Decision-making method and system for multi-expert hybrid model, storage medium and program product

Abstract

The invention provides a decision-making method and system for a multi-expert hybrid model, a storage medium and a program product, and belongs to the technical field of computers. The method comprises the following steps: constructing a plurality of parameter learnable expert models; constructing a global model with learnable parameters; obtaining a multi-modal task data set, and respectively training each expert model and the global model to obtain outputs of the global model and each expert model; according to the global model and the output of each expert model, utilizing an expert model selection mechanism to obtain the decision weight of the multi-expert hybrid model; and multiplying the decision-making right key point by the output of each expert model, and obtaining a model decision-making result through a weighted summation mode. According to the method, the process of the reasoning stage of the hybrid expert model is simplified, the complexity and parameter quantity of the model are reduced, and the whole model is more efficient.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular to a decision-making method, system, storage medium and program product of a multi-expert hybrid model. Background Art

[0002] In application scenarios such as multimodal dialogue systems and autonomous driving cabins, the expert mixture model is a complex problem-solving method based on knowledge from multiple fields. Its core concept is to combine multiple functional customized sub-models (i.e., "experts"), with each "expert" contributing within its specific function. When dealing with problems, different tasks are first separated and then distributed to "experts" who are good at related fields to handle. Finally, the "gated network" mechanism is used to decide which "expert" will participate in answering specific questions, and the conclusions of each "expert" are summarized. In this way, the expert mixture model provides an effective way to solve complex problems. The advantage of this method is that it can make full use of the expertise in various fields and give full play to the potential of each "expert" in solving complex cross-domain problems.

[0003] However, the development of traditional expert mixture models also faces difficulties, among which the model selection problem is a major difficulty. At present, the model selection method involved in answering specific questions is a mechanism called "gated network". This mechanism outputs a probability density vector, integrates the outputs of various expert models, and obtains the model decision result.

[0004] However, the "gated network" is usually obtained by neural network modeling, which makes the process of the expert mixture model reasoning stage cumbersome, increases the computational complexity and parameter quantity of the model, reduces efficiency, increases resource consumption, and increases user waiting time.

[0005] Therefore, there is an urgent need for a parameter-free modeling method to simplify the process of the hybrid expert model reasoning stage, reduce the complexity and number of parameters of the model, make the overall model more efficient, reduce resource consumption, and reduce user waiting time, in order to solve the above problems. Summary of the invention

[0006] In view of the problems existing in the above-mentioned prior art, the present invention proposes a decision-making method, system, storage medium and program product of a multi-expert hybrid model, which can simplify the process of the hybrid expert model reasoning stage, omit the reasoning process of the gating model, and reduce the complexity and parameter amount of the model.

[0007] In a first aspect, the present invention provides a decision-making method of a multi-expert hybrid model, comprising the following steps:

[0008] Build expert models with multiple learnable parameters;

[0009] Build a global model with learnable parameters;

[0010] Obtain a multimodal task dataset, train each expert model and the global model separately, and obtain the output of the global model and each expert model;

[0011] According to the output of the global model and each expert model, the decision weight of the multi-expert hybrid model is obtained by using the expert model selection mechanism;

[0012] The decision weight is multiplied by the output of each expert model, and the model decision result is obtained by weighted summation.

[0013] As a further improvement of the present invention, each of the expert models is implemented by a neural network, and the parameter quantity of each of the expert models remains consistent.

[0014] As a further improvement of the present invention, the global model is implemented by a neural network, and the number of parameters of the global model is greater than the number of parameters of the expert model.

[0015] As a further improvement of the present invention, the step of acquiring a multimodal task data set and training each expert model and a global model respectively includes:

[0016] Obtain multimodal task datasets;

[0017] Based on the multimodal task dataset and loss function, each expert model and the global model are trained separately.

[0018] As a further improvement of the present invention, the output of the global model is an encoding vector of the global model;

[0019] The output of each expert model is the last hidden layer vector of each expert model.

[0020] As a further improvement of the present invention, the decision weights of the multi-expert hybrid model are obtained by using the expert model selection mechanism according to the outputs of the global model and each expert model, including:

[0021] Construct a parameterized neural network;

[0022] Input the last hidden layer vector of each expert model, and train the parameterized neural network based on the multimodal task dataset and loss function;

[0023] The last hidden layer vector of each expert model, the encoding vector of the global model, the parameterized neural network and the loss function are taken as input;

[0024] Using the loss function, the encoding vector of the expert model encoded by the neural network and the encoding vector of the global model are calculated to form a set:

[0025]

[0026] Where f represents a set, represents the loss function, Z r represents the last hidden layer vector of each expert model, z represents the last hidden layer vector of the expert model, z G Represents the encoding vector of the global model, g φ represents a parameterized neural network;

[0027] Find the gradient of the elements in the set and form a matrix:

[0028]

[0029] Among them, U represents a matrix, and f represents a set;

[0030] Solve for the Lagrange multipliers:

[0031]

[0032] Among them, λ represents the Lagrange multiplier, the sum of λ is 1, λ i is the element in the vector λ, U represents the matrix, and f represents the set;

[0033] Get the decision weights of the multi-expert mixture model.

[0034] As a further improvement of the present invention, the expert model selection mechanism is an adaptive algorithm.

[0035] As a further improvement of the present invention, the decision weight is multiplied by the output of each expert model to obtain the model decision result by weighted summation, including:

[0036] Through the decision weights of the multi-expert hybrid model and the last hidden layer vectors of each expert model, the final model output result is obtained in the model inference stage:

[0037] Out 模型决策 =g φ (λ T Z r )

[0038] Among them, Out 模型决策 Represents the final model output result, g φ represents a parameterized neural network, λ represents the Lagrange multiplier, Z r Represents the last hidden layer vector of each expert model.

[0039] In a second aspect, the present invention provides a computer system, which is used to implement the steps of the method described in the first aspect, including an expert model construction module, a global model construction module, a model training module, a decision weight acquisition module and a decision result acquisition module;

[0040] The expert model building module is used to build an expert model with multiple learnable parameters;

[0041] The global model building module is used to build a global model with learnable parameters;

[0042] The model training module is used to obtain a multimodal task data set, train each expert model and the global model respectively, and obtain the output of the global model and each expert model;

[0043] The decision weight acquisition module is used to obtain the decision weight of the multi-expert hybrid model based on the output of the global model and each expert model using the expert model selection mechanism;

[0044] The decision result obtaining module is used to multiply the decision weight by the output of each expert model, and obtain the model decision result by weighted summation.

[0045] In a third aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described in the first aspect.

[0046] In a fourth aspect, the present invention provides a computer program product, comprising a computer program, which implements the steps of the method described in the first aspect when executed by a processor.

[0047] Compared with the prior art, the present invention has the following beneficial effects:

[0048] The present invention proposes a decision-making method, system, storage medium and program product of a multi-expert hybrid model, which simplifies the process of the hybrid expert model reasoning stage and reduces the complexity and parameter amount of the model. The present invention omits the reasoning process of the gated model, making the overall model more efficient. The simplified design helps to improve the reasoning speed of the model and reduce resource consumption on multimodal complex tasks. Especially in the interactive platform of the intelligent assistant, the non-gated expert hybrid model constructed by the present invention can realize the rapid output of the decision information of the intelligent assistant and reduce the user waiting time. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 A flowchart of a decision-making method of a multi-expert hybrid model disclosed in an embodiment of the present invention;

[0050] Figure 2A schematic diagram of a decision-making method of a multi-expert hybrid model disclosed in an embodiment of the present invention;

[0051] Figure 3 A framework diagram of a decision-making system of a multi-expert hybrid model disclosed in an embodiment of the present invention. DETAILED DESCRIPTION

[0052] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the present invention will combine the accompanying drawings to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments, wherein steps S1, S2... in the embodiments described in the present invention do not limit the only execution steps of the present invention; wherein the various models, mechanisms or functions described in the present invention are not the only limiting methods of the present invention. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0053] In the present invention, "module", "system", etc. refer to related entities applied to computers, such as hardware, a combination of hardware and software, software or software in execution, etc. In detail, for example, software includes but is not limited to a process running on a processor, a processor, an object, executable software, an execution thread, a program and / or a computer. In addition, an application or script program running on a server, and a server can all be software. One or more software can be in an execution process and / or thread, and the software can be localized on one computer and / or distributed between two or more computers, and can be run by various computer-readable media.

[0054] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.

[0055] In the first aspect of the embodiment of the present invention, a decision-making method of a multi-expert hybrid model is provided. Figure 1 and Figure 2 As shown, the following steps are included:

[0056] S1: Build an expert model with multiple learnable parameters;

[0057] Preferably, each of the expert models can be implemented by a neural network, and the parameter quantity of each of the expert models remains consistent.

[0058] In one embodiment of the present invention, N expert models are constructed The number of parameters of each expert model remains consistent.

[0059] S2: Build a global model with learnable parameters;

[0060] Preferably, the global model can be implemented by a neural network, and the number of parameters of the global model is greater than the number of parameters of the expert model.

[0061] In one embodiment of the present invention, N expert models are constructed Among them, the parameters of each expert model are consistent; build a global model Among them, the number of parameters of the global model is larger than that of the expert model.

[0062] Preferably, the number of parameters of the global model may be slightly larger than the number of parameters of the expert model. Obviously, the fact that the number of parameters of the global model described in the present invention is slightly larger than the number of parameters of the expert model is not the only way to limit the number of parameters of the global model and the expert model in the present invention. Technical solutions that can be obtained by those skilled in the art without creative labor on the basis of the present invention belong to the protection scope of the present invention.

[0063] S3: Obtain a multimodal task dataset, train each expert model and the global model respectively, and obtain the output of the global model and each expert model;

[0064] Preferably, the step of acquiring a multimodal task dataset and training each expert model and the global model separately may include:

[0065] S31: Obtain multimodal task dataset;

[0066] S32: Based on the multimodal task dataset and loss function, train each expert model and the global model separately.

[0067] Preferably, the output of the global model is an encoding vector of the global model;

[0068] The output of each expert model is the last hidden layer vector of each expert model.

[0069] In one embodiment of the present invention, N expert models are constructed Among them, the parameters of each expert model are consistent; build a global model Among them, the number of parameters of the global model is larger than that of the expert model. Get a multimodal task dataset D, based on the dataset D and the loss function Train each expert model separately and global model After the training is completed, the last hidden layer vector Z of each expert model is obtained r ={z1,z2,…,z N} and the encoding vector z of the global model G .

[0070] S4: Based on the outputs of the global model and each expert model, the decision weights of the multi-expert hybrid model are obtained using the expert model selection mechanism;

[0071] Preferably, the step of obtaining the decision weights of the multi-expert hybrid model by using an expert model selection mechanism based on the outputs of the global model and each expert model may include:

[0072] S41: Construct a parameterized neural network;

[0073] S42: Input the last hidden layer vector of each expert model, and train the parameterized neural network based on the multimodal task dataset and loss function;

[0074] S43: taking the last hidden layer vector of each expert model, the encoding vector of the global model, the parameterized neural network and the loss function as input;

[0075] S44: Using the loss function, the encoding vector of the expert model encoded by the neural network and the encoding vector of the global model are calculated to form a set:

[0076]

[0077] Where f represents a set, represents the loss function, Z r represents the last hidden layer vector of each expert model, z represents the last hidden layer vector of the expert model, z G Represents the encoding vector of the global model, g φ represents a parameterized neural network;

[0078] S45: Find the gradient of the elements in the set and form a matrix:

[0079]

[0080] Among them, U represents a matrix, and f represents a set;

[0081] S46: Solve for Lagrange multipliers:

[0082]

[0083] Among them, λ represents the Lagrange multiplier, the sum of λ is 1, λ i is the element in the vector λ, U represents the matrix, and f represents the set;

[0084] S47: Obtain the decision weight of the multi-expert hybrid model.

[0085] Preferably, the expert model selection mechanism is an adaptive algorithm.

[0086] In one embodiment of the present invention, N expert models are constructed Among them, the parameters of each expert model are consistent; build a global model Among them, the number of parameters of the global model is larger than that of the expert model. Get a multimodal task dataset D, based on the dataset D and the loss function Train each expert model separately and global model After the training is completed, the last hidden layer vector Z of each expert model is obtained r ={z1,z2,…,z N} and the encoding vector z of the global model G ;

[0087] Next, construct a parameterized neural network g φ ; Input the last hidden layer vector Z of each expert model r ={z1,z2,…,z N}, through the data set D and loss function Train the parameterized neural network g φ ;

[0088] The algorithm of adaptively selecting expert models is used as the algorithm for simulating the workspace mechanism, and the last hidden layer vector Z of each expert model is r ={z1,z2,…,z N}、The encoding vector z of the global model G , parameterized neural network g φ And the loss function As input; then do the following calculation:

[0089] Using loss function Calculation through neural network g φ The encoding vector of the encoded expert model and the encoding vector of the global model form a set f:

[0090]

[0091] Where f represents a set, represents the loss function, Z r represents the last hidden layer vector of each expert model, z represents the last hidden layer vector of the expert model, z G Represents the encoding vector of the global model, g φ represents a parameterized neural network.

[0092] Find the gradient of the elements in the set f and form the matrix U:

[0093]

[0094] Among them, U represents a matrix, and f represents a set;

[0095] Solve for the Lagrange multiplier λ:

[0096]

[0097] Among them, λ represents the Lagrange multiplier, the sum of λ is 1, λ i is the element in the vector λ, U represents the matrix, and f represents the set;

[0098] λ can be used as the decision weight of the multi-expert mixture model.

[0099] S5: Multiply the decision weight by the output of each expert model and obtain the model decision result through weighted summation.

[0100] Preferably, the step of multiplying the decision weight by the output of each expert model and obtaining the model decision result by weighted summation includes:

[0101] Through the decision weights of the multi-expert mixture model and the last hidden layer vector Z of each expert model r , in the model inference phase, the final model output result is obtained:

[0102] Out 模型决策 =g φ (λ T Z r )

[0103] Among them, Out 模型决策 Represents the final model output result, g φ represents a parameterized neural network, λ represents the Lagrange multiplier, Z r Represents the last hidden layer vector of each expert model.

[0104] In the embodiment of the present invention, Figure 3 As shown, in a second aspect, a computer system is provided, wherein the computer system is used to implement the steps of the method in the first aspect, including an expert model construction module, a global model construction module, a model training module, a decision weight acquisition module and a decision result acquisition module;

[0105] The expert model building module is used to build an expert model with multiple learnable parameters;

[0106] The global model building module is used to build a global model with learnable parameters;

[0107] The model training module is used to obtain a multimodal task data set, train each expert model and the global model respectively, and obtain the output of the global model and each expert model;

[0108] The decision weight acquisition module is used to obtain the decision weight of the multi-expert hybrid model based on the output of the global model and each expert model using the expert model selection mechanism;

[0109] The decision result obtaining module is used to multiply the decision weight by the output of each expert model, and obtain the model decision result by weighted summation.

[0110] In a third aspect, a computer-readable storage medium is provided in an embodiment of the present invention, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the method described in the first aspect are implemented.

[0111] In a fourth aspect, a computer program product is provided in an embodiment of the present invention, including a computer program, which implements the steps of the method described in the first aspect when executed by a processor.

[0112] The present invention proposes a decision-making method, system, storage medium and program product of a multi-expert hybrid model, which simplifies the process of the hybrid expert model reasoning stage and reduces the complexity and parameter amount of the model. The present invention omits the reasoning process of the gated model, making the overall model more efficient. The simplified design helps to improve the reasoning speed of the model and reduce resource consumption on multimodal complex tasks. Especially in the interactive platform of the intelligent assistant, the non-gated expert hybrid model constructed by the present invention can realize the rapid output of the decision information of the intelligent assistant and reduce the user waiting time.

Claims

1. A decision-making method of a multi-expert mixed model, characterized in that: The following steps are involved: Build expert models with multiple learnable parameters; Build a global model with learnable parameters; Obtain a multimodal task dataset, train each expert model and the global model separately, and obtain the output of the global model and each expert model; According to the output of the global model and each expert model, the decision weight of the multi-expert hybrid model is obtained by using the expert model selection mechanism; The decision weight is multiplied by the output of each expert model, and the model decision result is obtained by weighted summation.

2. The decision-making method of the multi-expert mixed model according to claim 1, characterized in that: Each of the expert models is implemented by a neural network, and the parameter quantity of each of the expert models remains consistent.

3. The decision-making method of the multi-expert mixed model according to claim 1 is characterized in that: The global model is implemented by a neural network, and the number of parameters of the global model is greater than the number of parameters of the expert model.

4. The decision-making method of the multi-expert hybrid model according to claim 1, characterized in that: The multimodal task dataset is obtained, and each expert model and the global model are trained separately, including: Obtain multimodal task datasets; Based on the multimodal task dataset and loss function, each expert model and the global model are trained separately.

5. The decision-making method of the multi-expert mixed model according to claim 1 or 4, characterized in that: The output of the global model is an encoding vector of the global model; The output of each expert model is the last hidden layer vector of each expert model.

6. The decision-making method of the multi-expert mixed model according to claim 1, characterized in that: The decision weights of the multi-expert hybrid model are obtained by using the expert model selection mechanism according to the outputs of the global model and each expert model, including: Construct a parameterized neural network; Input the last hidden layer vector of each expert model, and train the parameterized neural network based on the multimodal task dataset and loss function; The last hidden layer vector of each expert model, the encoding vector of the global model, the parameterized neural network and the loss function are taken as input; Using the loss function, the encoding vector of the expert model encoded by the neural network and the encoding vector of the global model are calculated to form a set: Where f represents a set, represents the loss function, Z r represents the last hidden layer vector of each expert model, z represents the last hidden layer vector of the expert model, z G Represents the encoding vector of the global model, g φ represents a parameterized neural network; Find the gradient of the elements in the set and form a matrix: Among them, U represents a matrix, and f represents a set; Solve for the Lagrange multipliers: Among them, λ represents the Lagrange multiplier, the sum of λ is 1, λ i is the element in the vector λ, U represents the matrix, and f represents the set; Get the decision weights of the multi-expert mixture model.

7. The decision-making method of the multi-expert mixed model according to claim 1 or 6, characterized in that: The expert model selection mechanism is an adaptive algorithm.

8. The decision-making method of the multi-expert hybrid model according to claim 1, characterized in that: The decision weight is multiplied by the output of each expert model, and the model decision result is obtained by weighted summation, including: Through the decision weights of the multi-expert hybrid model and the last hidden layer vectors of each expert model, the final model output result is obtained in the model inference stage: Out 模型决策 =g φ (λ T Z r ) Among them, Out 模型决策 Represents the final model output result, g φ represents a parameterized neural network, λ represents the Lagrange multiplier, Z r Represents the last hidden layer vector of each expert model.

9. A computer system, characterized in that: The computer system is used to implement the steps of the method described in any one of claims 1 to 8, comprising an expert model construction module, a global model construction module, a model training module, a decision weight acquisition module and a decision result acquisition module; The expert model building module is used to build an expert model with multiple learnable parameters; The global model building module is used to build a global model with learnable parameters; The model training module is used to obtain a multimodal task data set, train each expert model and the global model respectively, and obtain the output of the global model and each expert model; The decision weight acquisition module is used to obtain the decision weight of the multi-expert hybrid model based on the output of the global model and each expert model using the expert model selection mechanism; The decision result obtaining module is used to multiply the decision weight by the output of each expert model, and obtain the model decision result by weighted summation.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 8 are implemented.

11. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 8 are implemented.

Images