0034] The target environment and the candidate environment are street blocks.
[0035] [Structure 8]
[0036] A model creation method is provided, which creates a target model that is suitable for a target environment and is configured to output a target output when a target input is received.
[0037] The model creation method includes the following processes:
[0038] Multiple candidate models are selected, each of which is a model suitable for multiple candidate environments different from the target environment and is configured to output a reference output when a reference input is input. The reference input and the reference output satisfy one or both of the following conditions: the reference input is different from the target input and the reference output is different from the target output.
[0039] A baseline environment is determined from the plurality of candidate environments based on the reference output data obtained by inputting the teaching data of the reference input associated with the target environment into the plurality of candidate models, and the teaching data of the reference output corresponding to the teaching data of the reference input.
[0040] The model that is adapted to the benchmark environment and is constructed in a manner that outputs the target output when the target input is input is determined as the benchmark model;
[0041] The target model is created based on the baseline model.
[0042] [Structure 9]
[0043] A model creation system is used to create a target model, which is a model adapted to a target environment and configured to output a target output when a target input is received.
[0044] The model creation system has the following features:
[0045] The candidate model determination unit is configured to determine a plurality of candidate models, which are models that are respectively suitable for a plurality of candidate environments different from the target environment, and are configured to output a reference output when a reference input is input. The reference input and the reference output satisfy one or both of the following conditions: the reference input is different from the target input and the reference output is different from the target output.
[0046] The teaching data acquisition unit is configured to acquire teaching data of the reference input associated with the target environment and teaching data of the reference output corresponding to the teaching data of the reference input;
[0047] The reference environment determination unit is configured to determine a reference environment from the plurality of candidate environments based on the reference output data obtained by inputting the reference input teaching data into the plurality of candidate models, and the teaching data of the reference output.
[0048] The benchmark model determination unit is configured to determine a model that is suitable for the benchmark environment and is configured to output the target output when the target input is input as the benchmark model;
[0049] The target model creation unit is configured to create the target model based on the baseline model.
[0050] Invention Effects
[0051] It makes it easier to create models that give more appropriate results. Attached Figure Description
[0052] Figure 1 This is a general overview diagram of a model creation system implemented by the embodiments of this disclosure.
[0053] Figure 2 This is a schematic diagram of a user device implemented according to an embodiment of the present disclosure.
[0054] Figure 3 This is a schematic diagram of a server implemented by the present disclosure.
[0055] Figure 4 This is a schematic diagram illustrating the model creation method implemented by the present disclosure.
[0056] Figure 5 A flowchart illustrating the target model creation routine of an embodiment of this disclosure.
[0057] Figure 6 This is a functional block diagram of the processor of a user device implemented by an embodiment of the present disclosure.
[0058] Figure 7 This is a functional block diagram of the processor of a server implemented according to an embodiment of the present disclosure. Detailed Implementation
[0059] Figure 1 A model creation system 1, an embodiment implemented by this disclosure, is illustrated schematically. When referring to... Figure 1 In this embodiment, the model creation system 1 includes a user device 10 and a server 20. The user device 10 is configured to associate with a target environment ET. Figure 1 In the example shown, user device 10 is located within the target environment ET. On the other hand, in Figure 1 In the example shown, server 20 is located outside the target environment ET. User device 10 and server 20, implemented in the embodiments of this disclosure, are connected together in a manner that enables them to communicate with each other via a communication network N such as the Internet.
[0060] like Figure 2 As shown, the user device 10 implemented by the present disclosure includes one or more processors 11, one or more memories 12, storage devices 13 and input / output interfaces (IF) 14, and these components are connected together in a manner that enables them to communicate with each other via a bidirectional bus.
[0061] The memory 12 implemented in the embodiments of this disclosure includes volatile or non-volatile memory. Various programs are stored in the memory 12, and these programs are executed by the processor 11. The storage device 13 implemented in the embodiments of this disclosure stores completed models, etc.
[0062] In the input / output IF14 of the embodiments implemented by this disclosure, a communication device 15, an input / output device 16, and one or more sensors 17 are communicatively connected. The communication device 15 of the embodiments implemented by this disclosure is communicatively connected to the aforementioned communication network N. The input / output device 16 of the embodiments implemented by this disclosure includes, for example, a keyboard, mouse, media reader / writer, and display. The sensor 17 of the embodiments implemented by this disclosure acquires one or more data related to the target environment ET. In one example, the sensor 17 is installed within the target environment ET. The sensor 17 detects, for example, one or more of the following data related to the target environment ET: weather-related data (temperature, precipitation, humidity, etc.), traffic volume, power consumption, etc.
[0063] On the other hand, such as Figure 3 As shown, the server 20 implemented by the present disclosure includes one or more processors 21, one or more memories 22, storage devices 23, and input / output interfaces (IFs) 24, and these components are connected together in a manner that enables them to communicate with each other via a bidirectional bus.
[0064] The memory 22 implemented in the embodiments of this disclosure includes volatile or non-volatile memory. Various programs are stored in the memory 22, and these programs are executed by the processor 21. The storage device 23 implemented in the embodiments of this disclosure stores completed models, etc.
[0065] In the input / output IF24 of the embodiments implemented by this disclosure, a communication device 25 and an input / output device 26 are connected in a communicative manner. The communication device 25 of the embodiments implemented by this disclosure is connected to the aforementioned communication network N in a communicative manner. The input / output device 26 of the embodiments implemented by this disclosure includes, for example, a keyboard, a mouse, a media reader / writer, and a display.
[0066] Furthermore, the model creation system 1 implemented by this disclosure creates a model suitable for the target environment ET. In the embodiments implemented by this disclosure, a model suitable for a certain environment is suitable when, upon input associated with that environment, an output corresponding to and associated with that environment is produced. In other words, a model suitable for a certain environment is suitable for representing the relationship between the input associated with that environment and the output associated with that environment.
[0067] In embodiments implemented by this disclosure, the target environment ET is a neighborhood, such as a smart city or connected city utilizing big data, etc. In one example, the target environment ET is a new smart city.
[0068] Furthermore, the model creation system 1 implemented in this disclosure creates a target model MT. The target model MT implemented in this disclosure is a model that outputs a target output OT when given a target input IT, or a model that takes a target input IT as input and outputs a target output OT. Additionally, in embodiments implemented in this disclosure, the model is created, for example, through AI or artificial intelligence, particularly machine learning or deep learning. Furthermore, the model uses, for example, neural networks, support vector machines, random forests, etc. In one example, the model is used for inferring feature quantities for smart cities, calculating control parameters for autonomous driving of vehicles, etc. Furthermore, the model's input contains more than one parameter. Similarly, the model's output contains more than one parameter.
[0069] In one example, the target input IT is the temperature in the target environment ET, and the target output OT is the power consumption in the target environment ET.
[0070] In the embodiments implemented by this disclosure, the target model MT is created as follows: First, the target input IT and the target output OT are input to the user device 10 by the user via the input / output device 16. Next, in the user device 10, the reference input IR and the reference output OR are set. In the embodiments implemented by this disclosure, the reference input IR and the reference output OR satisfy one or both of the following conditions: the reference input IR is different from the target input IT, and the reference output OR is different from the target output OT. In other words, the reference input IR and the reference output OR are set to satisfy "IR≠IT and OR≠OT", or "IR=IT and OR≠OT", or "IR≠IT and OR=OT". Further, the reference input IR and the reference output OR are set to not satisfy "IR=IT and OR=OT".
[0071] In one example, with the target input IT being the temperature in the target environment ET and the target output OT being the power consumption in the target environment ET, the reference input IR being the day of the week in the target environment ET, and the reference output OR being the traffic volume in the target environment ET.
[0072] In embodiments implemented by this disclosure, teaching data associated with the target environment ET is then acquired. This teaching data includes teaching data IRT for a reference input IR and teaching data ORT for a reference output OR corresponding to the teaching data of the reference input IR. In one example, the teaching data is acquired via sensor 17 located within the target environment ET. In another example, pre-acquired teaching data associated with the target environment ET is acquired from storage device 13 or input / output device 16. When the acquisition of the teaching data is complete, i.e., for example, when the amount of acquired teaching data reaches a predetermined value, the teaching data, along with an instruction to create a target model MT, is sent from user device 10 to server 20.
[0073] When server 20 receives the instruction, it selects multiple candidate models MC from the models stored in storage device 23 of server 20. Each candidate model MC is a model suitable for multiple candidate environments EC, different from the target environment ET, and is a model that outputs a reference output OR when a reference input IR is given as input. In one example, the candidate environment EC is an existing smart city.
[0074] exist Figure 4 The image shows an example of the candidate model MC. Figure 4 The candidate models MC shown in the example include models MECX(IR, OR), MECY(IR, OR), and MECZ(IR, OR). Model MECX(IR, OR) is a model suitable for the candidate environment ECX, and outputs a reference output OR when a reference input IR is given. Model MECY(IR, OR) is a model suitable for the candidate environment ECY, and outputs a reference output OR when a reference input IR is given. Model MECZ(IR, OR) is a model suitable for the candidate environment ECZ, and outputs a reference output OR when a reference input IR is given. Here, the candidate environments ECX, ECY, and ECZ are different from the target environment ET.
[0075] Next, the teaching data of the reference input IR is input into the candidate model MC, and the reference output OR data is output from the candidate model MC.
[0076] exist Figure 4In the example shown, the teaching data IRT, which is a reference input IR, is input into the candidate model MECX(IR, OR), and the data ORX, which is a reference output OR, is output from the candidate model MECX(IR, OR). Similarly, the teaching data IRT, which is a reference input IR, is input into the candidate models MECY(IR, OR) and MECZ(IR, OR), and the data ORY and ORZ, which are reference output OR, are output from the candidate models MECY(IR, OR) and MECZ(IR, OR), respectively.
[0077] Next, the baseline environment EB is determined from multiple candidate environments EC based on the reference output OR data and the teaching data ORT of the reference output OR. In embodiments implemented by this disclosure, the correlation between the reference output OR data and the teaching data ORT of the reference output OR is calculated. In one example, this correlation is represented by a correlation coefficient CC.
[0078] exist Figure 4 In the example shown, the correlation coefficient CCX between the output data ORX of the candidate model MECX(IR, OR) and the teaching data ORT of the reference output OR is calculated. Similarly, the correlation coefficients CCY and CCZ between the output data ORY and ORZ of the candidate models MECY(IR, OR) and MECZ(IR, OR) and the teaching data ORT of the reference output OR are calculated.
[0079] Next, the candidate model MC with the highest correlation between the reference output OR data of the candidate model EC and the teaching data ORT of the reference output OR is determined. Furthermore, the candidate environment EC to which the candidate model MC is suitable is determined as the baseline environment EB.
[0080] exist Figure 4 In the example shown, when the correlation coefficient CCX is significantly larger than the correlation coefficients CCY and CCZ, the candidate model MECX (IR, OR) is determined to be the candidate model MC with the highest correlation. In this case, the candidate environment ECX is determined to be the baseline environment EB.
[0081] Next, the model MEB(IT, OT) is determined as the baseline model MB, which is a model suitable for the baseline environment EB and outputs the target output OT when the target input IT is input. At least one model suitable for the baseline environment EB is stored in the storage device 23 of the server 20 implemented by this disclosure, and the baseline model MB is determined from these models.
[0082] exist Figure 4In the example shown, when the candidate environment ECX is determined to be the baseline environment EB, the model MECX(IT, OT) is determined to be the baseline model MB. The model MECX(IT, OT) is a model that is suitable for the baseline environment ECX and outputs the target output OT when the target input IT is input.
[0083] Next, the target model MT is created based on the baseline model MB.
[0084] The embodiments of this disclosure will be further illustrated by specific examples. The target environment ET is a new smart city, the target input IT is the day of the week, and the target output OT is traffic volume. In this case, the target model MT is a model suitable for the new smart city, and is a model MN (day of the week, traffic volume) that outputs traffic volume when the input includes the day of the week. On the other hand, the reference input IR is temperature, and the reference output OR is power consumption. The candidate environments EC are existing smart cities X, Y, and Z. The candidate models are model MX (temperature, power consumption), model MY (temperature, power consumption), and model MZ (temperature, power consumption), wherein model MX (temperature, power consumption) is a model suitable for existing smart city X, and outputs power consumption when the input includes temperature; model MY (temperature, power consumption) is a model suitable for existing smart city Y, and outputs power consumption when the input includes temperature; and model MZ (temperature, power consumption) is a model suitable for existing smart city Z, and outputs power consumption when the input includes temperature. Next, the temperature in the new smart city is input into the candidate model MX (temperature, power consumption), and power consumption data is output from the candidate model MX (temperature, power consumption). Similarly, the temperature in the new smart city is input into the candidate models MY (temperature, power consumption) and MZ (temperature, power consumption), and power consumption data is output from the candidate models MY (temperature, power consumption) and MZ (temperature, power consumption), respectively. Then, the correlation coefficients between the power consumption data output from the candidate models MX (temperature, power consumption), MY (temperature, power consumption), and MZ (temperature, power consumption) and the power consumption in the new smart city are calculated. If the candidate model with the highest correlation coefficient is MX (temperature, power consumption), the existing smart city X is determined as the baseline environment EB. Next, model MX (day of the week, traffic volume) is determined as the baseline model MB, which is a model suitable for the existing smart city X, and outputs traffic volume when a day of the week is input. The target model MN (day of the week, traffic volume) is created based on the baseline model MX (day of the week, traffic volume).
[0085] In the first creation example of the target model MT, the baseline model MB is set as the target model MT.
[0086] In the second creation example of the target model MT, the target model MT is created by fine-tuning the baseline model MB. In the fine-tuning of the embodiment implemented by this disclosure, the weighted average of the layers of the baseline model MB is relearned using teaching data (teaching data for target input and teaching data for target output) without changing the number of layers in the baseline model MB.
[0087] In the third example of creating the target model MT, the target model MT is created by performing transfer learning on the baseline model MB. In the transfer learning of the embodiments of this disclosure, at least one layer is added to the baseline model MB without changing the weighting of the layers, and the weighting of the added layer is learned using teaching data (teaching data for the target input and teaching data for the target output).
[0088] In an embodiment implemented by this disclosure, when the target model MT is created, it is then sent from server 20 and received by user device 10. The target model MT is stored in, for example, storage device 13 of user device 10.
[0089] In the embodiments implemented by this disclosure, the target model MT is then used in the user device 10. That is, the target input IT data is input into the target model MT, and the target output OT data is output from the target model MT.
[0090] Thus, in the embodiments implemented by this disclosure, since a new model (target model MT) is created using the already created model, it is easier to create a new model. Furthermore, since the baseline environment EB is considered to be highly correlated with the target environment ET, the target model MT created based on the baseline model MB, which is suitable for the baseline environment EB, can provide more appropriate results for the target environment ET.
[0091] Figure 5 A routine for creating a target model MT, as implemented in this disclosure, is shown. When referring to... Figure 5 In step 100, the target input IT and target output OT are acquired in user device 10. In the next step 101, the reference input IR and reference output OR are set in user device 10. In the next step 102, teaching data is acquired in user device 10. In the next step 103, user device 10 determines whether the acquisition of teaching data has been completed. If it is determined that the acquisition of teaching data has not been completed, the process returns to step 102. If it is determined that the acquisition of teaching data has been completed, the process proceeds to step 104, where user device 10 sends the teaching data and an instruction to create a target model MT to server 20.
[0092] In the next step 200, teaching data and creation instructions are received on server 20. In the next step 201, a candidate model MC is determined on server 20. In the next step 202, the reference output OR data output from the candidate model MC is obtained on server 20. In the next step 203, the correlation coefficient CC is calculated on server 20. In the next step 204, the baseline environment EB is determined on server 20. In the next step 205, the baseline model MB is determined on server 20. In the next step 206, the target model MT is created on server 20. In the next step 207, the target model MT is sent on server 20.
[0093] In the next step 300, the target model MT is received in the user device 10. In the next step 301, the target model MT is used in the user device 10.
[0094] Figure 6 This is a functional block diagram of the processor 11 of the user device 10, implemented by this disclosure, in relation to the creation of the target model MT. (Refer to...) Figure 6 At that time, the processor 11 includes a teaching data acquisition unit 11a. The teaching data acquisition unit 11a is configured to acquire teaching data of reference input associated with the target environment and teaching data of reference output corresponding to the teaching data of reference input.
[0095] Figure 7 This is a functional block diagram of the processor 11 of the server 20, implemented by this disclosure, in relation to the creation of the target model MT. (Refer to...) Figure 7At the time, the processor 21 includes a candidate model determination unit 21a, a reference environment determination unit 21b, a reference model determination unit 21c, and a target model creation unit 21d. The candidate model determination unit 21a is configured to determine multiple candidate models, each of which is suitable for multiple candidate environments different from the target environment, and is configured to output a reference output when a reference input is input. Here, the reference input and reference output satisfy one or both of the following: the reference input is different from the target input, and the reference output is different from the target output. The reference environment determination unit 21b is configured to determine a reference environment from the multiple candidate environments based on the reference output data obtained by inputting the teaching data of the reference input into the multiple candidate models, and the teaching data of the reference output. The reference model determination unit 21c is configured to determine the model that is suitable for the reference environment and is configured to output a target output when a target input is input as the reference model. The target model creation unit 21d is configured to create a target model based on the reference model. In addition, the server 20 implemented in the embodiments of this disclosure functions as a model creation apparatus.
[0096] Symbol Explanation
[0097] 1…Model creation system;
[0098] 10…User Equipment;
[0099] 20… server (model creation device);
[0100] 21… processors;
[0101] 21a… Candidate Model Decision Unit;
[0102] 21b…Benchmark Environment Decision Department;
[0103] 21c…Benchmark Model Determination Unit;
[0104] 21d…Target Model Creation Department.
Claims
1. A model creation apparatus for creating a target model, said target model being a model adapted to a target environment and configured to output a target output when a target input is input, wherein, The model creation device includes: The candidate model determination unit is configured to determine a plurality of candidate models, which are models that are respectively suitable for a plurality of candidate environments different from the target environment, and are configured to output a reference output when a reference input is input. The reference input and the reference output satisfy one or both of the following conditions: the reference input is different from the target input and the reference output is different from the target output. The reference environment determination unit is configured to determine a reference environment from the plurality of candidate environments based on reference output data obtained by inputting teaching data of the reference input associated with the target environment into the plurality of candidate models, and teaching data of the reference output corresponding to the teaching data of the reference input, wherein the teaching data is detection data associated with the target environment obtained by a sensor; The benchmark model determination unit is configured to determine a model that is suitable for the benchmark environment and is configured to output the target output when the target input is input as the benchmark model; The target model creation unit is configured to create the target model based on the baseline model.
2. The model creation apparatus as described in claim 1, wherein, The reference input is different from the target input, and the reference output is different from the target output.
3. The model creation apparatus as described in claim 1 or 2, wherein, The reference environment determination unit is configured to determine the candidate model with the highest correlation between the reference output data obtained by inputting the reference input teaching data into the plurality of candidate models and the reference output teaching data, and to determine the candidate environment suitable for the candidate model with the highest correlation as the reference environment.
4. The model creation apparatus as described in claim 1 or 2, wherein, The target model creation unit is configured to set the reference model as the target model.
5. The model creation apparatus as described in claim 1 or 2, wherein, The target model creation unit is configured to create the target model by fine-tuning the baseline model.
6. The model creation apparatus as described in claim 1 or 2, wherein, The target model creation unit is configured to create the target model by performing transfer learning on the baseline model.
7. The model creation apparatus as described in claim 1 or 2, wherein, The target environment and the candidate environment are street blocks.
8. A model creation method, comprising creating a target model, wherein the target model is a model suitable for a target environment and is configured to output a target output when a target input is received, wherein... The model creation method includes the following processes: Multiple candidate models are selected, each of which is a model suitable for multiple candidate environments different from the target environment and is configured to output a reference output when a reference input is input. The reference input and the reference output satisfy one or both of the following conditions: the reference input is different from the target input and the reference output is different from the target output. A reference environment is determined from the plurality of candidate environments based on the reference output data obtained by inputting the reference input teaching data associated with the target environment into the plurality of candidate models, and the reference output teaching data corresponding to the reference input teaching data, wherein the teaching data is detection data associated with the target environment obtained by sensors; The model that is adapted to the benchmark environment and is constructed in a manner that outputs the target output when the target input is input is determined as the benchmark model; The target model is created based on the baseline model.
9. A model creation system for creating a target model, said target model being a model adapted to a target environment and configured to output a target output when a target input is received, wherein... The model creation system has the following features: The candidate model determination unit is configured to determine a plurality of candidate models, which are models that are respectively suitable for a plurality of candidate environments different from the target environment, and are configured to output a reference output when a reference input is input, wherein the reference input and the reference output satisfy one or both of the following conditions: the reference input is different from the target input and the reference output is different from the target output. The teaching data acquisition unit is configured to acquire teaching data of the reference input associated with the target environment and teaching data of the reference output corresponding to the teaching data of the reference input, wherein the teaching data is detection data associated with the target environment acquired by a sensor; The reference environment determination unit is configured to determine a reference environment from the plurality of candidate environments based on the reference output data obtained by inputting the reference input teaching data into the plurality of candidate models, and the teaching data of the reference output. The benchmark model determination unit is configured to determine a model that is suitable for the benchmark environment and is configured to output the target output when the target input is input as the benchmark model; The target model creation unit is configured to create the target model based on the baseline model.
Citation Information
Patent Citations
Learned model provision method and learned model provision device
WO2018142766A1
Method, apparatus and storage medium for predicting key quality indicators of video service
CN109040744A
Learned model provision method and learned model provision device
US20190147361A1