Object screening method and apparatus, electronic device, storage medium, and program product

Abstract

The embodiment of the application provides a kind of object screening method, device, electronic equipment, storage medium and program product, it is related to artificial intelligence, image processing, big data and other fields, the method is based on the probability distribution information that each data subset of target object data set is screened out, wherein, for any data subset, the probability distribution information determined based on energy model to the data subset of target object sequential transformation is same, i.e. energy model has permutation invariance, can realize the object screening process based on energy model and the order of each target object in target object data set is irrelevant, and then according to the probability distribution information, at least one target data subset is determined from each data subset, to obtain object screening result, higher object screening precision can be realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of machine learning, and in particular, relates to an object screening method and device, electronic equipment, storage medium and program product. BACKGROUND

[0002] Subset sampling is an important research problem in the field of machine learning, and has a wide range of application scenarios in real life, such as product recommendation.

[0003] Existing subset sampling methods are often related to sampling order, for example, in the scenario of product recommendation, after the first product is recommended, the first product will affect the selection of the second product.

[0004] However, in real life, subset sampling is affected by the arrangement order of elements in the set, which reduces the sampling accuracy. SUMMARY

[0005] The embodiments of the present application aim to solve the problem of reduced sampling accuracy caused by the influence of the arrangement order of elements in the set on subset sampling.

[0006] According to an aspect of the embodiments of the present application, an object screening method is provided, which comprises:

[0007] obtaining a target object data set, the target object data set comprising a plurality of target objects to be screened;

[0008] determining probability distribution information of each data subset of the target object data set being screened based on an energy model that has been constructed, wherein for any data subset, the probability distribution information determined by the energy model for the data subset after the order of the target objects is transformed is the same;

[0009] determining at least one target data subset from each data subset according to the probability distribution information, to obtain an object screening result.

[0010] According to another aspect of the embodiments of the present application, an object screening device is provided, which comprises:

[0011] an obtaining module configured to obtain a target object data set, the target object data set comprising a plurality of target objects to be screened;

[0012] a determining module configured to determine probability distribution information of each data subset of the target object data set being screened based on an energy model that has been constructed, wherein for any data subset, the probability distribution information determined by the energy model for the data subset after the order of the target objects is transformed is the same;

[0013] The screening module is configured to determine at least one target data subset from each data subset according to the probability distribution information, so as to obtain an object screening result.

[0014] According to a further aspect of the embodiments of the present application, an electronic device is provided, which comprises a memory, a processor and a computer program stored in the memory. The processor executes the computer program to implement the steps of the object screening method provided by the embodiments of the present application.

[0015] According to a further aspect of the embodiments of the present application, a computer readable storage medium is provided, which stores a computer program. The computer program is executed by a processor to implement the steps of the object screening method provided by the embodiments of the present application.

[0016] According to a further aspect of the embodiments of the present application, a computer program product is provided, which comprises a computer program. The computer program is executed by a processor to implement the steps of the object screening method provided by the embodiments of the present application.

[0017] The object screening method, device, electronic device, storage medium and program product provided by the embodiments of the present application determine the probability distribution information of each data subset of the target object data set based on the constructed energy model. For any data subset, the probability distribution information determined by the energy model for the data subset after the target object is sequentially transformed is the same, that is, the energy model has permutation invariance. The object screening process based on the energy model is independent of the order of each target object in the target object data set. Then, at least one target data subset is determined from each data subset according to the probability distribution information, so as to obtain an object screening result. High object screening accuracy can be achieved. BRIEF DESCRIPTION OF DRAWINGS

[0018] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the description of the embodiments of the present application will be briefly introduced.

[0019] Figure 1 A flowchart of an object screening method provided by the embodiments of the present application;

[0020] Figure 2 A schematic diagram of the relationship between the target object data set and each data subset provided by the embodiments of the present application;

[0021] Figure 3 A schematic diagram of each permutation result of the subset provided by the embodiments of the present application;

[0022] Figure 4 A schematic diagram of modeling permutation invariance provided by the embodiments of the present application;

[0023] Figure 5 A schematic diagram of an image detection result provided by an embodiment of the present application;

[0024] Figure 6 A schematic diagram of compound selection in drug discovery provided by an embodiment of the present application;

[0025] Figure 7 A schematic diagram of an implementation environment provided by an embodiment of the present application;

[0026] Figure 8 A structural schematic diagram of an object screening device provided by an embodiment of the present application;

[0027] Figure 9 A structural schematic diagram of an electronic device provided by an embodiment of the present application. DETAILED DESCRIPTION

[0028] Embodiments of the present application will be described below in conjunction with the accompanying drawings. It should be understood that the embodiments described below in conjunction with the accompanying drawings are exemplary descriptions of the technical solutions of the embodiments of the present application, and do not constitute a limitation on the technical solutions of the embodiments of the present application.

[0029] Those skilled in the art can understand that the singular forms "a", "an" and "the" used herein include plural forms, unless specifically stated otherwise. It should be further understood that the terms "comprise" and "include" used in the embodiments of the present application mean that the corresponding features can be implemented as the presented features, information, data, steps, operations, elements and / or components, but do not exclude other features, information, data, steps, operations, elements, components and / or combinations thereof supported by the present technology. It should be understood that when we say that an element is "connected" or "coupled" to another element, the element can be directly connected or coupled to the other element, or it can mean that the element and the other element are connected through an intermediate element. In addition, "connected" or "coupled" used herein can include wireless connection or wireless coupling. The term "and / or" used herein indicates that at least one of the items defined by the term, for example, "A and / or B" can be implemented as "A", or as "B", or as "A and B".

[0030] In order to make the purpose, technical solutions and advantages of the present application clearer, the embodiments of the present application will be described in further detail below in conjunction with the accompanying drawings.

[0031] First, several terms related to the present application are introduced and explained:

[0032] (1)Artificial Intelligence (AI): AI is the theory, method, technology and application system that use digital computers or digital computer controlled machines to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results. It attempts to understand the essence of intelligence and produce a new intelligent machine that can react in a similar way to human intelligence. Artificial intelligence is the design principle and implementation method of various intelligent machines, so that machines have the functions of perception, reasoning and decision-making.

[0033] (2) Machine Learning (ML): ML is a multi-disciplinary subject that involves probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory, etc. It is a subject that studies how computers simulate or implement human learning behavior to acquire new knowledge or skills, and reorganize existing knowledge structure to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental approach to making computers intelligent, and its applications are widespread in various fields of artificial intelligence. Machine learning and deep learning usually include artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and rule-based learning.

[0034] (3) Set function: The processing object (i.e. domain) of the set function is a set.

[0035] (4) Variations: Variational method is a mathematical means for handling functionals (functions with functions as variables), which ultimately seeks the extremal function that makes the functional attain the maximum or minimum value. Variational inference (VI) is the application of variational method in inference problems. For the posterior probability density that cannot be directly obtained, by inferring a distribution to approximate the posterior probability density, the inference problem can be converted into a functional optimization problem.

[0036] (5) Mean Field (MF): Mean field theory is a paradigm and theory that approximates all influences on a single body in a model as an external field, thereby decomposing a many-body problem into multiple single-body problems for solving. In variational inference, combined with mean field theory, unobservable variables can be split into multiple groups of independent variables.

[0037] (6) KL divergence (Kullback-Leibler Divergence): Also known as relative entropy, information divergence or information gain, it is an asymmetric measure of the difference between two probability distributions.

[0038] (7)Inverse KL divergence: the KL divergence of distribution q with respect to distribution p is called inverse KL divergence, one advantage of using inverse KL divergence is that it can ensure to get local extremum.

[0039] (8)alpha divergence: by selecting different alpha values to measure the difference between two probability distributions, which has high flexibility.

[0040] (9)Wasserstein divergence: also known as EMD (Earth-Mover Distances), which measures the distance between two probability distributions by optimizing the minimum value of the function under linear constraints.

[0041] (10)Fixed point equation: by using the properties of the fixed point of the function, the function solving problem can be converted into using the fixed point equation to iterate for a limited number of times, so as to obtain the solution of the equation.

[0042] (11)MJC and MAP: in the field of target detection in machine learning, it is a very important measure to measure the performance of target detection algorithm, MAP is the weighted average of the average correct rate (AP) of all class detection, and the expression of MJC is

[0043] The object screening method provided by the embodiments of the present application aims to obtain an object screening result irrelevant to the order of target objects in a target object data set, and can also be understood as obtaining a subset sampling result irrelevant to the arrangement order of elements in a set. For example, in the scenario of commodity recommendation, the present scheme pre-implementation system only decides the set of recommended commodities according to the preferences of the general public, and is irrelevant to the order of commodity sorting in the commodity library.

[0044] For ease of understanding, in the embodiments of the present application, the subset sampling problem is defined as: given a set of data wherein the set S is a subset corresponding to the set V, and the target is to estimate the parameter θ, so that for all data satisfy:

[0045]

[0046] wherein, F θ (S) is a set function, and the function argmax is used to find the subset of the set function.

[0047] After the parameter estimation is completed, the optimal subset of a set can be obtained in the following way:

[0048]

[0049] In the prior art, in order to solve the above problem, a probability function method is proposed. Specifically, the first j elements in the subset S are denoted as Then the j+1th sample is sampled from the set V\S with a probability of j (i.e., any element ∈V in the set but ), where γ is a temperature coefficient. Therefore, after a series of sampling processes, the distribution corresponding to π={s1, s2, …, s m} is

[0050]

[0051] where However, it should be noted that the calculation of p(π; θ) depends on the order of the elements in the sequence π, which makes the parameter θ very sensitive to the order of sampling.

[0052] In another related art, in order to solve this problem, it is proposed that the sum of the probabilities of all possible permutations can be maximized:

[0053]

[0054] where Π S represents the permutation space corresponding to S. Obviously, the calculation cost of this objective is very expensive. In addition, this method can only reduce the sensitivity of the algorithm to the sampling order, but cannot truly achieve the permutation invariance of the algorithm.

[0055] The object screening method and device, electronic equipment, storage medium and program product provided by the present application aim to solve the above technical problems in the prior art. In view of at least one of the above problems existing in the related art, the corresponding solutions are given.

[0056] The technical solutions of the embodiments of the present application and the technical effects produced by the technical solutions of the present application will be described below through the description of several exemplary embodiments. It should be noted that the following embodiments can be mutually referenced, borrowed or combined. For the same terms, similar features and similar implementation steps in different embodiments, they will not be described repeatedly.

[0057] An object screening method is provided in the embodiments of the present application, as shown in Figure 1 The method comprises the following steps.

[0058] Step S101: Obtain a target object data set, wherein the target object data set comprises a plurality of target objects to be screened.

[0059] The target object data set can be understood as a set to be sampled, i.e., a data set from which a target object is needed to be sampled. The target object is an element in the target object data set. In the embodiments of the present application, the type and representation of the target object are not limited, and optionally, different types of target objects can have the same or different representations. For example, an image type target object can be represented as a matrix, and other types of target objects can also be represented as a vector, a sequence, etc., but are not limited thereto.

[0060] In step S102, based on the constructed energy model, probability distribution information of each data subset of the target object data set is determined. For any data subset, the probability distribution information determined by the energy model for the data subset after the target object is sequentially transformed is the same.

[0061] In the embodiments of the present application, the energy model is constructed based on the probability distribution information of each data subset formed for each target object in the target object data set. The energy model can be modeled based on an energy function capable of implementing the execution process.

[0062] As an example, an optional energy model is a set function wherein V represents the target object data set (i.e., the set to be sampled), |V| represents the number of target objects in the target object data set (i.e., the number of elements in the set to be sampled), 2 |V| and S represents each data subset of the target object data set. Wherein, 2 |V| Each target object in the target object data set has two possibilities of being extracted or not being extracted, and all possible combinations can form each data subset of the target object data set. As an example, taking the number of target objects in the target object data set as 2, the relationship between the target object data set V and each data subset (S0-S3) is shown in Figure 2 .

[0063] In the embodiments of the present application, the energy model can be modeled by a neural network with an input of each data subset of the target object data set (i.e., the set to be sampled) and an output of the probability value or probability distribution information of each data subset. In actual applications, a person skilled in the art can select a suitable type and structure of neural network according to actual needs, which is not limited in the embodiments of the present application.

[0064] In the embodiments of the present application, the energy model is modeled to have permutation invariance, that is, for any individual data subset of the target object data set, the probability value determined based on the energy model is the same. Wherein, the permutation result of any individual data subset refers to various results obtained by permuting the positions of the target objects in the individual data subset, that is, for any data subset, the probability distribution information determined by the energy model for the data subset after the order of the target objects is changed is the same. As an example, taking S3 in Figure 2 as an example, the permutation results of S3 are shown in Figure 3 , any permutation result is taken as the input of the energy model, and the corresponding probability value is unchanged.

[0065] Step S103: determining at least one target data subset from the individual data subsets according to the probability distribution information, to obtain the object screening result.

[0066] In the embodiments of the present application, the probability distribution information represents the distribution of the probability that the individual data subsets are screened out, for example, if ψ * is used to represent the probability that each data subset S i ∈V in the target object data set V should be screened out.

[0067] Therefore, in the embodiments of the present application, the screening of the subsets can be performed in the following manner:

[0068] S=topN(ψ * )

[0069] Wherein, topN(x) represents the subscript of the subset S i corresponding to the first N maximum values in the value x, and N is a positive integer, so that at least one target data subset can be obtained. Further, the target objects in the at least one target data subset can be determined as the object screening result, or the target objects in the at least one target data subset can be post-processed, for example, de-duplication or error elimination, but not limited thereto, and then the target objects obtained by the post-processing are determined as the object screening result.

[0070] In actual applications, the value of N can be set by the person skilled in the art according to the actual situation, which is not specifically limited in the embodiments of the present application.

[0071] The object screening method provided in the embodiments of the present application determines the probability distribution information of each data subset of the target object data set based on the constructed energy model, wherein the probability distribution information determined by the energy model for the data subset after the target object is sequentially transformed is the same for any data subset, that is, the energy model has permutation invariance, the object screening process based on the energy model is independent of the order of each target object in the target object data set, and then at least one target data subset is determined from each data subset according to the probability distribution information to obtain an object screening result, which can achieve high object screening accuracy.

[0072] In the embodiments of the present application, a feasible implementation manner is provided for the energy model, wherein the energy function can adopt a Deep Set function.

[0073] Specifically, the step S102 can include the steps of determining the probability representation of each data subset being screened out respectively, and performing exponential normalization on each probability representation to obtain the probability distribution information of each data subset being screened out. That is, the energy model can be modeled based on an energy function capable of implementing the execution process.

[0074] As an example, the energy function corresponding to the execution process can be represented as:

[0075]

[0076] wherein S represents each data subset, p(S; θ) is the probability distribution information of each data subset, F θ (S) is the set function described above, and θ is a model parameter of the energy model.

[0077] Specifically, the energy function maps the F θ (S) to a probability value of a non-negative probability by using the function exp, and Z normalizes the mapping results of each data subset by adding them (that is, the probabilities of each data subset are summed to 1). After the above processing, the energy function can exhibit the processing results of each data subset in the form of probability to obtain the probability value corresponding to each data subset, and the probability distribution information of each data subset can be obtained according to the distribution of the probability value of each data subset. In the embodiments of the present application, the probability distribution information can be a probability density distribution.

[0078] In other embodiments, the energy model can also be constructed in other manners, as long as the manner capable of obtaining the probability distribution information of each data subset can be applied to the present application, and therefore should also be included in the protection scope of the present application.

[0079] In the embodiments of the present application, in order to model the permutation invariance, the energy function based on which the modeling is performed can satisfy the permutation invariance.

[0080] Specifically, the inventors of the present application have found that when a probability density function p(S) of a set can be expressed as: p(S) is permutation invariant, where both p and f are transformation functions.

[0081] That is, in the embodiments of the present application, the probability of any data subset is proportional to the feature transformation result obtained by the following method: mapping each target object in any data subset to a vector of the first dimension to obtain a first feature transformation result of any data subset; taking the number of target objects in any data subset as the second dimension of the vector, and fusing the first feature transformation result based on the second dimension to obtain a fusion result; and performing nonlinear transformation on the fusion result to obtain a second feature transformation result. In simple terms, the probability value of any data subset has the following representation:

[0082]

[0083] Where S is any data subset, s is a target object in any data subset, p(S) is the probability value of any data subset, and both p and f are feature transformation functions. For example, as shown in Figure 4 f can map the target object to a vector of the first dimension (corresponding to W in Figure 4 After adding the vectors of the target objects in this any data subset (based on the second dimension, corresponding to H in Figure 4 ), p performs nonlinear transformation on the vector addition result. In actual applications, the fusion method is not limited to addition, for example, it can also include averaging, weighted addition, weighted averaging, etc., and those skilled in the art can extend it according to actual conditions, which is not limited in the embodiments of the present application.

[0084] For the embodiments of the present application, according to the above theorem, p(S; q) can be guaranteed to satisfy permutation invariance, so that the energy model satisfies permutation invariance.

[0085] In the embodiments of the present application, in order to avoid excessive calculation cost of p(S; q), for example, in order to avoid the constant term Z of the above energy function from being difficult to calculate, the method of using a variational distribution to approximate p(S; q) can be used. That is, in the embodiments of the present application, the output result of the energy model for each data subset can be variational inferred based on the optimized variational model to obtain the probability distribution information of each data subset being screened out.

[0086] In the embodiments of the present application, in order to make the probability distribution information (for ease of description, referred to as variational distribution here) approximate p(S; q), the variational lower bound can be maximized by optimizing the variational model, that is, the optimized variational model minimizes the distance between the probability distribution information and the output result, so that the optimal variational distribution q(S; q* )。

[0087] In the embodiments of the present application, the variational distribution can represent the distribution of the probability of each data subset being screened out, and specifically, a score can be used to represent the probability of the corresponding data subset being screened out, that is, the score ψ * in the variational distribution q(S; ψ * represents a score of each data subset S i in the target object data set V that should be screened out.

[0088] The object screening method provided in the embodiments of the present application can obtain an object screening result based on an energy model, calculate the probability distribution information of the probability (or score) of each data subset being screened out by using the mean field variational inference method, and then select the data subset, so that a high reliability of object screening can be achieved.

[0089] In the embodiments of the present application, when the variational distribution is used to approximate p(S; θ), the variational distribution can be constructed as:

[0090]

[0091] wherein ψ ∈ [0, 1] |V| , that is, the score ψ i corresponding to each target object in the data subset S and the score ψ j corresponding to each target object in the set V\S are in the range of [0, 1], and the score ψ corresponding to the data subset S is a feature (for example, a vector) composed of |V| values in the range of [0, 1].

[0092] In the embodiments of the present application, the mean field theory is introduced when the variational distribution is constructed, that is, it is assumed that each target object in the data subset S and each target object in the set V\S are independent of each other.

[0093] Further, the distance between the output result and the variational distribution can be measured based on at least one of the following ways: KL divergence, inverse KL divergence, alpha divergence, Wasserstein divergence, but is not limited thereto, and other measurement ways can also be used in other embodiments, and those skilled in the art can extend according to actual conditions, which is not limited in the embodiments of the present application.

[0094] In the embodiments of the present application, the distance between the output result and the variational distribution is measured by taking the KL divergence as an example, and the variational distribution q(S; ψ) can be approximated to p(S; θ) by minimizing the KL distance KL(q||p). Since KL(q||p) is non-negative, the variational lower bound can have the following representation:

[0095]

[0096] where, is the Shannon entropy, that is, the greater the uncertainty of the variable, the greater the entropy, and the greater the amount of information required.

[0097] ψ∈[0,1] |V| is the set function F θ is the multilinear extension of (S).

[0098] By maximizing the above variational lower bound, the KL distance between q(S; ψ) and p(S; θ) can be minimized, so that q(S; ψ) approximates p(S; θ).

[0099] Specifically, by optimizing the parameters θ in the variational model, the maximum variational lower bound and the optimal variational distribution q(S; ψ) can be calculated.

[0100] In the embodiment of the application, the variational model is optimized by the following method: obtaining training samples related to the type of the target object, the training samples including a first set of samples and a second set of samples, the second set of samples being a subset of the first set of samples; based on the training samples, iteratively optimizing the parameters of the variational model; wherein for each optimization, based on the variational model, calculating the variational parameters of the first set of samples, if based on the variational parameters and the second set of samples, it is determined that the training end condition is met, then the optimized variational model is obtained, if based on the variational parameters and the second set of samples, it is determined that the training end condition is not met, then the optimization of the parameters of the variational model is continued.

[0101] It can be understood that, assuming that the training samples are V is the first set of samples, S * is the second set of samples, S * is a data subset of V. Wherein each data subset of the first set of samples V can still be expressed as S, S * is the annotation result of the data subset sampling in the training samples.

[0102] In the embodiment of the application, the gradient ascent algorithm can be used to maximize the variational lower bound, and in actual application, the way that can be used is not limited thereto, and this algorithm should not be understood as a limitation of the application.

[0103] Specifically, continuing to take the KL divergence as an example to measure the distance between the above output result and the variational distribution, for ψ i :

[0104] The partial derivative of the multilinear extension with respect to it is:

[0105]

[0106] The partial derivative of the entropy term with respect to it is:

[0107]

[0108] Therefore, the fixed point equation of the variational lower bound is:

[0109]

[0110] where σ is an activation function, which can be a sigmoid activation function, but is not limited thereto.

[0111] That is, in the embodiment of the present application, the variational parameters of the set sample can be calculated based on the variational model in the following manner in combination with the fixed point equation of the variational lower bound: determining initial variational parameters; performing iterative updating of the initial variational parameters a predetermined number of times based on the variational model; and obtaining the variational parameters of the set sample according to the iterative updating result.

[0112] The process of performing iterative updating of the initial variational parameters a predetermined number of times based on the variational model specifically includes: processing the initial variational parameters based on the variational model to obtain intermediate variational parameters; and repeatedly performing the step of processing the intermediate variational parameters obtained in the last iteration based on the variational model until the predetermined number of times is reached.

[0113] Further, for each of the initial variational parameters and the intermediate variational parameters, the process of processing the variational parameter based on the variational model specifically includes: performing multi-linear expansion on the variational parameter to obtain a multi-linear expansion result; determining a partial differential result of the multi-linear expansion result; and performing activation processing on the partial differential result.

[0114] Specifically, the initial variational parameter ψ (0) can be initialized as ψ ∈ [0, 1] |V| That is, the initialization process of the initial variational parameter can include:

[0115] ψ (0) ← ψ ∈ [0, 1] |V|

[0116] Alternatively, ψ (0) can also be directly initialized as:

[0117] ψ (0) ← [1 / |V|] |V|

[0118] Further, the variational model can have the following representation:

[0119]

[0120] where 1≤k≤K, K is a positive integer, is a multi-linear expansion of the set function, is a partial differential of the multilinear expansion, and σ is an activation function.

[0121] It should be noted that the predetermined number K can be set by a person skilled in the art according to actual conditions, and the embodiments of the present application are not limited herein.

[0122] Then, based on the variational model, the process of performing the predetermined number of iteration updates on the initial variational parameter can include:

[0123]

[0124] That is, ψ (1) is obtained by using the variational model based on ψ (0) (the initial variational parameter), and ψ (2) Initially, the kth iteration update process is based on ψ (k-1) (the intermediate variational parameter obtained by the last processing), and ψ (k) is obtained by using the variational model.

[0125] Further, according to the iteration update result, the process of obtaining the variational parameter of the set sample can include:

[0126] ψ * = ψ (K)

[0127] As an example, taking the sigmoid activation function as an example, the parameter ψ of the variational distribution can be updated in the following manner:

[0128] ψ (0) ← [1 / |V|] |V|

[0129]

[0130] ψ * = ψ (K)

[0131] For ease of description, the above function (i.e., the variational model) similar to the RNN (Recurrent Neural Network) can be denoted as MFI(ψ;V,K). It can be seen that MFI(ψ;V,K) is a model (also can be understood as a differentiable function) about the parameter θ. Therefore, by optimizing the parameter θ, the maximum variational lower bound and the optimal variational distribution can be obtained.

[0132] In the embodiments of the present application, the optimization of the parameters of the variational model can be based on, but not limited to, at least one of the following training methods:

[0133] (1) Cross Entropy Loss;

[0134] (2) Constrastive Divergence;

[0135] (3) Noise Contrastive Estimation;

[0136] (4) Score Matching.

[0137] The following is an example of optimizing the parameter θ by minimizing the cross entropy loss function, which can be represented as:

[0138]

[0139] The process of optimizing the parameter θ based on the cross entropy loss can include:

[0140]

[0141] In the embodiments of the present application, the training end condition can include but is not limited to at least one of the following:

[0142] (1) reaching the maximum number of training times;

[0143] (2) loss convergence or invariance;

[0144] (3) the test effect of the training sample reaches the expectation.

[0145] Through the above training process, the purpose of end-to-end training of the energy model (by training the same parameter θ in the variational model as the energy model) can be achieved.

[0146] The following gives an algorithm example corresponding to the above training process:

[0147] Algorithm DiffMF(θ; K):

[0148] 1: Obtain training sample data

[0149]

[0150] 2: Calculate variational parameters

[0151] ψ * ←MFI(ψ; V, K)

[0152] 3: Calculate the loss function

[0153]

[0154] 4: Update the parameter θ

[0155]

[0156] where the algorithm

[0157] 1: Initialize variational parameter ψ

[0158] ψ (0) ← ψ ∈ [0, 1] |V|

[0159] 2: Perform the following for k = 1, …, K

[0160] 3: Update variational parameter

[0161]

[0162] 4: End

[0163] In the embodiments of the application, the average field variational inference is used to sample the data subsets, and the sampling method of the discrete energy model, such as Gibbs sampling, MH (Metropolis-Hasting) sampling, and the like, can also be used, but is not limited thereto.

[0164] The sampling method provided in the embodiments of the application is used to solve the problem that the sampling accuracy is reduced due to the influence of the arrangement order of the target objects in the set on the data subset sampling, and a set function learning method irrelevant to the order is proposed. The model is modeled as an energy model, and the network structure of the deep set function is introduced, so as to realize the probability model irrelevant to the input order of the target objects in the object screening process and the target object data set. Then, the average field variational inference is used to calculate the score of each data subset in the set, so as to select the data subset according to the score, and the end-to-end training of the energy model is realized. The inventors of the application have proved through a large number of tests that the algorithm can realize high subset sampling accuracy.

[0165] The object screening method provided in the embodiments of the application can be applied to many scenes requiring subset extraction, and some application scenes are exemplified below.

[0166] Scene one: commodity recommendation

[0167] For this scene, the application provides an object recommendation method, which comprises the following steps:

[0168] Step S201: acquiring a target object data set;

[0169] In the embodiments of the application, the target object refers to an object to be recommended, for example, a commodity, a service, a place, and the like to be recommended, but is not limited thereto. Taking the target object as a commodity as an example, the target object data set can be a commodity library.

[0170] Step S202: taking the to-be-recommended object dataset as the target object dataset, and using the object screening method provided in the above at least one embodiment to obtain the to-be-recommended objects screened from the to-be-recommended object dataset;

[0171] In the above at least one embodiment, the target object dataset is created for the to-be-recommended object dataset, which can be a set composed of vectors respectively created according to information of each to-be-recommended object in the to-be-recommended object dataset. Taking the to-be-recommended object as a commodity as an example, a vector corresponding to the commodity can be created according to information such as an identity (ID and / or name, etc.), attributes, and description information of the commodity, and the target object dataset can be obtained according to vectors of all commodities in the commodity library.

[0172] In the above at least one embodiment, taking the to-be-recommended object as a commodity as an example, the object screening method provided in the above at least one embodiment is used to realize the screening process irrelevant to the order of the commodities in the commodity library, and the selected commodities from the commodity library are consistent with the real choices of the general customers, thereby achieving high commodity screening precision.

[0173] Step S203: generating recommendation information based on the screened to-be-recommended objects.

[0174] Taking the to-be-recommended object as a commodity as an example, the selected commodities in the commodity library can be recommended to the corresponding objects. For example, taking the commodity as a toy as an example, when a request of browsing the toy library is received, toy commodities screened from the toy library by using the object screening method provided in the above at least one embodiment can be recommended and displayed, but the present application is not limited thereto.

[0175] Specifically, for this scenario, the training samples used to train the energy model can be related to the type of the to-be-recommended object. For example, a commodity recommendation dataset is used. It can be understood that different training samples will train energy models with different recommended contents, and a person skilled in the art can select appropriate training samples for training according to actual conditions. Based on the object screening method provided in the above at least one embodiment, the energy model can achieve good performance on the required commodity recommendation function after training.

[0176] The inventors of the present application tested the commodity recommendation provided in the above at least one embodiment on the disclosed commodity recommendation dataset. For a given commodity library, such as toys, furniture, and the like, the target is to make the commodities recommended by the energy model as consistent as possible with the real choices of the general customers.

[0177] Specifically, two indexes MJC and MAP are used as experimental indexes, and the average test results of commodity recommendation are shown in Table 1:

[0178]

[0179] Table 1

[0180] It can be seen that the energy model provided by the embodiments of the present application has better performance improvement compared with the probability greedy model in the prior art, which proves the effectiveness of the scheme of the present application.

[0181] Scenario two: picture detection and / or anomaly detection

[0182] For this scenario, the embodiments of the present application provide an image detection method, which comprises:

[0183] Step S301: obtaining an image data set;

[0184] In the embodiments of the present application, the images in the image data set are images of which the target features need to be detected, wherein the target features can refer to image anomaly features, can refer to different features between images, can refer to specified features of specific objects in images, etc. Taking the human face images in the image data set as an example, the target features can refer to the specified attributes of the human face, but are not limited thereto.

[0185] Step S302: taking the image data set as a target object data set, and using the object screening method provided in the above at least one embodiment to obtain target images with target features screened from the image data set;

[0186] Wherein, the target object data set is created for the image data set, which can be a matrix created according to the pixel point information of each image in the image data set to form a set.

[0187] In the embodiments of the present application, the object screening method provided in the above at least one embodiment is used to realize that the screening process is irrelevant to the order of the images in the image data set, to improve the possibility that the images screened from the image data set are target images with target features, and to achieve higher image detection accuracy.

[0188] Step S303: determining the target images as the detection result of the image data set.

[0189] Similarly, different training samples will train energy models with different detection contents, and those skilled in the art can select appropriate image training samples for training according to actual conditions. Based on the object screening sampling method provided in the embodiments of the present application, the energy model can achieve better performance on the required image detection function and / or anomaly detection function after training.

[0190] The inventors of the present application tested the picture detection and / or anomaly detection method provided in the embodiments of the present application on the public handwritten digital picture data set and the human face data set.

[0191] Wherein, for a handwritten digit picture dataset, the goal is to make the image detected by the energy model have the same data, for example, as shown in Figure 5 In this example, the image with the same number 38 is the image detection result.

[0192] For a face dataset, the goal is to make the energy model detect abnormal face images that do not have the target features. As an example, a normal face image needs to have two attributes of "with bangs" and "without beard", and then an abnormal face image that does not have these two attributes will be detected. In practical applications, the number and specific content of the attributes can be set according to actual conditions.

[0193] Alternatively, the abnormal face image detection can be for multiple target features at the same time. As an example, the abnormal image detection is for three target features: target feature 1: having two attributes of "with bangs" and "without beard"; target feature 2: having two attributes of "big nose" and "male"; target feature 3: having two attributes of "smile" and "black hair", then the normal face image is the face image that meets any of the above target features, and the abnormal face image is the face image that does not meet any of the three target features. In practical applications, the type and specific content of the target features can be set according to actual conditions.

[0194] Specifically, by taking the index MJC as the experimental index, the average test results of the image detection results and / or the abnormal detection results are shown in Table 2:

[0195]

[0196]

[0197] Table 2

[0198] It can be seen that the energy model provided by the embodiments of the present application has better performance improvement compared with the related art, which proves the effectiveness of the scheme of the present application.

[0199] Scenario three: compound selection

[0200] For this scenario, the embodiments of the present application provide a compound selection method, which comprises:

[0201] Step S401: obtaining a compound dataset;

[0202] In the embodiments of the present application, the compounds in the compound dataset are compounds to be predicted to have an effect on a predetermined disease, wherein the prediction mode of the compounds to have an effect on the predetermined disease can be set by those skilled in the art according to actual conditions, for example, predicting the activity on a target protein, but is not limited thereto.

[0203] Step S402: taking the compound dataset as a target object dataset, and using the object screening method provided in the above at least one embodiment to obtain target compounds related to the predetermined disease screened from the compound dataset;

[0204] The target object dataset is created for the compound dataset, and can be a set composed of vectors, matrices or sequences created according to the molecular, structural and other characteristics of each compound in the compound dataset.

[0205] In the embodiments of the present application, the object screening method provided in the above at least one embodiment is used to realize the screening process regardless of the order of the compounds in the compound dataset, improve the effect of the screened compounds in the compound dataset being related to the predetermined disease, and achieve higher compound selection accuracy.

[0206] Step S403: determining a drug for treating the predetermined disease based on the target compound.

[0207] The compound selection refers to a process of evaluating the biological activity, pharmacological effect and medicinal value of a substance that can be used as a compound by using appropriate methods, for example, the evaluation basis shown in Figure 6 The use of artificial intelligence to assist in compound selection in drug development has obvious advantages in time and cost.

[0208] The inventors of the present application tested the picture detection and / or anomaly detection provided in the embodiments of the present application on the disclosed loaded protein ligand complex dataset and binding database.

[0209] Specifically, the test results of the compound selection are shown in Table 3 by taking the index MJC as an experimental index.

[0210] Method Loading a protein ligand complex dataset Binding database Random 0.0725 0.0267 Prior art related technology 0.3499±0.0087 0.1760±0.0055 The present solution 0.3534±0.0143 0.1894±0.0021

[0211] Table 3

[0212] It can be seen that the energy model provided in the embodiments of the present application has better performance improvement compared with the related art, which proves the effectiveness of the present application.

[0213] The implementation of other details in the above scenarios and the resulting optimization effects can be referred to the introduction of the object screening method provided in the embodiments of the present application, which will not be described here.

[0214] Those skilled in the art should understand that the above several scenarios are only examples, and the use scenarios of various sampling methods with appropriate changes based on these examples can also be applied to the present application, so they should also be included in the protection scope of the present application.

[0215] In actual applications, the computer device with computing capability that executes the methods provided in the embodiments of the present application can be a terminal, a server, and the like. The terminal can be a notebook computer, a tablet computer, a desktop computer, a set-top box, a smart speaker, a mobile device (for example, a mobile phone, a portable music player, a personal digital assistant, a portable game device, a smart watch, a smart voice interaction device, a vehicle-mounted terminal, and the like), a smart home appliance (for example, but not limited to, a smart television), a smart robot, or a plant device, and the like, but is not limited thereto. The server can be a stand-alone physical server, a server cluster or a distributed system composed of multiple physical servers, or a cloud server that provides cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN (Content Delivery Network), and basic cloud computing services such as big data and artificial intelligence platforms.

[0216] Cloud computing refers to a delivery and use model of IT infrastructure, which refers to obtaining required resources in a scalable manner as needed through a network. Broadly, cloud computing refers to a delivery and use model of services, which refers to obtaining required services in a scalable manner as needed through a network. Such services can be IT and software, Internet related, or other services. Cloud computing is a product of the development and integration of traditional computer and network technologies such as grid computing, distributed computing, parallel computing, utility computing, network storage technologies, virtualization, load balancing, and the like.

[0217] With the development of the Internet, real-time data flow, and diversified connected devices, and the promotion of requirements such as search services, social networks, mobile commerce, and open collaboration, cloud computing has rapidly developed. Unlike previous parallel distributed computing, the generation of cloud computing will revolutionize the entire Internet model and enterprise management model in terms of concept.

[0218] Further, the methods provided in the embodiments of the present application can also be completed by a plurality of computer devices or devices with computing capability in cooperation, for example, different computer devices or devices each complete a part of the steps of the methods provided in the embodiments of the present application. As an example, the methods provided in the embodiments of the present application can be implemented in an environment as shown in FIG. 1, which is described with reference to FIG. 1. Figure 7 Figure 7 ​The implementation environment can include a server and a terminal. The terminal is connected to and interacts with the server through a communication network such as a Wi-Fi network or a cellular network. The server can also manage data through a data storage system.

[0219] In the embodiments of the present application, the computer device described above can also be one or more nodes in a blockchain network, or the data involved in the technical solutions provided by the embodiments of the present application can be saved in the blockchain.

[0220] Blockchain is a new application mode of distributed data storage, peer-to-peer transmission, consensus mechanism, and encryption algorithm. Blockchain is essentially a decentralized database, a series of data blocks associated using cryptographic methods, each containing information about a batch of network transactions, used to verify the validity of the information (anti-fake) and generate the next block. Blockchain can include a blockchain underlying platform, a platform product service layer, and an application service layer.

[0221] The blockchain underlying platform can include account management, basic services, smart contracts, and operation control processing modules. The account management module is responsible for the identity information management of all blockchain participants, including maintaining public and private key generation (account management), key management, and the maintenance of the correspondence between real identity information and blockchain addresses (permission management), and in the case of authorization, supervising and auditing the transaction of certain real identities, providing risk control rule configuration (risk audit); the basic service module is deployed on all blockchain node devices to verify the validity of business requests, and after consensus, the valid requests are recorded on the storage. For a new business request, the basic service first performs interface adaptation analysis and authentication processing (interface adaptation), then encrypts the business information through a consensus algorithm (consensus management), and after encryption, the complete and consistent transmission is transmitted to the shared ledger (network communication) and recorded and stored; the smart contract module is responsible for contract registration and issuance, contract triggering and contract execution. Developers can define contract logic through a certain programming language, publish it to the blockchain (contract registration), call keys or other events to trigger execution according to the logic of the contract terms, complete the contract logic, and also provide contract upgrade and cancellation functions; the operation control module is mainly responsible for deployment, configuration modification, contract setting, cloud adaptation during product release, and real-time state visualization output during product operation, such as alarms, network conditions, and node device health status.

[0222] The platform product service layer provides basic capabilities and implementation frameworks for typical applications. Developers can add business features based on these basic capabilities to complete the blockchain implementation of business logic. The application service layer provides application services based on the blockchain solution for business participants to use.

[0223] As introduced above, the technical solutions provided in the embodiments of the present application are related to big data (such as information recommendation, commodity recommendation, etc.), computer vision of artificial intelligence (such as picture detection and / or anomaly detection), machine learning and the like.

[0224] It should be noted that artificial intelligence is a comprehensive technology of computer science, involving a wide range of fields, both hardware and software technologies. Artificial intelligence basic technologies generally include technologies such as sensors, special artificial intelligence chips, cloud computing, distributed storage, big data processing technology, operation / interaction system, mechatronics and the like. Artificial intelligence software technology mainly includes computer vision technology, speech processing technology, natural language processing technology, and machine learning / deep learning, autonomous driving, intelligent transportation and the like.

[0225] Computer vision technology (Computer Vision, CV) is a science that studies how to make machines "see". Further, it refers to using cameras and computers to replace human eyes to identify and measure targets and further process images so that the computer processing becomes more suitable for human eye observation or image transmission to instrument detection. As a scientific discipline, computer vision researches related theories and technologies, trying to establish artificial intelligence systems that can obtain information from images or multidimensional data. Computer vision technology usually includes image processing, image recognition, image semantic understanding, image retrieval, OCR, video processing, video semantic understanding, video content / behavior recognition, three-dimensional object reconstruction, 3D technology, virtual reality, augmented reality, simultaneous localization and mapping, autonomous driving, intelligent transportation and the like. It also includes common face recognition, fingerprint recognition and other biometric identification technologies.

[0226] Big data refers to a collection of data that cannot be captured, managed and processed within a certain time frame using conventional software tools, and is a massive, high-growth and diversified information asset that requires new processing models to have stronger decision-making, insight discovery and process optimization capabilities. With the advent of the cloud era, big data has attracted more and more attention, and big data requires special technology to effectively process large amounts of data over time. Technologies suitable for big data include large-scale parallel processing databases, data mining, distributed file systems, distributed databases, cloud computing platforms, the Internet and scalable storage systems.

[0227] In addition, the technical solutions provided in the embodiments of the present application can also be applied to the field of intelligent transportation. Specifically, for the case of picture detection, anomaly detection and the like in intelligent transportation, the above object screening method can be used for detection, and further, the detection results can be displayed or other corresponding processing.

[0228] Intelligent Traffic System (ITS) is also called Intelligent Transportation System, which is a comprehensive transportation system that integrates advanced science and technology (information technology, computer technology, data communication technology, sensor technology, electronic control technology, automatic control theory, operations research, artificial intelligence, etc.) in transportation, service control and vehicle manufacturing, and strengthens the connection between vehicles, roads and users, so as to form a comprehensive transportation system that ensures safety, improves efficiency, improves environment and saves energy.

[0229] With the research and progress of artificial intelligence technology, artificial intelligence technology is researched and applied in many fields, such as common smart home, smart wearable device, virtual assistant, smart speaker, smart marketing, unmanned vehicle, autonomous vehicle, unmanned aerial vehicle, robot, smart medical treatment, smart customer service, Internet of Vehicles, autonomous driving, intelligent transportation, etc. It is believed that with the development of technology, artificial intelligence technology will be applied in more fields and play an increasingly important role.

[0230] The energy model-based data subset sampling method provided by the embodiment of the present application can design the energy function to have the property of permutation invariance through the energy model, so as to realize the property independent of the sampling process and order. In addition, in order to realize sampling of the energy model, the method of mean field variational estimation is proposed, which realizes the purpose of data subset sampling and end-to-end training. The energy model of the embodiment of the present application can achieve good sampling performance.

[0231] The embodiment of the present application provides an object screening device, as shown in the figure, the object screening device 80 can include: an acquisition module 801, a determination module 802 and a screening module 803, wherein, Figure 8

[0232] The acquisition module 801 is configured to acquire a target object data set, the target object data set including a plurality of target objects to be screened;

[0233] The determination module 802 is configured to determine probability distribution information of each data subset of the target object data set based on the constructed energy model, wherein the probability distribution information determined by the energy model for the data subset after the target object is sequentially transformed is the same for any data subset;

[0234] The screening module 803 is configured to determine at least one target data subset from each data subset according to the probability distribution information, so as to obtain an object screening result.

[0235] ​In an optional implementation, the probability of any data subset is proportional to a feature transformation result obtained by:

[0236] In an optional implementation, when determining the probability distribution information of each data subset of the target object data set being screened out based on the constructed energy model, the determining module 802 is specifically configured to:

[0237] determine a probability representation of each data subset being screened out, respectively;

[0238] exponentially normalize each probability representation to obtain the probability distribution information of each data subset being screened out.

[0239] In an optional implementation, when determining the probability distribution information of each data subset of the target object data set being screened out based on the constructed energy model, the determining module 802 is specifically configured to:

[0240] variational inference output results of the energy model for each data subset based on the optimized variational model to obtain the probability distribution information of each data subset being screened out, wherein the optimized variational model minimizes the distance between the probability distribution information and the output results.

[0241] In an optional implementation, the apparatus further includes an optimizing module 804 configured to optimize the variational model by:

[0242] obtain training samples related to the type of the target object, the training samples including a first set of samples and a second set of samples, the second set of samples being a subset of the first set of samples;

[0243] based on the training samples, iteratively optimize parameters of the variational model;

[0244] wherein for each optimization, based on the variational model, calculate variational parameters of the first set of samples, if based on the variational parameters and the second set of samples, it is determined that a training end condition is met, then the optimized variational model is obtained, and if based on the variational parameters and the second set of samples, it is determined that the training end condition is not met, then the optimization of the parameters of the variational model is continued.

[0245] In an optional implementation, when calculating the variational parameters of the set of samples based on the variational model, the optimizing module 804 is specifically configured to:

[0246] determine initial variational parameters;

[0247] based on the variational model, iteratively update the initial variational parameters for a predetermined number of times;

[0248] based on the result of the iterative update, obtain the variational parameters of the set of samples.

[0249] In an optional implementation, the optimization module 804, when performing the predetermined number of iterations of updating the initial variational parameter based on the variational model, is specifically configured to:

[0250] processing the initial variational parameter based on the variational model to obtain an intermediate variational parameter;

[0251] repeating the step of processing the intermediate variational parameter obtained in the last processing based on the variational model until the predetermined number of iterations is reached;

[0252] In the step of processing each of the initial variational parameter and the intermediate variational parameters based on the variational model, the optimization module 804 is specifically configured to:

[0253] performing a multi-linear expansion on the variational parameter to obtain a multi-linear expansion result;

[0254] determining a partial differential result of the multi-linear expansion result;

[0255] performing an activation processing on the partial differential result.

[0256] Embodiments of the present application provide an object recommendation device, which can include a first object obtaining module 501, a first screening module 502, and a recommendation module 503, wherein,

[0257] The first object obtaining module 501 is configured to obtain a to-be-recommended object data set;

[0258] The first screening module 502 is configured to take the to-be-recommended object data set as a target object data set, and use the object screening method provided in the embodiments of the present application to obtain a to-be-recommended object screened from the to-be-recommended object data set;

[0259] The recommendation module 503 is configured to generate recommendation information based on the screened to-be-recommended object.

[0260] Embodiments of the present application provide an image detection device, which can include a second obtaining module 601, a second screening module 602, and a detection module 603, wherein,

[0261] The second obtaining module 601 is configured to obtain an image data set;

[0262] The second screening module 602 is configured to take the image data set as a target object data set, and use the object screening method provided in the embodiments of the present application to obtain a target image with a target feature screened from the image data set;

[0263] The detection module 603 is configured to determine the target image as a detection result of the image data set.

[0264] This application provides a compound selection device 70, which may include: a third acquisition module 701, a third screening module 702, and a drug development module 703, wherein...

[0265] The third acquisition module 701 is used to acquire the compound dataset;

[0266] The third screening module 702 is used to take the compound dataset as the target object dataset and use the object screening method provided in the embodiments of this application to obtain the target compounds related to the predetermined disease screened from the compound dataset;

[0267] Drug development module 703 is used to identify drugs for treating predetermined diseases based on target compounds.

[0268] Each device in the embodiments of this application can execute the method provided in the embodiments of this application, and their implementation principles are similar. The actions performed by each module in the devices of each embodiment of this application correspond to the steps in the methods of each embodiment of this application. For detailed functional descriptions of each module of the device, please refer to the descriptions in the corresponding methods shown above, which will not be repeated here.

[0269] This application provides an electronic device including a memory, a processor, and a computer program stored in the memory. The processor executes the computer program to implement the steps of the foregoing method embodiments. Optionally, the electronic device may refer to the aforementioned server, or it may refer to the aforementioned terminal.

[0270] In one alternative embodiment, an electronic device is provided, such as Figure 9 As shown, Figure 9 The illustrated electronic device 900 includes a processor 901 and a memory 903. The processor 901 and the memory 903 are connected, for example, via a bus 902. Optionally, the electronic device 900 may further include a transceiver 904, which can be used for data interaction between the electronic device and other electronic devices, such as sending and / or receiving data. It should be noted that in practical applications, the transceiver 904 is not limited to one type, and the structure of the electronic device 900 does not constitute a limitation on the embodiments of this application.

[0271] The processor 901 can be a CPU (Central Processing Unit), a general-purpose processor, a DSP (Digital Signal Processor), an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array) or other programmable logic device, transistor logic device, hardware component, or any combination thereof. It can implement or execute the various exemplary logical blocks, modules and circuits described in connection with the disclosure. The processor 901 can also be a combination of computing functions, such as a combination of one or more microprocessors, a combination of a DSP and a microprocessor, etc.

[0272] The bus 902 can include a path for transmitting information between the above-mentioned components. The bus 902 can be a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus, etc. The bus 902 can be divided into an address bus, a data bus, a control bus, etc. For convenience of representation, Figure 9 In the figure, only one thick line is used to represent the bus, but it does not mean that there is only one bus or only one type of bus.

[0273] The memory 903 can be a ROM (Read Only Memory) or other type of static storage device that can store static information and instructions, a RAM (Random Access Memory) or other type of dynamic storage device that can store information and instructions, an EEPROM (Electrically Erasable Programmable Read Only Memory), a CD-ROM (Compact Disc Read Only Memory) or other optical disk storage, an optical disk storage (including a compact disk, a laser disk, an optical disk, a digital versatile disk, a Blu-ray disk, etc.), a magnetic disk storage medium, other magnetic storage device, or any other medium that can be used to carry or store computer programs and can be read by a computer, without limitation.

[0274] The memory 903 is configured to store a computer program for implementing the embodiments of the present application, and the processor 901 is configured to control the execution of the computer program stored in the memory 903. The processor 901 is configured to execute the computer program stored in the memory 903 to implement the steps shown in the foregoing method embodiments.

[0275] The electronic device includes, but is not limited to, a terminal, a server, and the like. The terminal can be a notebook computer, a tablet computer, a desktop computer, a set-top box, a smart speaker, a mobile device (for example, a mobile phone, a portable music player, a personal digital assistant, a portable game device, a smart watch, a smart voice interaction device, a vehicle terminal, and the like), a smart home appliance (for example, but not limited to, a smart television), a smart robot, or a plant terminal, but is not limited thereto. The server can be a stand-alone physical server, a server cluster or a distributed system composed of multiple physical servers, or a cloud server providing cloud computing services.

[0276] The embodiments of the present application provide a computer readable storage medium, which stores a computer program. The computer program is executed by a processor to implement the steps and corresponding contents of the foregoing method embodiments.

[0277] The embodiments of the present application also provide a computer program product, which includes a computer program. The computer program is executed by a processor to implement the steps and corresponding contents of the foregoing method embodiments.

[0278] The terms "first", "second", "1", "2", and the like in the specification and claims of the present application and the above-described drawings are used to distinguish similar objects, and do not necessarily indicate a specific order or sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments of the present application described herein can be implemented in an order other than that shown or described.

[0279] It should be understood that, although the flowcharts of the embodiments of the present application indicate the respective operation steps by arrows, the implementation order of the steps is not limited to the order indicated by the arrows. Unless otherwise specified herein, the implementation steps in each flowchart can be executed in other orders as required in some implementation scenarios of the embodiments of the present application. In addition, part or all of the steps in each flowchart can include multiple sub-steps or multiple stages based on the actual implementation scenario. Part or all of the sub-steps or stages can be executed at the same time, and each of the sub-steps or stages can also be executed at different times. In the scenario where the execution times are different, the execution order of the sub-steps or stages can be flexibly configured according to requirements, and the embodiments of the present application do not limit this.

[0280] The above are only optional implementations of some implementation scenarios of the present application. It should be noted that, for those skilled in the art, other similar implementation manners based on the technical concept of the present application without departing from the technical concept of the present application also belong to the protection scope of the embodiments of the present application.

Claims

1. An object filtering method, characterized in that, include: Obtain a target object dataset, which includes multiple target objects to be filtered. The target object dataset is a product library, an image dataset, or a compound dataset. The product information in the product library is created based on the product's corresponding identity, attributes, and / or description information. Based on the constructed energy model, the probability distribution information of each data subset of the target object dataset being selected is determined. Each target object in the target object dataset has two possibilities: being selected or not being selected. The data subsets are obtained based on various possible combinations. For any data subset, the probability distribution information determined by the data subset after the target object is sequentially transformed based on the energy model is the same. Based on the probability distribution information, at least one target data subset is determined from each of the data subsets to obtain the object screening result; The determination of the probability distribution information for each data subset of the target object dataset selected based on the constructed energy model includes: The energy model is used to perform variational inference on the output of the energy model for each data subset based on the optimized variational model, thereby obtaining the probability distribution information of each data subset. The optimized variational model minimizes the distance between the probability distribution information and the output.

2. The object screening method according to claim 1, characterized in that, The probability of any subset of data is proportional to the feature transformation result obtained in the following manner: Each target object in any subset of data is mapped to a vector of the first dimension to obtain the first feature transformation result of any subset of data. The number of target objects in any subset of data is used as the second dimension of the vector, and the first feature transformation result is fused based on the second dimension to obtain the fusion result; The fusion result is subjected to a nonlinear transformation to obtain the second feature transformation result.

3. The object screening method according to claim 1, characterized in that, The determination of the probability distribution information for each data subset of the target object dataset selected based on the constructed energy model includes: Determine the probability representation of each data subset being selected; By exponentially normalizing each of the probability representations, the probability distribution information of each selected data subset is obtained.

4. The object screening method according to claim 1, characterized in that, The variational model is optimized in the following way: Obtain training samples related to the type of the target object, the training samples including a first set of samples and a second set of samples, the second set of samples being a subset of the first set of samples; Based on the training samples, the parameters of the variational model are iteratively optimized; In each optimization, based on the variational model, the variational parameters of the first set of samples are calculated. If the training termination condition is met based on the variational parameters and the second set of samples, the optimized variational model is obtained. If the training termination condition is not met based on the variational parameters and the second set of samples, the parameters of the variational model are further optimized.

5. The object screening method according to claim 4, characterized in that, The step of calculating the variational parameters of the first set of samples based on the variational model includes: Determine the initial variational parameters; Based on the variational model, the initial variational parameters are iteratively updated a predetermined number of times; Based on the iterative update results, the variational parameters of the first set of samples are obtained.

6. The object screening method according to claim 5, characterized in that, The step of iteratively updating the initial variational parameters based on the variational model a predetermined number of times includes: Based on the variational model, the initial variational parameters are processed to obtain intermediate variational parameters; Repeat the steps of processing the intermediate variational parameters obtained from the previous processing based on the variational model until the predetermined number of times is reached; Specifically, for each variational parameter among the initial variational parameters and each intermediate variational parameter, the variational parameter is processed based on the variational model, including: The variational parameter is then subjected to a multilinear extension to obtain the multilinear extension result; Determine the partial derivative of the multilinear extension result; The partial derivative result is then activated.

7. An object screening device, characterized in that, include: The acquisition module is used to acquire a target object dataset, which includes multiple target objects to be filtered. The target object dataset is a product library, an image dataset, or a compound dataset. The product information in the product library is created based on the product's corresponding identity identifier, attribute, and / or description information. The determination module is used to determine the probability distribution information of each data subset of the target object dataset being selected based on the constructed energy model. Each target object in the target object dataset has two possibilities: being selected or not being selected. The data subsets are obtained based on various possible combinations. For any data subset, the probability distribution information determined by the data subset after the target object is sequentially transformed based on the energy model is the same. The filtering module is used to determine at least one target data subset from each data subset based on the probability distribution information, so as to obtain the object filtering result; When the determining module is used to determine the probability distribution information of each data subset of the target object dataset selected based on the constructed energy model, it is specifically used for: The energy model is used to perform variational inference on the output of the energy model for each data subset based on the optimized variational model, thereby obtaining the probability distribution information of each data subset. The optimized variational model minimizes the distance between the probability distribution information and the output.

8. The object screening device according to claim 7, characterized in that, The probability of any subset of data is proportional to the feature transformation result obtained in the following manner: Each target object in any subset of data is mapped to a vector of the first dimension to obtain the first feature transformation result of any subset of data. The number of target objects in any subset of data is used as the second dimension of the vector, and the first feature transformation result is fused based on the second dimension to obtain the fusion result; The fusion result is subjected to a nonlinear transformation to obtain the second feature transformation result.

9. The object screening device according to claim 7, characterized in that, When determining the probability distribution information of each data subset of the target object dataset based on the constructed energy model, the determining module 802 is specifically used for: Determine the probability representation of each data subset being selected; By exponentially normalizing each of the probability representations, the probability distribution information of each selected data subset is obtained.

10. The object screening device according to claim 7, characterized in that, It also includes an optimization module for optimizing the variational model in the following ways: Obtain training samples related to the type of the target object, the training samples including a first set of samples and a second set of samples, the second set of samples being a subset of the first set of samples; Based on the training samples, the parameters of the variational model are iteratively optimized; In each optimization, based on the variational model, the variational parameters of the first set of samples are calculated. If the training termination condition is met based on the variational parameters and the second set of samples, the optimized variational model is obtained. If the training termination condition is not met based on the variational parameters and the second set of samples, the parameters of the variational model are further optimized.

11. The object screening device according to claim 10, characterized in that, When the optimization module is used to calculate the variational parameters of the first set of samples based on the variational model, it is specifically used for: Determine the initial variational parameters; Based on the variational model, the initial variational parameters are iteratively updated a predetermined number of times; Based on the iterative update results, the variational parameters of the first set of samples are obtained.

12. The object screening device according to claim 11, characterized in that, When the optimization module is used to iteratively update the initial variational parameters a predetermined number of times based on the variational model, it is specifically used for: Based on the variational model, the initial variational parameters are processed to obtain intermediate variational parameters; Repeat the steps of processing the intermediate variational parameters obtained from the previous processing based on the variational model until the predetermined number of times is reached; Specifically, when the optimization module processes each variational parameter based on the variational model for the initial variational parameter and each intermediate variational parameter, it is used as follows: The variational parameter is then subjected to a multilinear extension to obtain the multilinear extension result; Determine the partial derivative of the multilinear extension result; The partial derivative result is then activated.

13. An electronic device comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method according to any one of claims 1-6.

14. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1-6.

15. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1-6.

Images