A multi-objective learning method, system, storage medium and terminal device

Abstract

The embodiment of the present invention discloses a multi-objective learning method, system, storage medium and terminal device, which are applied to the field of information processing technology based on artificial intelligence. The multi-objective learning system will determine that the multi-objective model includes a shared underlying sub-network, multiple underlying task sub-networks, multiple task feature modules and multiple task output modules, and determine that the training sample includes multiple sample information. The overall loss function related to the multi-objective initial model calculated based on the multi-objective initial model and the training sample includes: the first related information between the input features and output features of the shared underlying sub-network based on the sample information, and the second related information between the output features of the multiple underlying task sub-networks based on the sample information and the output features of the shared underlying sub-network, and then adjust the multi-objective initial model according to the overall loss function. The effect of the multi-objective model learned in the embodiment of the present invention is greatly improved.

Description

Technical Field

[0001] The present invention relates to the field of information processing technology based on artificial intelligence, and in particular to a multi-objective learning method, system, storage medium and terminal device. Background Art

[0002] Most machine learning models learn independently, a practice known as single-task learning. This involves designing a model for a specific task and then iteratively optimizing the model. For more complex tasks, the task can be broken down into its subtasks, which can be modeled individually. However, when modeling each subtask, it's easy to overlook relationships, conflicts, and constraints between tasks, resulting in a poor overall performance across multiple subtasks.

[0003] An existing multi-task learning method can achieve good results in learning complex tasks. Specifically, multi-task learning mainly combines multiple related tasks for learning. For example, in a recommendation system, the ranking model simultaneously estimates the click-through rate and browse time of a candidate. There are many existing models based on multi-task learning, such as the Multi-gate Mixture-of-Experts (MMoE), Customized Gate Control (CGC), and Progressive Layered Extraction (PLE). These models involve the learning of shared modules during the multi-task learning process, which leads to information redundancy and makes the learned multi-task model less effective. Summary of the Invention

[0004] The embodiments of the present invention provide a multi-objective learning method, system, storage medium and terminal device, which improve the performance of the trained multi-objective model.

[0005] An embodiment of the present invention provides a multi-objective learning method, including:

[0006] Determine a multi-objective initial model, wherein the multi-objective initial model includes: a shared underlying subnetwork, multiple underlying task subnetworks, multiple task feature modules, and multiple task output modules;

[0007] Determine a training sample, wherein the training sample includes a plurality of sample information;

[0008] Calculate an overall loss function associated with the multi-objective initial model based on the multi-objective initial model and the training samples; the overall loss function includes: first correlation information between the input features and output features of the shared underlying sub-network based on the sample information, and second correlation information between the output features of the multiple underlying task sub-networks based on the sample information and the output features of the shared underlying sub-network;

[0009] The multi-objective initial model is adjusted according to the overall loss function to obtain a final multi-objective model, where the multi-objective model is used to obtain multiple task execution results based on the data to be processed.

[0010] Another embodiment of the present invention provides a multi-objective learning system, including:

[0011] A model determination unit is used to determine a multi-objective initial model, wherein the multi-objective initial model includes: a shared underlying subnetwork, multiple underlying task subnetworks, multiple task feature modules, and multiple task output modules;

[0012] A sample determination unit, configured to determine a training sample, wherein the training sample includes a plurality of sample information;

[0013] A loss function unit is configured to calculate an overall loss function associated with the multi-objective initial model based on the multi-objective initial model and the training samples; the overall loss function includes: first correlation information between the input features and output features of the shared underlying sub-network based on the sample information, and second correlation information between the output features of the multiple underlying task sub-networks based on the sample information and the output features of the shared underlying sub-network;

[0014] An adjustment unit is used to adjust the multi-objective initial model according to the overall loss function to obtain a final multi-objective model, wherein the multi-objective model is used to obtain multiple task execution results based on the data to be processed.

[0015] Another aspect of an embodiment of the present invention further provides a computer-readable storage medium, which stores a plurality of computer programs, and the computer programs are suitable for being loaded by a processor and executing the multi-objective learning method as described in one aspect of an embodiment of the present invention.

[0016] Another aspect of the present invention provides a terminal device, including a processor and a memory;

[0017] The memory is used to store multiple computer programs, and the computer programs are used to be loaded by the processor and execute the multi-objective learning method as described in one aspect of an embodiment of the present invention; the processor is used to implement each computer program in the multiple computer programs.

[0018] It can be seen that in the training process of the multi-objective model in this embodiment, it is determined that the multi-objective model includes a shared underlying sub-network, multiple underlying task sub-networks, multiple task feature modules and multiple task output modules, and it is determined that the training sample includes multiple sample information. The overall loss function related to the multi-objective initial model calculated based on the multi-objective initial model and the training sample includes: the first related information between the input features and output features of the shared underlying sub-network based on the sample information, and the second related information between the output features of the multiple underlying task sub-networks based on the sample information and the output features of the shared underlying sub-network, and then the multi-objective initial model is adjusted according to the overall loss function. Since the learning of the shared module, i.e., the shared underlying sub-network, is involved in the training process of the multi-objective model, the first related information and the second related information related to the shared underlying sub-network in the overall loss function can well guide the training process of the multi-objective model, so that information redundancy will not be caused, thereby improving the effect of the learned multi-objective model. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0020] Figure 1 is a schematic diagram of a multi-objective learning method provided by an embodiment of the present invention;

[0021] Figure 2 is a flowchart of a multi-objective learning method provided by one embodiment of the present invention;

[0022] Figure 3 is a schematic diagram of a multi-objective initial model determined in one embodiment of the present invention;

[0023] Figure 4 is a flow chart of a method for calculating an overall loss function in one embodiment of the present invention;

[0024] Figure 5 This is a flowchart of a multi-objective learning method provided by an application embodiment of the present invention;

[0025] Figure 6 is a schematic diagram of a multi-objective initial model determined in an application embodiment of the present invention;

[0026] Figure 7 is a schematic diagram of a distributed system to which a multi-objective learning method is applied in another application embodiment of the present invention;

[0027] Figure 8 is a schematic diagram of a block structure in another application embodiment of the present invention;

[0028] Figure 9 1 is a schematic diagram of the logical structure of a multi-objective learning system provided by an embodiment of the present invention;

[0029] Figure 10 This is a schematic diagram of the logical structure of a terminal device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0030] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0031] The terms "first", "second", "third", "fourth", etc. (if any) in the description and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the numbers used in this way can be interchanged where appropriate, so that the embodiments of the invention described herein can, for example, be implemented in orders other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or apparatus that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or apparatus.

[0032] The embodiment of the present invention provides a multi-objective learning method, which can mainly train a multi-objective model. The multi-objective model is a composite task execution model. Figure 1 As shown in the figure, the multi-objective learning system can implement the training of the multi-objective model in the following way:

[0033] Determine a multi-objective initial model, wherein the multi-objective initial model includes: a shared underlying subnetwork, multiple underlying task subnetworks, multiple task feature modules, and multiple task output modules;

[0034] Determine a training sample, wherein the training sample includes a plurality of sample information;

[0035] Calculate an overall loss function associated with the multi-objective initial model based on the multi-objective initial model and the training samples; the overall loss function includes: first correlation information between the input features and output features of the shared underlying sub-network based on the sample information, and second correlation information between the output features of the multiple underlying task sub-networks based on the sample information and the output features of the shared underlying sub-network;

[0036] The multi-objective initial model is adjusted according to the overall loss function to obtain a final multi-objective model, where the multi-objective model is used to obtain multiple task execution results based on the data to be processed.

[0037] The multi-objective model described above is a machine learning model based on artificial intelligence (AI), which can be trained using specific training methods. Artificial intelligence (AI) refers to the theories, methods, techniques, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. In other words, AI is a comprehensive technology within computer science that seeks to understand the essence of intelligence and produce new intelligent machines that can respond in a manner similar to human intelligence. AI is the study of the design principles and implementation methods of various intelligent machines, enabling them to possess the capabilities of perception, reasoning, and decision-making.

[0038] Artificial intelligence (AI) technology is a comprehensive discipline encompassing a wide range of fields, encompassing both hardware and software technologies. Foundational AI technologies generally include sensors, specialized AI chips, cloud computing, distributed storage, big data processing, operating / interaction systems, and mechatronics. AI software technologies primarily encompass computer vision, speech processing, natural language processing, and machine learning / deep learning.

[0039] Machine learning (ML) is a multidisciplinary field that encompasses probability theory, statistics, approximation theory, convex analysis, and algorithmic complexity theory. It specifically studies how computers can simulate or implement human learning behaviors to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is at the core of artificial intelligence and the fundamental way to make computers intelligent. Its applications span all areas of AI. Machine learning and deep learning typically include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and learning through demonstration.

[0040] In this way, since the training process of the multi-objective model involves the learning of the shared module, that is, the shared underlying sub-network, the first related information and the second related information related to the shared underlying sub-network in the overall loss function can well guide the training process of the multi-objective model, so that information redundancy will not be caused, and the effect of the learned multi-objective model can be better improved.

[0041] The embodiment of the present invention provides a multi-objective learning method, which is mainly performed by the multi-objective learning system described above. The flow chart is as follows: Figure 2 Shown, including:

[0042] Step 101: determine a multi-objective initial model, which includes: a shared underlying sub-network, multiple underlying task sub-networks, multiple task feature modules, and multiple task output modules.

[0043] It can be understood that when determining the structure of the multi-objective initial model, the multi-objective learning system can determine the multi-layer structure included in the multi-objective initial model and the initial values of the parameters in each layer of the structure.

[0044] Specifically, the multi-objective learning system can determine the multi-objective initial model with the following structure, such as Figure 3 As shown, the multi-objective initial model may include: a shared underlying subnetwork 10, multiple underlying task subnetworks 11, multiple task feature modules 12, and multiple task output modules 13. The number of underlying task subnetworks 11, task feature modules 12, and task output modules 13 is the same, wherein:

[0045] The inputs of the shared underlying sub-network 10 and the multiple underlying task sub-networks 11 are the same, which are all input features of the sample information. The output of the shared underlying sub-network 10 is connected to one input of any task feature module 12, and the output of any underlying task sub-network 11 is connected to the input of one task feature module 12. The output of any task feature module 12 is connected to the input of one task output module 13. Specifically:

[0046] The shared underlying subnetwork 10 is used to obtain shared feature information for sample information, and the any underlying task subnetwork 11 is used to obtain feature information for sample information. This feature information is primarily used for executing a task. The any task feature module 12 is used to fuse the feature information obtained by the any underlying task subnetwork 11 with the shared feature information obtained by the shared underlying subnetwork 10 to obtain feature information suitable for executing a task. The any task output module 13 is used to execute the corresponding task based on the feature information obtained by the any task feature module 12 and obtain the task execution result. A task refers to a specific operation, such as classification or prediction, and can be a task in any application scenario, such as predicting the click-through rate and viewing time of news.

[0047] Furthermore, the multi-objective initial model also includes a shared low-order feature module 14 and multiple low-order task modules 15. The inputs of the shared low-order feature module 14 and the low-order task module 15 are the same as the inputs of the shared underlying sub-network 10, which are all input features of sample information (not shown in the figure). The output of the shared low-order feature module 14 is connected to any task output module 13, and the output of any low-order task module 15 is connected to one task output module 13. Among them:

[0048] The shared low-order feature module 14 is used to obtain low-order shared features of the sample information, and the low-order task module 15 is used to obtain low-order task features of the sample information, so that the above-mentioned task output module 13 will execute the corresponding task according to the feature information obtained by the task feature module 12, the low-order task features obtained by the low-order task module 15, and the low-order shared features obtained by the shared low-order feature module 14, and obtain the task execution result. Specifically, the low-order task features obtained by each low-order task module 15 can refer to first-order non-intersection information based on the sample information, and the low-order shared features obtained by the shared low-order feature module 14 can be second-order intersection information based on the sample information, wherein non-intersection information refers to information of certain individual calculation items involved in the calculation process of the low-order task module 15, these individual calculation items include a single sub-feature of the sample information, and can also be called first-order non-intersection information; and intersection information refers to information of other individual calculation items involved in the calculation process of the shared low-order feature module 14, these individual calculation items include multiple sub-features of the sample information, and if they include two sub-features, they can be called second-order intersection information.

[0049] For example, in the calculation process shown in Formula 1 below, is the cross information, that is, the second-order cross information, and Non-cross information is first-order non-cross information:

[0050]

[0051] Generally, when acquiring features of sample information, deep machine learning models (such as the shared underlying subnetwork 10 and underlying task subnetwork 11 described above) primarily acquire high-order features. In this embodiment, the shared low-order feature module 14 and low-order task module 15 supplement the low-order features, enabling a more complete and accurate description of the sample information. Furthermore, the parameters of the multi-objective initial model refer to fixed parameters used by each layer of the multi-objective initial model during the calculation process that do not require constant assignment, such as parameter scale, number of network layers, user vector length, and other parameters.

[0052] Step 102: determine a training sample, where the training sample includes a plurality of sample information.

[0053] Furthermore, the training samples also include multiple task annotations corresponding to each sample information. For example, if a sample information is a sample image, the corresponding task annotation can be the location annotation of a specific object (such as a face) in the sample image. The task annotations here are mainly used to supervise the learning of the parameter values of each parameter in the multi-objective initial model during the subsequent training process.

[0054] Step 103, based on the multi-objective initial model and the training samples, calculate the overall loss function related to the multi-objective initial model, the overall loss function includes: the first correlation information between the input features and output features of the shared underlying sub-network 10 based on the sample information, and the second correlation information between the output features of the multiple underlying task sub-networks 11 based on the sample information and the output features of the shared underlying sub-network 10.

[0055] Specifically, when calculating the overall loss function, such as Figure 4 As shown, this can be achieved through the following steps:

[0056] A. The shared underlying subnetwork 10 and each underlying task subnetwork 11 in the multi-objective initial model respectively obtain the shared feature information of any sample information and the feature information adapted to the corresponding task, and the shared feature information and the feature information obtained by the underlying task subnetwork 11 are fused through the task feature module 12 to obtain the fused feature information, and then the task output module 13 executes the corresponding task according to the fused feature information to obtain the task execution result.

[0057] Among them, in the process of obtaining the task execution results, if the multi-objective initial model also includes the above-mentioned shared low-order feature module 14 and the low-order task module 15, the low-order shared features of the sample information can be obtained through the shared low-order feature module 14 and the low-order task features of the sample information can be obtained through a low-order task module 15. In this way, the corresponding task output module 13 will specifically execute the corresponding task according to the above-mentioned fused feature information, low-order shared features and low-order task features to obtain the task execution results.

[0058] B. Calculate the first relevant information based on the input features and output shared feature information of the shared underlying sub-network 10, and maximize the first relevant information to obtain a first association loss function based on the shared underlying sub-network 10.

[0059] C. Calculate the second relevant information based on the shared feature information output by the shared underlying sub-network 10 and the feature information output by each underlying task sub-network 11, and minimize the second relevant information to obtain a second association loss function based on the underlying task sub-network 11.

[0060] Specifically, when calculating the first association loss function and the second association loss function, it is necessary to measure the information correlation between two arbitrary variables based on the mutual information (MI) method, and MI maximization can maximize the correlation between the two variables, and MI minimization can minimize the correlation between the two variables.

[0061] For example, a neural network can be used to implement a KL-divergence-based representation to estimate the lower bound of the MI between two variables, as shown in the following formula 2. A floating-point score can be obtained, denoted as score:

[0062]

[0063] In this embodiment, in order to maximize the MI between variables to obtain the first association loss function, the two variables involved are the input features and output shared feature information of the shared underlying sub-network 10, so the obtained first association loss function can be expressed by the following formula 3:

[0064] log(1+e -score ) (3)

[0065] In this embodiment, in order to minimize the MI between variables to obtain the second association loss function, the two variables involved are the shared feature information output by the shared underlying sub-network 10 and the feature information output by any underlying task sub-network 11. Therefore, the obtained second association loss function can be expressed by the following formula 4:

[0066] log(1+e score ) (4)

[0067] Further, perform the following step D to calculate the third loss function, and then perform step E to calculate the overall loss function.

[0068] D. Calculate the third loss function corresponding to each task output module 13 based on the task execution results output by each task output module 13 and the task labels corresponding to the sample information.

[0069] The third loss function is used to indicate the error between the task execution results obtained by the multi-objective initial model based on each sample information and the actual information of each sample information (obtained according to the task annotation in the training sample), such as the cross entropy loss function.

[0070] E. Calculate the overall loss function associated with the multi-objective initial model based on the above-mentioned first associated loss function, second associated loss function and third loss function.

[0071] Specifically, the weighted sum of the first associated loss function, the second associated loss function and the third loss function can be used as the overall loss function.

[0072] Step 104 , adjusting the multi-objective initial model according to the overall loss function to obtain a final multi-objective model, where the multi-objective model is used to obtain multiple task execution results based on the data to be processed.

[0073] It should be noted that the training process of the multi-objective model is to minimize the value of the above-mentioned overall loss function. The training process continuously optimizes the parameter values of the parameters in the multi-objective initial model determined in the above-mentioned step 101 through a series of mathematical optimization methods such as backpropagation derivation and gradient descent, and reduces the calculated value of the above-mentioned overall loss function to the minimum.

[0074] Specifically, when the function value of the calculated loss function is large, such as greater than the preset value, it is necessary to change the parameter value, such as reducing the weight value of a certain neuron connection, so that the function value of the loss function calculated according to the adjusted parameter value is reduced.

[0075] It should be noted that the above steps 103 to 104 are an adjustment of the parameter values in the multi-objective initial model through the task execution results corresponding to each sample information output by the multi-objective initial model. In actual applications, it is necessary to continuously loop through the above steps 103 to 104 until the adjustment of the parameter values meets certain stopping conditions.

[0076] Therefore, after executing steps 101 to 104 of the above embodiment, the multi-objective learning system needs to determine whether the current adjustment of the parameter value meets the preset stopping condition. If it does, the process ends and the parameter value adjusted in step 104 is used as the parameter value in the multi-objective model finally trained. If it does not meet the conditions, the system returns to executing steps 103 to 104 for the multi-objective initial model after adjusting the parameter value. The preset stopping condition includes, but is not limited to, any one of the following conditions: the difference between the currently adjusted parameter value and the last adjusted parameter value is less than a threshold, i.e., the adjusted parameter value reaches convergence; and the number of parameter value adjustments is equal to the preset number, etc.

[0077] It can be seen that in the training process of the multi-objective model in this embodiment, it is determined that the multi-objective model includes a shared underlying sub-network, multiple underlying task sub-networks, multiple task feature modules and multiple task output modules, and it is determined that the training sample includes multiple sample information. The overall loss function related to the multi-objective initial model calculated based on the multi-objective initial model and the training sample includes: the first related information between the input features and output features of the shared underlying sub-network based on the sample information, and the second related information between the output features of the multiple underlying task sub-networks based on the sample information and the output features of the shared underlying sub-network, and then the multi-objective initial model is adjusted according to the overall loss function. Since the learning of the shared module, i.e., the shared underlying sub-network, is involved in the training process of the multi-objective model, the first related information and the second related information related to the shared underlying sub-network in the overall loss function can well guide the training process of the multi-objective model, so that information redundancy will not be caused, thereby improving the effect of the learned multi-objective model.

[0078] The following is a specific application example to illustrate the multi-objective learning method of the present invention. In this embodiment, two tasks performed on sample information are used as an example. Figure 5 As shown, the following steps may be included:

[0079] Step 201: determine the multi-objective initial model, wherein the structure of the multi-objective initial model can be as follows: Figure 6 As shown, it includes a shared expert network, expert network A, expert network B, task A network (task A tower), task B network, task output A, task output B, logistic regression network A, logistic regression (LR) network B and shared FM.

[0080] Among them, the shared expert network is the above-mentioned shared underlying subnetwork, each expert network is the above-mentioned underlying task subnetwork, each task network is the above-mentioned task feature module, the task output is the above-mentioned task output module, each logistic regression network is the above-mentioned low-order task module, and the shared FM is the above-mentioned shared low-order feature module.

[0081] Step 202: determine training samples, which may specifically include multiple sample information and task labels corresponding to each sample information.

[0082] Step 203: Obtain shared feature information Emb of any sample information through the shared expert network and each expert network in the multi-objective initial model. share and feature information adapted to the corresponding task (denoted as Emb in this embodiment) A and Emb B), and the shared feature information Emb is shared through each task network (such as task A network or task B network) share And the feature information obtained by an expert network (Emb A or Emb B ) to obtain the fused feature information (denoted as or ).

[0083] Step 204: low-order shared features (denoted as O) of the sample information are obtained through the shared FM and each logistic regression network. fm ) and low-level task features (denoted as or ).

[0084] Step 205: Each task outputs the fused feature information ( or ), low-order shared features O fm and low-level task features ( or ) executes the corresponding task and obtains the task execution result (denoted as O A or O B ).

[0085] Step 206 , calculate the overall loss function associated with the multi-objective initial model.

[0086] Specifically, the multi-objective learning system is based on the input features of the shared expert network and the output shared feature information Emb share Calculate the first relevant information MI1, and maximize the first relevant information MI1 to obtain the first correlation loss function loss based on the shared expert network max ; According to the shared feature information Emb output by the shared expert network share And the feature information output by each expert network (Emb A or Emb B ) calculates the second relevant information MI2, and minimizes the second relevant information MI2 to obtain the second associated loss function based on the underlying task sub-network (denoted as loss min1 or loss min2 ); According to the task execution results output by each task ( A or O B ) and the task labels corresponding to the sample information, and calculate the third loss function (denoted as loss1 or loss2) corresponding to the output of each task.

[0087] Then the multi-objective learning system is based on the first associated loss function loss max , the second associated loss function (loss min1 or lossmin2 ) and the third loss function (loss1 or loss2) to calculate the overall loss function L, which can be expressed as follows:

[0088] L = loss max +loss min1 +loss min2 +loss1+loss2 (5)

[0089] Step 207: Adjust the parameter values of the parameters in the multi-objective initial model according to the overall loss function calculated in the above steps.

[0090] In step 208, it is determined whether the current adjustment of the parameter value satisfies the preset stop condition. If so, the parameter value obtained by adjusting in the above step 207 is used as the parameter value in the final multi-objective model to obtain the trained multi-objective model; if not, the process returns to execute step 203 as above.

[0091] It can be seen that the expert networks included in the multi-objective initial model determined in this embodiment are not many, and only three are used, namely the shared expert network, expert network A, and expert network B, and the performance of the multi-objective model obtained through training in the above steps is greatly improved. Specifically, when the existing MMoE, CGC, PLE models and the multi-objective model trained in this embodiment are used to perform two tasks (task A and task B) on the data to be processed, and the evaluation index based on each model, namely the area under the curve (AUC), is calculated, as shown in Table 1 below:

[0092] Model / Indicator AUC-A AUC-B MMoE 0.762 0.781 CGC 0.765 0.783 PLE 0.767 0.784 Model of this embodiment 0.768 0.787

[0093] Table 1

[0094] The AUC can represent the probability that, for any pair of positive and negative samples, the model predicts the positive sample as a positive example more likely than the negative example as a positive example. The higher the AUC value, the higher the performance of the model. It can be seen that the performance of the multi-target model trained by the method in this embodiment is greatly improved.

[0095] It should be noted that the multi-objective model trained in this embodiment can be applied to a variety of application scenarios. For example, in the information recommendation scenario, in the process of sorting recommended information, the multi-objective model can be used to estimate the click-through rate and browsing time of the candidate information, and then the information can be sorted according to the estimated information, and the sorted information can be pushed to the user terminal.

[0096] The following is another specific application example to illustrate the multi-objective learning method in the present invention. The multi-objective learning system in the embodiment of the present invention is mainly a distributed system 100, which may include a client 300 and multiple nodes 200 (any form of computing device in the access network, such as a server, a user terminal), and the client 300 and the node 200 are connected through network communication.

[0097] Taking the distributed system as the blockchain system as an example, see Figure 7 This is a schematic diagram of an optional architecture for a distributed system 100 provided in an embodiment of the present invention, applied to a blockchain system. The system consists of multiple nodes 200 (any type of computing device connected to a network, such as a server or user terminal) and clients 300. The nodes form a peer-to-peer (P2P) network. The P2P protocol is an application layer protocol that runs on top of the Transmission Control Protocol (TCP). In a distributed system, any machine, such as a server or terminal, can join and become a node. Nodes include hardware, middleware, operating system, and application layers.

[0098] See also Figure 7 The functions of each node in the blockchain system shown include:

[0099] 1) Routing: A basic function of a node, used to support communication between nodes.

[0100] In addition to the routing function, nodes can also have the following functions:

[0101] 2) Applications, deployed in the blockchain, implement specific services based on actual business needs, record data related to the implementation of functions to form record data, carry digital signatures in the record data to indicate the source of the task data, and send the record data to other nodes in the blockchain system for other nodes to add the record data to a temporary block when they successfully verify the source and integrity of the record data.

[0102] For example, the services implemented by the application include code for implementing a multi-objective learning function, which mainly includes:

[0103] Determine a multi-objective initial model, the multi-objective initial model including: a shared underlying subnetwork, multiple underlying task subnetworks, multiple task feature modules and multiple task output modules; determine a training sample, the training sample including multiple sample information; calculate an overall loss function related to the multi-objective initial model based on the multi-objective initial model and the training sample; the overall loss function includes: first related information between the input features and output features of the shared underlying subnetwork based on the sample information, and second related information between the output features of the multiple underlying task subnetworks based on the sample information and the output features of the shared underlying subnetwork respectively; adjust the multi-objective initial model according to the overall loss function to obtain a final multi-objective model, the multi-objective model is used to obtain multiple task execution results based on the data to be processed.

[0104] 3) Blockchain, including a series of blocks that are connected to each other in the order of their generation. Once a new block is added to the blockchain, it will not be removed. The block records the record data submitted by the nodes in the blockchain system.

[0105] See also Figure 8 This is an optional schematic diagram of the block structure provided by an embodiment of the present invention. Each block includes the hash value of the transaction records stored in the block (the hash value of the current block) and the hash value of the previous block. The blocks are connected by hash values to form a blockchain. In addition, the block may also include information such as the timestamp when the block was generated. Blockchain is essentially a decentralized database, a series of data blocks generated using cryptographic methods. Each data block contains relevant information used to verify the validity of the information (anti-counterfeiting) and generate the next block.

[0106] The embodiment of the present invention also provides a multi-objective learning system, the structural diagram of which is shown in FIG. Figure 9 Specifically, it may include:

[0107] The model determination unit 20 is used to determine a multi-objective initial model, wherein the multi-objective initial model includes: a shared underlying sub-network, multiple underlying task sub-networks, multiple task feature modules and multiple task output modules.

[0108] The model determination unit 20 is specifically used to determine a multi-objective initial model with the following structure, in which: the inputs of the shared underlying sub-network and multiple underlying task sub-networks are the same, and the outputs of the shared underlying sub-networks are respectively connected to an input of any task feature module, the outputs of any underlying task sub-networks are respectively connected to the input of a task feature module, and the outputs of any task feature module are connected to the input of a task output module. Furthermore, the multi-objective initial model also includes: a shared low-order feature module and multiple low-order task modules; the inputs of the shared low-order feature module and the low-order task module are the same as the inputs of the shared underlying sub-network, the outputs of the shared low-order feature modules are respectively connected to any task output module, and the outputs of any of the low-order task modules are connected to a task output module.

[0109] The sample determination unit 21 is configured to determine a training sample, where the training sample includes a plurality of sample information.

[0110] The loss function unit 22 is used to calculate the overall loss function related to the multi-objective initial model based on the multi-objective initial model determined by the model determination unit 20 and the training samples determined by the sample determination unit 21; the overall loss function includes: the first related information between the input features and output features of the shared underlying sub-network based on the sample information, and the second related information between the output features of the multiple underlying task sub-networks based on the sample information and the output features of the shared underlying sub-network.

[0111] The loss function unit 22 is specifically used to obtain the shared feature information of any sample information and the feature information adapted to the corresponding task through the shared underlying subnetwork and each underlying task subnetwork in the multi-objective initial model, and fuse the shared feature information and the feature information obtained by an underlying task subnetwork through the task feature module to obtain the fused feature information, and the task output module executes the corresponding task according to the fused feature information to obtain the task execution result; calculate the first related information according to the input feature of the shared underlying subnetwork and the output shared feature information, and maximize the first related information to obtain a first associated loss function based on the shared underlying subnetwork; calculate the second related information according to the shared feature information output by the shared underlying subnetwork and the feature information output by each underlying task subnetwork, and minimize the second related information to obtain a second associated loss function based on the underlying task subnetwork; calculate the overall loss function according to the first associated loss function and the second associated loss function.

[0112] If the multi-objective initial model also includes: a shared low-order feature module and multiple low-order task modules; then when calculating the overall loss function, the loss function unit 22 is specifically used to obtain the low-order shared features and low-order task features of the sample information respectively through the shared low-order feature module and each low-order task module; the task output module executes the corresponding task according to the fused feature information, low-order shared features and low-order task features to obtain the task execution result.

[0113] If the training sample also includes task labels corresponding to each sample information, then when calculating the overall loss function, the loss function unit 22 is specifically used to calculate the third loss function corresponding to each task output module based on the task execution results output by each task output module and the task labels corresponding to the sample information; and calculate the overall loss function based on the first associated loss function, the second associated loss function and the third loss function.

[0114] The adjustment unit 23 is used to adjust the multi-objective initial model according to the overall loss function calculated by the loss function unit 22 to obtain a final multi-objective model, and the multi-objective model is used to obtain multiple task execution results based on the data to be processed.

[0115] The adjusting unit 23 is specifically configured to adjust the parameter values of the parameters in the multi-objective initial model according to the overall loss function;

[0116] Furthermore, the adjusting unit 23 is further configured to stop adjusting the parameter value when the number of times the parameter value is adjusted is equal to a preset number, or when the difference between the currently adjusted parameter value and the last adjusted parameter value is less than a threshold.

[0117] It can be seen that in the system of this embodiment, the model determination unit 20 will determine that the multi-objective model includes a shared underlying sub-network, multiple underlying task sub-networks, multiple task feature modules and multiple task output modules, and the sample determination unit 21 determines that the training sample includes multiple sample information, and the loss function unit 22 calculates the overall loss function related to the multi-objective initial model based on the multi-objective initial model and the training sample, which includes: the first related information between the input features and output features of the shared underlying sub-network based on the sample information, and the second related information between the output features of the multiple underlying task sub-networks based on the sample information and the output features of the shared underlying sub-network, and then the adjustment unit 23 adjusts the multi-objective initial model according to the overall loss function. Since the learning of the shared module, i.e., the shared underlying sub-network, is involved in the training process of the multi-objective model, the first related information and the second related information related to the shared underlying sub-network in the overall loss function can well guide the training process of the multi-objective model, so that information redundancy will not be caused, thereby improving the effect of the learned multi-objective model.

[0118] The embodiment of the present invention further provides a terminal device, the structural diagram of which is shown in FIG. Figure 10 As shown, the terminal device may have relatively large differences due to different configurations or performances, and may include one or more central processing units (CPUs) 30 (for example, one or more processors) and memory 31, and one or more storage media 32 (for example, one or more mass storage devices) storing application programs 321 or data 322. Memory 31 and storage medium 32 may be temporary storage or permanent storage. The program stored in the storage medium 32 may include one or more modules (not shown), each module may include a series of instruction operations in the terminal device. Furthermore, the central processing unit 30 may be configured to communicate with the storage medium 32 to execute a series of instruction operations in the storage medium 32 on the terminal device.

[0119] Specifically, the application 321 stored in the storage medium 32 includes a multi-objective learning application, and the application may include the model determination unit 20, sample determination unit 21, loss function unit 22, and adjustment unit 23 in the multi-objective learning system described above, which are not described in detail here. Furthermore, the central processing unit 30 can be configured to communicate with the storage medium 32 and execute a series of operations corresponding to the multi-objective learning application stored in the storage medium 32 on the terminal device.

[0120] The terminal device may also include one or more power supplies 33, one or more wired or wireless network interfaces 34, one or more input and output interfaces 35, and / or one or more operating systems 323, such as Windows Server™, Mac OS X™, Unix™, Linux™, FreeBSD™, etc.

[0121] The steps performed by the multi-objective learning system in the above method embodiment can be based on the Figure 10 The structure of the terminal device shown.

[0122] Furthermore, another aspect of an embodiment of the present invention provides a computer-readable storage medium, which stores multiple computer programs, and the computer programs are suitable for being loaded by a processor and executing the multi-objective learning method described in the multi-objective learning system as described above.

[0123] Another aspect of the present invention provides a terminal device, including a processor and a memory;

[0124] The memory is used to store multiple computer programs, and the computer programs are used to be loaded by the processor and execute the multi-objective learning method described in the above-mentioned multi-objective learning system; the processor is used to implement each computer program in the multiple computer programs.

[0125] In addition, according to one aspect of the present application, a computer program product or computer program is provided, the computer program product or computer program including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the multi-objective learning method provided in the various optional implementations described above.

[0126] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be completed by instructing related hardware through a program, and the program can be stored in a computer-readable storage medium, which may include: read-only memory (ROM), random access memory (RAM), magnetic disk or optical disk, etc.

[0127] The above is a detailed introduction to a multi-objective learning method, system, storage medium and terminal device provided by an embodiment of the present invention. Specific examples are used in this article to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core ideas. At the same time, for those skilled in the art, according to the ideas of the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A multi-objective learning method, characterized in that: include: Determine a multi-objective initial model, the multi-objective initial model comprising: a shared underlying subnetwork, multiple underlying task subnetworks, multiple task feature modules, multiple task output modules, a shared low-order feature module, and multiple low-order task modules; the inputs of the shared underlying subnetwork and the multiple underlying task subnetworks are the same, and the outputs of the shared underlying subnetworks are respectively connected to an input of any task feature module, the outputs of any underlying task subnetworks are respectively connected to the input of a task feature module, and the outputs of any task feature module are connected to the input of a task output module; the inputs of the shared low-order feature module and the low-order task module are the same as the inputs of the shared underlying subnetwork, the outputs of the shared low-order feature modules are respectively connected to any task output module, and the outputs of any low-order task modules are connected to a task output module; Determine a training sample, wherein the training sample includes a plurality of sample information; wherein the sample information is a sample image; Calculate an overall loss function associated with the multi-objective initial model based on the multi-objective initial model and the training samples; the overall loss function includes: first correlation information between the input features and output features of the shared underlying sub-network based on the sample information, and second correlation information between the output features of the multiple underlying task sub-networks based on the sample information and the output features of the shared underlying sub-network; The multi-objective initial model is adjusted according to the overall loss function to obtain a final multi-objective model, wherein the multi-objective model is used to obtain multiple task execution results based on the data to be processed, including: through the multi-objective model, multiple tasks are executed based on the data to be processed to obtain the execution results of multiple tasks.

2. The method according to claim 1, wherein The calculating of the overall loss function related to the multi-objective initial model according to the multi-objective initial model and the training samples specifically includes: The shared feature information of any sample information and the feature information adapted to the corresponding task are respectively obtained through the shared underlying subnetwork and each underlying task subnetwork in the multi-objective initial model, and the shared feature information and the feature information obtained by an underlying task subnetwork are fused through the task feature module to obtain fused feature information, and the task output module executes the corresponding task according to the fused feature information to obtain a task execution result; Calculating first relevant information based on the input features and output shared feature information of the shared underlying sub-network, and maximizing the first relevant information to obtain a first association loss function based on the shared underlying sub-network; Calculating second relevant information based on the shared feature information output by the shared underlying subnetwork and the feature information output by each underlying task subnetwork, and minimizing the second relevant information to obtain a second association loss function based on the underlying task subnetwork; The overall loss function is calculated according to the first associated loss function and the second associated loss function.

3. The method according to claim 2, wherein The multi-objective initial model also includes: a shared low-level feature module and multiple low-level task modules; the task output module performs corresponding tasks according to the fused feature information to obtain task execution results, specifically including: Obtaining low-order shared features and low-order task features of the sample information respectively through the shared low-order feature module and each low-order task module; The task output module executes the corresponding task according to the fused feature information, low-order shared features and low-order task features to obtain a task execution result.

4. The method according to claim 2, wherein The training samples also include task labels corresponding to each sample information, and then calculating the overall loss function based on the first associated loss function and the second associated loss function specifically includes: Calculating a third loss function corresponding to each task output module according to the task execution results output by each task output module and the task labels corresponding to the sample information; The overall loss function is calculated according to the first associated loss function, the second associated loss function and the third loss function.

5. The method according to claim 2, wherein The adjusting the multi-objective initial model according to the overall loss function specifically includes: adjusting parameter values of parameters in the multi-objective initial model according to the overall loss function; The method further includes: stopping adjusting the parameter value when the number of times the parameter value is adjusted is equal to a preset number, or when the difference between the currently adjusted parameter value and the last adjusted parameter value is less than a threshold.

6. A multi-objective learning system, characterized in that include: A model determination unit is used to determine a multi-objective initial model, wherein the multi-objective initial model includes: a shared underlying subnetwork, multiple underlying task subnetworks, multiple task feature modules, multiple task output modules, a shared low-order feature module, and multiple low-order task modules; the inputs of the shared underlying subnetwork and the multiple underlying task subnetworks are the same, and the outputs of the shared underlying subnetworks are respectively connected to an input of any task feature module, the outputs of any underlying task subnetworks are respectively connected to the input of a task feature module, and the outputs of any task feature modules are connected to the input of a task output module; the inputs of the shared low-order feature modules and the low-order task modules are the same as the inputs of the shared underlying subnetwork, the outputs of the shared low-order feature modules are respectively connected to any task output modules, and the outputs of any low-order task modules are connected to a task output module; A sample determination unit, configured to determine a training sample, wherein the training sample includes a plurality of sample information; wherein the sample information is a sample image; A loss function unit is configured to calculate an overall loss function associated with the multi-objective initial model based on the multi-objective initial model and the training samples; the overall loss function includes: first correlation information between the input features and output features of the shared underlying sub-network based on the sample information, and second correlation information between the output features of the multiple underlying task sub-networks based on the sample information and the output features of the shared underlying sub-network; An adjustment unit is used to adjust the multi-objective initial model according to the overall loss function to obtain a final multi-objective model, wherein the multi-objective model is used to obtain multiple task execution results based on the data to be processed, including: through the multi-objective model, multiple tasks are executed based on the data to be processed to obtain the execution results of multiple tasks.

7. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a plurality of computer programs, and the computer programs are suitable for being loaded by a processor and executing the multi-objective learning method according to any one of claims 1 to 5.

8. A terminal device, characterized in that: including processor and memory; The memory is used to store multiple computer programs, and the computer programs are used to be loaded by the processor and execute the multi-objective learning method according to any one of claims 1 to 5; the processor is used to implement each computer program in the multiple computer programs.

9. A computer program product, characterized in that The computer program product includes computer instructions, which are stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device performs the multi-objective learning method according to any one of claims 1 to 5.

Images