Interactive learning model training method and system based on machine learning
By introducing an interactive learning model training method into the neural network model, using the central processing library to evaluate the matching degree and optimize the neural network response model, the problem of insufficient instruction matching degree when the training data of the neural network model is insufficient, and effective optimization of the neural network model is achieved.
Patent Information
- Application Number
- CN202411692667.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-25
- Publication Date
- 2025-02-28
AI Technical Summary
When the training data of the neural network model is insufficient, the matching degree of the instructions generated and the input data is insufficient. The existing optimization methods are usually optimized for continuous function models. For the interruption of the neural network model, its optimization method is difficult to work.
Using an interactive learning model training method based on machine learning, by establishing a neural network reaction model, receiving interactive data and signal transmission, the central processing library evaluates the matching degree between the execution command and the interactive data, scores the matching degree, and optimizes the low-rated neural network reaction model.
By optimizing the neural network response model and ensuring that it can return high matching execution commands for the input data, the problem that existing optimization methods are difficult to optimize the neural network model is solved.
Smart Images
Figure CN119539125A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of data processing technology, and in particular to a method and system for training an interactive learning model based on machine learning. Background Art
[0002] Traditional Network Attached Storage (NAS) designs a sub-neural network model by using a recurrent neural network (RNN) to generate a configuration string containing the architecture information of the neural network model, which is applied or will be applied on the device. Thereafter, the performance of the sub-neural network model is measured by using a validation set, and the accuracy value R obtained as a result can be used as a reward in reinforcement learning to update the parameters of the sub-neural network model. In addition, techniques such as gradient descent optimization and evolutionary search are also used to achieve optimal network performance.
[0003] Optimizing the neural network model architecture by using neural networks is useful for installation on autonomous mobile machines (such as autonomous vehicles, drones, robots, etc.) that require deep learning technology. However, the traditional method of using neural network models for intended purposes has the following problems: when the neural network model training data is insufficient, the instructions it generates do not match the input data sufficiently, that is, the instructions are not the instructions required by the input data. However, existing optimization methods are usually optimized for continuous function models, and are difficult to work for the discontinuous case of neural network models. Summary of the invention
[0004] In order to solve the above technical problems, a method and system for training an interactive learning model based on machine learning are provided. The technical solution solves the problem raised in the above background technology that when the neural network model training data is insufficient, the instructions generated by it do not match the input data sufficiently, that is, the instructions are not the instructions required by the input data. However, the existing optimization methods are usually optimized for continuous function models, and the optimization methods are difficult to work for the discontinuous situation of the neural network model.
[0005] In order to achieve the above purpose, the technical solution adopted by the present invention is:
[0006] An interactive learning model training method based on machine learning, comprising:
[0007] Establishing a neural network response model, the neural network response model includes an input layer, a hidden layer and an output layer;
[0008] The neural network reaction model receives the interaction data, which is imported into the input layer. The input layer extracts at least one keyword from the interaction data. The input layer compares the keywords stored in the input neurons with the extracted keywords, and selects the keywords stored in the input neurons that are consistent or similar.
[0009] The keywords stored in each selected input neuron transmit signals unidirectionally in the input layer, the hidden layer and the output layer according to the unidirectional flow path. When the signal is transmitted to the output layer, the execution command stored in the output neuron where the output signal is located, for keywords stored in multiple selected input neurons, they correspond to execution commands stored in different output neurons respectively. For keywords stored in each selected input neuron, there is a set consisting of at least one output neuron corresponding to each keyword stored in the selected input neuron. The intersection of the sets of these output neurons is determined, and the execution command of the intersection of the sets of the output neurons is the execution command finally output by the neural network reaction model;
[0010] The central processing library calculates the similarity between the execution command and the keywords of the interaction data, builds a matching value calculation model, determines whether the neural network response model has defects, and optimizes the neural network response model;
[0011] The central processing library collects at least one execution information corresponding to the interactive data through the network, the central processing library collects a network event set of the execution information through the network, calculates the conditional probability of the interactive data under the condition that the execution information appears, and determines the pre-execution command from the at least one execution information according to the determination rule;
[0012] The central processing library compares whether the pre-execution command is included in the execution command stored in the output neuron and takes different execution operations.
[0013] Preferably, the input layer is provided with a plurality of parallel input neurons, which store keywords; the hidden layer is provided with a plurality of levels of judgment neurons, which store conditional judgment information; the output layer is provided with a plurality of parallel output neurons, which store execution commands; and the input layer, the hidden layer and the output layer are connected by a unidirectional flow path.
[0014] Preferably, the input neuron is connected to the judgment neuron with a unidirectional flow path, the multi-level judgment neurons are connected with a unidirectional flow path, the judgment neuron is connected to the output neuron with a unidirectional flow path, the signal is transmitted unidirectionally between the neurons connected with the unidirectional flow path, the signal of the neuron not connected with the unidirectional flow path cannot be transmitted, and when the signal is transmitted to the output layer, the output layer outputs the execution command stored in the output neuron where the signal is located;
[0015] The multi-level judgment neurons are composed of multiple judgment neuron circuits, and a unidirectional flow path is connected between the multiple judgment neurons. The judgment neuron circuits are composed of multiple judgment neurons, and the judgment neurons are arranged from low to high levels. The lowest-level judgment neurons are unidirectionally connected to the input neurons, and the highest-level input neurons are unidirectionally connected to the output neurons.
[0016] Preferably, the central processing library calculates the similarity between the execution command and the interaction data keyword as follows:
[0017]
[0018] Where Mi is the matching similarity between the execution command and the i-th keyword in the interaction data, aij is the j-th matching index value of the i-th keyword in the execution command, bij is the j-th matching index value of the i-th keyword in the interaction data, and n is the total number of matching index values.
[0019] Preferably, the calculation formula of the matching value model is:
[0020]
[0021] Where Y is the matching value between the execution command and the interaction data, αi is the weight value of the similarity of the i-th keyword, and h is the total number of keywords.
[0022] Preferably, the determining whether the neural network reaction model has defects and optimizing the neural network reaction model specifically comprises the following steps:
[0023] Calculate the matching value through the calculation formula of the matching model;
[0024] Comparing the matching value with a first preset threshold and a second preset threshold;
[0025] If the matching value is greater than 0 and less than the first preset threshold, the matching degree between the execution command and the interaction data is level one;
[0026] If the matching value is greater than or equal to the first preset threshold value, but less than the second preset threshold value, the matching degree between the execution command and the interaction data is level 2;
[0027] If the matching value is greater than or equal to the second preset threshold, the matching degree between the execution command and the interaction data is level three;
[0028] When the matching degree between the execution command and the interaction data is level one, there are defects in the neural network reaction model, and the neural network reaction model is optimized.
[0029] Preferably, if the central processing library compares the pre-execution command to be included in the execution command stored in the output neuron, the pre-execution command corresponds to the input neuron, finds the position of the highest-level judgment neuron that transmits the interactive data signal to the multi-level judgment neuron, and there is at least one position of the highest-level judgment neuron where the signal is located, deletes the unidirectional flow path connecting the highest-level judgment neuron where the signal is located and the output neuron, and generates a new unidirectional flow path between the highest-level judgment neuron where the signal is located and the output neuron corresponding to the pre-execution command.
[0030] Preferably, if the central processing library compares the pre-execution command and it is not included in the execution command stored in the output neuron, then the position of the highest-level judgment neuron that transmits the interactive data signal to the multi-level judgment neuron is found, and there is at least one position of the highest-level judgment neuron where the signal is located, and the one-way flow path connecting the highest-level judgment neuron where the signal is located and the output neuron is deleted, a new output neuron is established in the output layer, and the pre-execution command is stored in the new output neuron as an execution command, and a new one-way flow path is generated between the highest-level judgment neuron where the signal is located and the new output neuron.
[0031] Preferably, the calculation formula of the conditional probability of the interaction data under the condition that the execution information appears is:
[0032] L=Y / X
[0033] Where L is the conditional probability, X is the number of elements in the network event set where the central processing library collects execution information through the network; Y is the number of network events where interactive data appears in the network event set counted by the central processing library;
[0034] The determination rule is to arrange at least one execution information from small to large according to the conditional probability of the interactive data under the condition of the execution information appearing, and take the execution information with the largest conditional probability as the pre-execution command.
[0035] An interactive learning model training system based on machine learning, comprising:
[0036] A model building module, wherein the model building module is used to build a neural network reaction model, wherein the neural network reaction model includes an input layer, a hidden layer, and an output layer;
[0037] An interactive data receiving module, the interactive data receiving module is used to control the neural network reaction model to receive interactive data;
[0038] An optimization module, the optimization module is used to calculate the similarity between the execution command and the keyword of the interaction data, build a matching value calculation model, determine whether the neural network reaction model has defects, and optimize the neural network reaction model;
[0039] A comparison module is used to compare whether the pre-execution command is included in the execution command stored in the output neuron.
[0040] Compared with the prior art, the present invention provides an interactive learning model training method and system based on machine learning, which has the following beneficial effects:
[0041] The present invention sets a central processing library to evaluate the matching degree between execution commands and interaction data, scores them according to the matching degree, optimizes neural network reaction models with low ratings, corrects the unidirectional flow paths of hidden layers and output layers in the neural network reaction models or increases output neurons in the output layer, thereby ensuring that the re-optimized neural network reaction model can return execution commands with high matching degree for the input data, thereby solving the problem that existing optimization methods are difficult to optimize neural network reaction models. BRIEF DESCRIPTION OF THE DRAWINGS
[0042] Figure 1 A schematic diagram of the interactive learning model training method of the present invention;
[0043] Figure 2 A schematic diagram of a method for optimizing a neural network reaction model in the present invention;
[0044] Figure 3 A schematic diagram of a method for comparing a pre-execution command in a central processing library of the present invention to determine whether the pre-execution command is included in the execution command stored in the output neuron. DETAILED DESCRIPTION
[0045] The following description is used to disclose the present invention so that those skilled in the art can implement the present invention. The preferred embodiments described below are only examples, and those skilled in the art may think of other obvious variations.
[0046] Example 1
[0047] Please refer to Figure 1-Figure 3 As shown, an interactive learning model training method based on machine learning includes:
[0048] Establishing a neural network response model, the neural network response model includes an input layer, a hidden layer and an output layer;
[0049] The neural network reaction model receives the interaction data, which is imported into the input layer. The input layer extracts at least one keyword from the interaction data. The input layer compares the keywords stored in the input neurons with the extracted keywords, and selects the keywords stored in the input neurons that are consistent or similar.
[0050] The keywords stored in each selected input neuron transmit signals unidirectionally in the input layer, hidden layer and output layer according to the unidirectional flow path. When the signal is transmitted to the output layer, the execution command stored in the output neuron where the output signal is located, for keywords stored in multiple selected input neurons, they correspond to execution commands stored in different output neurons respectively. For keywords stored in each selected input neuron, there will be at least one set of output neurons corresponding to each keyword. The intersection of the sets of these output neurons is determined, and the execution command of the intersection of the sets of the output neurons is the execution command finally output by the neural network reaction model.
[0051] The central processing library calculates the similarity between the execution command and the keywords of the interaction data, builds a matching value calculation model, determines whether the neural network response model has defects, and optimizes the neural network response model;
[0052] The central processing library collects at least one execution information corresponding to the interactive data through the network, the central processing library collects a network event set of the execution information through the network, calculates the conditional probability of the interactive data under the condition that the execution information appears, and determines the pre-execution command from the at least one execution information according to the determination rule;
[0053] The central processing library compares whether the pre-execution command is included in the execution command stored in the output neuron and takes different execution operations.
[0054] It can be understood by those skilled in the art that, for the revised neural network reaction model, when the interactive data is input again, when its signal is transmitted to the highest-level judgment neuron, the next step will be to transmit it to the output neuron along the unidirectional flow path. However, due to the above steps, the original unidirectional flow path that leads to the output neuron with insufficient matching degree is cut off, so the signal will be transmitted to the output neuron corresponding to the pre-execution command. Since the neural network reaction model ultimately outputs the execution command of the intersection of the set of established output neurons, according to the correction method of the unidirectional flow path, the output neuron corresponding to the pre-execution command will inevitably replace the output neuron that originally appeared in the intersection, and then output the execution command in the output neuron corresponding to the pre-execution command.
[0055] The input layer is provided with a plurality of parallel input neurons, which store keywords; the hidden layer is provided with a plurality of levels of judgment neurons, which store conditional judgment information; the output layer is provided with a plurality of parallel output neurons, which store execution commands; and the input layer, the hidden layer and the output layer are connected by a unidirectional flow path.
[0056] The input neuron and the judgment neuron are connected with a one-way flow path, the multi-level judgment neurons are connected with a one-way flow path, the judgment neuron and the output neuron are connected with a one-way flow path, the signal is transmitted in one direction between the neurons connected with the one-way flow path, the signal of the neuron not connected with the one-way flow path cannot be transmitted, when the signal is transmitted to the output layer, the output neuron where the output layer outputs the signal is stored in the execution command;
[0057] The multi-level judgment neuron is composed of multiple judgment neuron circuits, and a unidirectional flow path is connected between the multiple judgment neurons. The judgment neuron circuit is composed of multiple judgment neurons, and the judgment neurons are arranged from low to high levels. The lowest level judgment neuron is unidirectionally connected to the input neuron, and the highest level input neuron is unidirectionally connected to the output neuron.
[0058] The calculation formula of the central processing library for the similarity between the execution command and the interaction data keyword is:
[0059]
[0060] Where Mi is the matching similarity between the execution command and the i-th keyword in the interaction data, aij is the j-th matching index value of the i-th keyword in the execution command, bij is the j-th matching index value of the i-th keyword in the interaction data, and n is the total number of matching index values.
[0061] The calculation formula of the matching value model is:
[0062]
[0063] Where Y is the matching value between the execution command and the interaction data, αi is the weight value of the similarity of the i-th keyword, and h is the total number of keywords.
[0064] Determining whether the neural network response model has defects and optimizing the neural network response model specifically include the following steps:
[0065] Calculate the matching value through the calculation formula of the matching model;
[0066] Comparing the matching value with a first preset threshold and a second preset threshold;
[0067] If the matching value is greater than 0 and less than the first preset threshold, the matching degree between the execution command and the interaction data is level one;
[0068] If the matching value is greater than or equal to the first preset threshold value, but less than the second preset threshold value, the matching degree between the execution command and the interaction data is level 2;
[0069] If the matching value is greater than or equal to the second preset threshold, the matching degree between the execution command and the interaction data is level three;
[0070] When the matching degree between the execution command and the interaction data is level one, there are defects in the neural network reaction model, and the neural network reaction model is optimized.
[0071] If the central processing library compares the pre-execution command to be included in the execution command stored in the output neuron, the pre-execution command corresponds to the input neuron, and the position of the highest-level judgment neuron that transmits the interactive data signal to the multi-level judgment neuron is found. There is at least one position of the highest-level judgment neuron where the signal is located, and the unidirectional flow path connecting the highest-level judgment neuron where the signal is located and the output neuron is deleted, and a new unidirectional flow path is generated between the highest-level judgment neuron where the signal is located and the output neuron corresponding to the pre-execution command.
[0072] If the central processing library compares the pre-execution command and it is not included in the execution command stored in the output neuron, then find the position of the highest-level judgment neuron that transmits the interactive data signal to the multi-level judgment neuron, and there is at least one position of the highest-level judgment neuron where the signal is located. Delete the one-way flow path connecting the highest-level judgment neuron where the signal is located and the output neuron, establish a new output neuron in the output layer, store the pre-execution command as an execution command in the new output neuron, and generate a new one-way flow path between the highest-level judgment neuron where the signal is located and the new output neuron.
[0073] The calculation formula for the conditional probability of interaction data under the condition of execution information appearance is:
[0074] L=Y / X
[0075] Where L is the conditional probability, X is the number of elements in the network event set where the central processing library collects execution information through the network; Y is the number of network events where interactive data appears in the network event set counted by the central processing library;
[0076] The determination rule is to arrange at least one execution information from small to large according to the conditional probability of the interactive data under the condition of the execution information appearing, and take the execution information with the largest conditional probability as the pre-execution command.
[0077] It can be understood by those skilled in the art that the conditional probability of interaction data under the condition that execution information appears is mainly used to evaluate the correlation between interaction data and execution information. When execution information appears, the higher the conditional probability of interaction data appearing, the higher the correlation between the two. Conversely, the lower the correlation, the lower the correlation. A low correlation means a low match between the execution information and the interaction data. Therefore, when selecting, it is necessary to eliminate execution information with low correlation.
[0078] An interactive learning model training system based on machine learning, comprising:
[0079] A model building module, wherein the model building module is used to build a neural network reaction model, wherein the neural network reaction model includes an input layer, a hidden layer, and an output layer;
[0080] An interactive data receiving module, the interactive data receiving module is used to control the neural network reaction model to receive interactive data;
[0081] An optimization module, the optimization module is used to calculate the similarity between the execution command and the keyword of the interaction data, build a matching value calculation model, determine whether the neural network reaction model has defects, and optimize the neural network reaction model;
[0082] A comparison module is used to compare whether the pre-execution command is included in the execution command stored in the output neuron.
[0083] To summarize, by setting up a central processing library to evaluate the matching degree between execution commands and interaction data, scoring is performed based on the matching degree, optimizing the neural network reaction models with low ratings, correcting the unidirectional flow paths of the hidden layer and the output layer in the neural network reaction model, or increasing the output neurons in the output layer, thereby ensuring that the re-optimized neural network reaction model can return execution commands with high matching degree for the input data, thereby solving the problem that the existing optimization methods are difficult to optimize the neural network reaction models.
[0084] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions only describe the principles of the present invention. The present invention may be subject to various changes and improvements without departing from the spirit and scope of the present invention. These changes and improvements fall within the scope of the present invention. The scope of protection claimed by the present invention is defined by the attached claims and their equivalents.
Claims
1. An interactive learning model training method based on machine learning, characterized in that: include: Establishing a neural network response model, the neural network response model includes an input layer, a hidden layer and an output layer; The neural network reaction model receives the interaction data, which is imported into the input layer. The input layer extracts at least one keyword from the interaction data. The input layer compares the keywords stored in the input neurons with the extracted keywords, and selects the keywords stored in the input neurons that are consistent or similar. The keywords stored in each selected input neuron transmit signals unidirectionally in the input layer, the hidden layer and the output layer according to the unidirectional flow path. When the signal is transmitted to the output layer, the execution command stored in the output neuron where the output signal is located, for keywords stored in multiple selected input neurons, they correspond to execution commands stored in different output neurons respectively. For keywords stored in each selected input neuron, there is a set consisting of at least one output neuron corresponding to each keyword stored in the selected input neuron. The intersection of the sets of these output neurons is determined, and the execution command of the intersection of the sets of the output neurons is the execution command finally output by the neural network reaction model; The central processing library calculates the similarity between the execution command and the keywords of the interaction data, builds a matching value calculation model, determines whether the neural network response model has defects, and optimizes the neural network response model; The central processing library collects at least one execution information corresponding to the interactive data through the network, the central processing library collects a network event set of the execution information through the network, calculates the conditional probability of the interactive data under the condition that the execution information appears, and determines the pre-execution command from the at least one execution information according to the determination rule; The central processing library compares whether the pre-execution command is included in the execution command stored in the output neuron and takes different execution operations.
2. The interactive learning model training method based on machine learning according to claim 1, characterized in that: The input layer is provided with a plurality of parallel input neurons, in which keywords are stored; the hidden layer is provided with a plurality of levels of judgment neurons, in which conditional judgment information is stored; the output layer is provided with a plurality of parallel output neurons, in which execution commands are stored; and the input layer, the hidden layer and the output layer are connected by a unidirectional flow path.
3. The interactive learning model training method based on machine learning according to claim 2, characterized in that: The input neuron and the judgment neuron are connected with a unidirectional flow path, the multi-level judgment neurons are connected with a unidirectional flow path, the judgment neuron and the output neuron are connected with a unidirectional flow path, the signal is transmitted in one direction between the neurons connected with the unidirectional flow path, the signal of the neuron not connected with the unidirectional flow path cannot be transmitted, and when the signal is transmitted to the output layer, the output layer outputs the execution command stored in the output neuron where the signal is located; The multi-level judgment neurons are composed of multiple judgment neuron circuits, and a unidirectional flow path is connected between the multiple judgment neurons. The judgment neuron circuits are composed of multiple judgment neurons, and the judgment neurons are arranged from low to high levels. The lowest-level judgment neurons are unidirectionally connected to the input neurons, and the highest-level input neurons are unidirectionally connected to the output neurons.
4. The interactive learning model training method based on machine learning according to claim 3, characterized in that: The calculation formula of the central processing library for the similarity between the execution command and the interaction data keyword is: Where Mi is the matching similarity between the execution command and the i-th keyword in the interaction data, aij is the j-th matching index value of the i-th keyword in the execution command, bij is the j-th matching index value of the i-th keyword in the interaction data, and n is the total number of matching index values.
5. The interactive learning model training method based on machine learning according to claim 4, characterized in that: The calculation formula of the matching value model is: Where Y is the matching value between the execution command and the interaction data, αi is the weight value of the similarity of the i-th keyword, and h is the total number of keywords.
6. The interactive learning model training method based on machine learning according to claim 5, characterized in that: The determining whether the neural network reaction model has defects and optimizing the neural network reaction model specifically comprises the following steps: Calculate the matching value through the calculation formula of the matching model; Comparing the matching value with a first preset threshold and a second preset threshold; If the matching value is greater than 0 and less than the first preset threshold, the matching degree between the execution command and the interaction data is level one; If the matching value is greater than or equal to the first preset threshold value, but less than the second preset threshold value, the matching degree between the execution command and the interaction data is level 2; If the matching value is greater than or equal to the second preset threshold, the matching degree between the execution command and the interaction data is level three; When the matching degree between the execution command and the interaction data is level one, there are defects in the neural network reaction model, and the neural network reaction model is optimized.
7. The interactive learning model training method based on machine learning according to claim 6, characterized in that: If the central processing library compares the pre-execution command to be included in the execution command stored in the output neuron, the pre-execution command corresponds to the input neuron, and the position of the highest-level judgment neuron that transmits the interactive data signal to the multi-level judgment neuron is found. There is at least one position of the highest-level judgment neuron where the signal is located, and the unidirectional flow path connecting the highest-level judgment neuron where the signal is located and the output neuron is deleted, and a new unidirectional flow path is generated between the highest-level judgment neuron where the signal is located and the output neuron corresponding to the pre-execution command.
8. The interactive learning model training method based on machine learning according to claim 7, characterized in that: If the central processing library compares the pre-execution command and it is not included in the execution command stored in the output neuron, then find the position of the highest-level judgment neuron that transmits the interactive data signal to the multi-level judgment neuron, and there is at least one position of the highest-level judgment neuron where the signal is located. Delete the one-way flow path connecting the highest-level judgment neuron where the signal is located and the output neuron, establish a new output neuron in the output layer, store the pre-execution command as an execution command in the new output neuron, and generate a new one-way flow path between the highest-level judgment neuron where the signal is located and the new output neuron.
9. The interactive learning model training method based on machine learning according to claim 8, characterized in that: The calculation formula of the conditional probability of the interaction data under the condition of the occurrence of the execution information is: L=Y / X Where L is the conditional probability, X is the number of elements in the network event set where the central processing library collects execution information through the network; Y is the number of network events where interactive data appears in the network event set counted by the central processing library; The determination rule is to arrange at least one execution information from small to large according to the conditional probability of the interactive data under the condition of the execution information appearing, and take the execution information with the largest conditional probability as the pre-execution command.
10. An interactive learning model training system based on machine learning, used to implement an interactive learning model training method based on machine learning as described in any one of claims 1 to 9, characterized in that: include: A model building module, wherein the model building module is used to build a neural network reaction model, wherein the neural network reaction model includes an input layer, a hidden layer, and an output layer; An interactive data receiving module, the interactive data receiving module is used to control the neural network reaction model to receive interactive data; An optimization module, the optimization module is used to calculate the similarity between the execution command and the keyword of the interaction data, build a matching value calculation model, determine whether the neural network reaction model has defects, and optimize the neural network reaction model; A comparison module is used to compare whether the pre-execution command is included in the execution command stored in the output neuron.