TWI739124B - Cloud transaction system and method for providing neural network training model in supervised state - Google Patents

Abstract

A cloud trading system and method that provide a neural network-like training model in a supervised state. It mainly trains multiple data data with different training programs, and then obtains different training models, which can be used by the client to remotely The terminal conducts transactions through a third-party trading platform, and downloads the aforementioned different training models according to the transaction results under the supervision of the supervision unit, so that the user terminal can compare the accuracy differences between different training models to facilitate training The accuracy of the model.

Description

Cloud transaction system and method for providing neural network training model in supervised state

The present invention is related to the cloud transaction technology of the neural network training model, and particularly refers to a kind of data provided in a supervised state for users to freely use training data and conduct paid transactions in a huge graphical medical data database. A cloud trading system and method similar to neural network training model.

The stacking relationship between artificial intelligence, machine learning, and neural network-like deep learning is inseparable. The result is the accuracy of functional prediction in each field. Among them, the most basic deep learning is the soul of machine learning, which can show human beings. Of wisdom. However, deep learning requires a meticulous neural network-like architecture to perform characterization learning algorithms on data to achieve high-accuracy functional predictions. At present, there is no complete, complete, and massively accurate graphical medical treatment. The data database is used by artificial intelligence talents all over the world, which makes the development and application of artificial intelligence in the medical field suffer a bottleneck; on the other hand, a unit that has a complete, complete, massive and accurate graphical medical data database is usually unwilling to disclose it. These data are used for research and development on their own. However, the talents are limited. There is a complete, complete, and massive and accurate medical image labeling data database. However, there is no talent to develop new algorithms for training high-accuracy models and high-precision models. Accurate application results in slow improvement of model accuracy and small application range.

For a piece of data, the technique of forming a model through neural network-like training is also a well-known technique. For example, my country's announcement of Patent No. TW I645303, its claim 18 discloses the technology of training a feature data with a similar neural network. Another example is my country's Announcement No. TW I662511, and its request items 5 and 6 reveal the technology of using different deep convolutional neural network (DCNN) training to form different classification models.

It is that the inventor of the present invention is based on the concept of continuous product improvement and innovation, based on years of practical experience in product design and development, and active research and development thinking, through countless actual development experiences, leading to the invention of this invention.

In the current massive graphical medical data database, there is no mechanism that allows users to obtain a training model trained with a neural network through a paid transaction mechanism, which also causes the training model to fail to further improve the accuracy. space.

In order to solve the above problems, the present invention proposes a cloud transaction system and method for providing neural network-like training models in a supervised state, which can provide a huge graphical medical data database as data, and provide completed The trained pre-training model and the training program provided to allow the user to obtain the completed training model trained by this training program, so that the user can understand the difference in accuracy of different training models.

In order to achieve the above effects, the present invention proposes a cloud transaction system that provides a neural network-like training model in a supervised state. A pre-training model completed by training a plurality of data data, a first training program of at least one type of neural network, and the training of the at least one first training program At least one completed training model completed by a plurality of data data, the accuracy of the at least one completed training model is different from the accuracy of the pre-training model; a supervision unit, coupled to the storage unit; a transaction port, coupled to The monitoring unit, the transaction port is used for the monitoring unit to be coupled to an external third-party transaction platform; and an ingress port is coupled to the monitoring unit, and the port is used for a client to be coupled via the cloud ; Wherein: the monitoring unit is used to receive a transaction information from the transaction port, and according to the transaction information to allow a specific client to download the pre-training model and the at least one first training program from the storage unit via the connection port And at least one of the at least one completed training model.

In addition, the present invention also proposes a transaction method that provides a neural network-like training model in a supervised state, which includes the following steps: a client is coupled to the port via the cloud; the client selects the desired purchase At least one of the pre-training model, the at least one first training program, and the at least one completed training model; the client pays the third-party trading platform to purchase the pre-training model, the at least one first training program, and For the fee corresponding to at least one of the three at least one completed training model, the third-party trading platform sends a transaction information to the supervision unit, and the supervision unit opens the downloading authority of the client according to the transaction information; and the user The terminal downloads at least one of the pre-training model, the at least one first training program, and the at least one completed training model that it has purchased under the supervision of the supervision unit through the connection port.

In this way, the present invention can provide a huge graphical medical data database for users to use, and at the same time provide a pre-trained model that has been trained, and provide at least one training program for the user to obtain a new training model. In order to improve the accuracy of the model.

In addition, the present invention can also allow users to upload additional training programs by themselves to train the same data, thereby obtaining additional training models. That is, in addition to the aforementioned pre-trained model and completed training model, the present invention also allows the user to provide a training program to The aforementioned data is trained to obtain a new training model, so that the accuracy of the training model can be continuously improved by improving the neural network-like machine learning and deep learning functions.

When uploading, the client uploads a second training program of a type of neural network through the connection port, the supervision unit releases the second training program to be stored in the storage unit, and the cloud server uses the first training program to store it in the storage unit. The second training program trains the plurality of data data to generate an improved training model, and stores the improved training model in the storage unit.

10: A cloud trading system that provides a neural network training model in a supervised state

100: Cloud server

1001: physical computer server

101: Data Sheet

103: storage unit

1031: storage equipment

104: The first training program

106: trading port

108: Connect to Port

201: Client

203: Service Control Interface

300: pre-trained model

400: Complete the training model

600: Supervision Unit

702: Third-party trading platform

703: Transaction Information

10’: A cloud trading system that provides a neural network-like training model in a supervised state

100’: Cloud server

101’: Data

103’: Storage unit

108’: Connected to the port

201’: Client

500’: Improved training model

503’: The second training program

600’: Supervision Unit

703’: Transaction Information

Fig. 1 is a block diagram of the first preferred embodiment of the present invention.

Figure 2 is a flowchart of the first preferred embodiment of the present invention.

Fig. 3 is another block diagram of the first preferred embodiment of the present invention.

Fig. 4 is a block diagram of the second preferred embodiment of the present invention.

Fig. 5 is a flowchart of the second preferred embodiment of the present invention.

In order to explain in detail the technical features of the present invention, the following preferred embodiments are described in conjunction with the drawings. Among them: as shown in Figures 1 to 2, the first preferred embodiment of the present invention is proposed in The cloud trading system 10, which provides a neural network training model in a supervised state, is mainly composed of a storage unit 103, a supervision unit 600, a transaction port 106, and a connection port 108, of which: The storage unit 103 stores a plurality of data data 101, a pre-trained model 300 completed by training the plurality of data data 101 by a predetermined type of neural network (not shown in the figure), and at least one type of neural network The first training program 104 and at least one completed training model 400 completed by the at least one first training program 104 training the plurality of data data 101, the accuracy of the at least one completed training model 400 is different from that of the pre-training model 300 Accuracy. The predetermined neural network may be a training program of a high-accuracy algorithm developed by the manufacturer and not disclosed to the public, and the predetermined neural network and the at least one first training program 104 may be a convolutional neural network (CNN , Convolutional Neural Network) or Recurrent Neural Network (RNN) or extreme gradient boosting (eXtreme Gradient Boosting, XGBoost) or random forest (Random Forest) or gradient boosting decision tree (Gradient Boosting Machine) or support vector machine ( Support Vector Machine), etc., and the predetermined neural network and the algorithm of the at least one first training program 104 are different, so that the accuracy of the pre-training model 300 and the at least one completed training model 400 can be different. Among them, the plurality of data materials 101 are graphical medical data materials. The at least one first training program 104 can be one or more in number. In the case of multiple first training programs 104, each first training program 104 will be based on the aforementioned different neural network types in practice. (Convolutional Neural Network (CNN, Convolutional Neural Network) or Recurrent Neural Network (RNN) or Extreme Gradient Boosting (XGBoost) or Random Forest or Gradient Boosting Decision Tree (Gradient) Boosting Machine (Boosting Machine) or Support Vector Machine (Support Vector Machine)) have different algorithms, and there are differences between each first training program 104. The following descriptions are based on plural first training programs 104, so The at least one completed training model 400 is also taken as an example, and the accuracy of each model will be different.

The monitoring unit 600 is coupled to the storage unit 103.

The transaction port 106 is coupled to the monitoring unit 600, and the transaction port 106 is used for the monitoring unit 600 to be coupled to an external third-party transaction platform 702.

The connection port 108 is coupled to the monitoring unit 600, and the connection port 108 is used for a client 201 to be coupled to the cloud. In actual implementation, the storage unit 103, the monitoring unit 600, the transaction port 106, and the connection port 108 can be integrated into a cloud server 100, where the cloud server 100 can be a single physical computer server. It can also be a large-scale system formed by a combination of multiple physical computer servers. In this embodiment, a physical computer server is taken as an example for illustration, and the physical computer server itself has a function of computing processing. The client 201 can be a computer or a smart phone in practice. In addition, after the client 201 is coupled to the inbound port 108, in practice, there will be a service control interface 203 for the client 201 to operate. This service control interface 203 can be a part of the system, that is, the system 10 also includes the service control interface 203, and when the client 201 is coupled to the inbound port 108, the service control interface 203 is provided by the inbound port 108 to the client 201 to open, for example, with a web page The form of is displayed on the client 201. In addition, in other embodiments, the service control interface 203 can also be a program directly installed on the client 201 and displayed after being coupled to the port 108. In this case, the The service control interface 203 is not included in the system, but is a program installed on the client 201. The service control interface 203 provides the pre-training model 300, the plurality of first training programs 104, and the plurality of completion training model 400 options for the client 201 to purchase, and the service control interface 203 provides options for the client 201 to purchase. The link connected to the three-party transaction platform 702 allows the user terminal 201 to click on the link to couple with the third-party transaction platform 702 and make payment when deciding to purchase an item and want to pay.

Wherein, the monitoring unit 600 is used for receiving a transaction information 703 sent by the third-party transaction platform 702 from the transaction port 106, and allowing the client 201 according to the transaction information 703 At least one of the pre-training model 300, the plurality of first training programs 104, and the plurality of completion training models 400 is downloaded from the storage unit 103 via the connection port 108.

As shown in Figure 2, in the first embodiment, when a transaction is performed, the transaction method is the following steps:

S1: The client 201 is coupled to the connection port 108 via the cloud.

S2: The client 201 uses the service control interface 203 to select at least one of the pre-trained model 300, the plurality of first training programs 104, and the plurality of completed training models 400 to be purchased. The user terminal 201 selects all of the foregoing three as an example.

S3: The client 201 pays the third-party transaction platform 702 for the corresponding fees for purchasing the pre-trained model 300, the plural first training program 104, and the plural completed training model 400, and the third-party transaction platform 702 sends a The transaction information 703 is given to the supervisory unit 600, and the supervisory unit 600 grants the download authority of the client 201 based on the transaction information 703.

S4: The client 201 downloads the pre-trained model 300, the plurality of first training programs 104, and the plurality of completed training models 400 that it purchased under the supervision of the supervision unit 600 through the connection port 108.

After completing the above steps, the client 201 downloads the purchased subject matter under the supervision of the supervision unit 600. The aforementioned pre-training model 300 and the complex completion training model 400 are different in accuracy, mainly for the purpose of allowing the client 201 to understand the difference in accuracy of training models trained with different training programs after downloading. Then, based on this difference, it is possible to decide whether to develop a more accurate training program to improve the accuracy of the training model. It can be seen that, while providing a huge graphical medical data database for users to use, the present invention also provides a pre-trained model 300 that has been trained, and provides the plurality of first training programs 104 for users to obtain This complex number completes the training The model 400 can improve the accuracy of the training model through the difference in accuracy of different training models and the difference in the machine learning and deep learning functions of the neural network-like network of the plural first training program 104. In addition, in the case where there is only one first training program 104, there is no difference between the plural first training programs 104, and furthermore, only one first training program 104 will complete the training model. 400. Similarly, there will be no difference in accuracy of multiple completed training models 400. In such a situation, only the accuracy of the completed training model 400 and the pre-trained model 300 can be compared with each other in accuracy.

In addition, it should be added that, as shown in Figure 3, the storage unit 103, the supervision unit 600, the transaction port 106 and the connection port 108, if integrated in a combination of multiple physical computer servers 1001 In the system, the storage unit 103 will also be jointly formed by the storage devices 1031 in a plurality of physical computer servers 1001. In this case, the pre-training model 300, the plural first training programs 104 and The plurality of completed training models 400 are stored in a storage device 1031 of a physical computer server 1001, and only the physical computer server 1001 is used to set the transaction port 106 and the connection port 108, and the plurality of data data 101 are stored The storage device 1031 of the other physical computer server 1001 in the joint system. In this way, the effect of storage can also be achieved, and the plurality of data data 101 can be separated without being stored in the same storage together with the pre-training model 300, the plurality of first training programs 104, and the complex training model 400 In the device 1031, in this way, it can be ensured that the user can only download the pre-trained model 300, the plural first training program 104, and the plural completed training model 400, but has no chance to download the plural data 101, which achieves additional protection The effect of the plural data.

Please refer to FIGS. 4 to 5 again. The second preferred embodiment of the present invention proposes a cloud trading system 10' that provides a neural network-like training model in a supervised state, which is basically the same as the first embodiment disclosed above. The difference is that: in the second embodiment, the client 201' uploads a second training program 503' of a type of neural network through the connection port 108', and the monitoring unit 600' releases the second training program 503' The training program 503' is stored in the storage unit 103', the cloud server 100' trains the plurality of data 101' with the second training program 503' to generate an improved training model 500', and trains the improvement The model 500' is stored in the storage unit 103', and the cloud server 100' is composed of a physical computer server with processing and computing capabilities, so the above-mentioned training behavior can be performed. Since the second training program 503' is uploaded by the client 201' to the cloud server 100', the cloud server 100' can allow the client 201' to download the improved training model 500', that is, In the absence of the transaction information 703', the supervision unit 600' only allows the client 201' to download from the storage unit 103' via the connection port 108' and is trained by the second training program 503' The improved training model 500'.

When the system 10' of the second embodiment conducts transactions, in addition to the above-mentioned steps S1 to S4, the transaction method may further include a step Sn: the client 201' uploads the data through the connection port 108' The second training program 503', the supervision unit 600' releases the second training program 503' to be stored in the storage unit 103', and the cloud server 100' trains the plurality of data with the second training program 503' The data 101' generates the improved training model 500', and the improved training model 500' is stored in the storage unit 103'. This step Sn only needs to be after step S1, and it can be before or after or between S2 and S4. In FIG. 5, Sn is after S4 as an example.

It can be seen from the above that in addition to allowing the client 201’ to download the purchased subject matter under supervision, the second embodiment can obtain the training model it wants to obtain and understand the difference in accuracy. In addition, the user terminal 201' can also upload the second training program 503' that it thinks is better, and train the improved training model 500' through the second training program 503'. In this way, the user can freely use the plurality of data 101' to further develop a better training program, thereby improving the accuracy of the training model to a greater extent.

The remaining technical features and achievable effects of the second embodiment are the same as those of the first embodiment disclosed above, and will not be repeated here.

The plural data data 101 of the present invention are medical imaging data collected by China Medical University and affiliated hospitals, and are clinical trials approved by China Medical University & Hospital Research Ethics Committee (China Medical University & Hospital Research Ethics Committee) plan.

The above-mentioned embodiments of the present invention are provided for the convenience of explanation only, and should not be used to limit the meaning of this case, that is, all the various alterations and designs based on the scope of the listed patent applications should be included in the scope of patents in this case.

10: A cloud trading system that provides a neural network training model in a supervised state

100: Cloud server

101: Data Sheet

103: storage unit

104: The first training program

106: trading port

108: Connect to Port

201: Client

203: Service Control Interface

300: pre-trained model

400: Complete the training model

600: Supervision Unit

702: Third-party trading platform

703: Transaction Information

Claims (11)

A cloud trading system that provides a neural network training model in a supervised state, comprising: a storage unit storing at least a plurality of data data, and a predetermined neural network training the plurality of data data is completed The pre-training model, the first training program of at least one type of neural network, and at least one completed training model completed by training the plurality of data data by the at least one first training program, the accuracy of the at least one completed training model is different On the accuracy of the pre-trained model; a monitoring unit coupled to the storage unit; a transaction port coupled to the monitoring unit, the transaction port for the monitoring unit to be coupled to an external third-party trading platform ; And a connection port, coupled to the monitoring unit, the connection port for a client to be coupled via the cloud; wherein: the monitoring unit is used to receive a transaction information from the transaction port, and based on the transaction information Allow a specific client to download at least one of the pre-training model, the at least one first training program, and the at least one completed training model from the storage unit via the connection port. According to the first item of the scope of patent application, the cloud transaction system that provides neural network training model in a supervised state, wherein: the storage unit, the supervision unit, the transaction port and the connection port are integrated in a cloud Server. According to the second item of the scope of patent application, the cloud transaction system that provides neural network-like training models in a supervised state, wherein: the cloud server has at least one physical computer server. The cloud trading system that provides neural network training models in a supervised state according to item 2 of the scope of patent application, wherein: the client uploads a second training program of a neural network through the connection port, The supervision unit releases the second training program to be stored in the storage unit, and the The cloud server trains the plurality of data data with the second training program to generate an improved training model, and stores the improved training model in the storage unit. According to item 4 of the scope of patent application, the cloud transaction system that provides neural network training model in a supervised state, wherein: the supervising unit only allows the client to connect through the connection without the transaction information The port downloads the improved training model trained by the second training program from the storage unit. According to the first item of the scope of patent application, the cloud transaction system that provides neural network training model under supervision, which further includes a service control interface, when the client is coupled to the ingress port, the The service control interface is provided by the connection port to the client to open, and the service control interface provides the pre-training model, the at least one first training program, and the at least one option to complete the training model for the client Optional, and the service control interface provides a link to the third-party trading platform. A trading method that uses the system described in item 1 to provide a neural network-like training model in a supervised state includes the following steps: a client is coupled to the port via the cloud; the client selects all At least one of the pre-trained model to be purchased, the at least one first training program, and the at least one completed training model; the client pays the third-party transaction platform to purchase the pre-trained model, the at least one first training model A training program and the cost corresponding to at least one of the three at least one completed training model, the third-party trading platform sends a transaction information to the monitoring unit, and the monitoring unit opens the download authority of the client based on the transaction information ;as well as The client downloads at least one of the pre-trained model, the at least one first training program, and the at least one completed training model selected by the client under the supervision of the supervision unit through the connection port. According to item 7 of the scope of patent application, a neural network training model-like transaction method is provided in a supervised state, wherein: the storage unit, the supervision unit, the transaction port, and the connection port are integrated into a cloud server器中。 According to item 8 of the scope of patent application, the trading method of providing neural network training model in a supervised state further includes a step: the client uploads the second type of neural network through the connection port A training program, where the supervision unit releases the second training program to be stored in the storage unit, the cloud server trains the plurality of data data with the second training program to generate an improved training model, and uses the improved training model Stored in the storage unit. According to item 9 of the scope of patent application, the method of providing neural network training model in a supervised state, wherein: the supervising unit only allows the client to pass through the connection port without the transaction information The improved training model trained by the second training program is downloaded from the storage unit. According to item 7 of the scope of patent application, the trading method that provides a neural network training model in a supervised state further includes a service control interface. When the client is coupled to the ingress port, the service The control interface is provided by the connection port to the client to open, and the service control interface provides the pre-trained model, the at least one first training program, and the at least one option to complete the training model for the client to select And the service control interface provides a link to the third-party trading platform.

Images