1. A neural and symbolic based multi-modal big data machine auto-learning system, the system comprising: the system comprises a characteristic engineering automatic construction module, a mechanism model automatic construction module, a hyper-parameter optimizing and optimizing module and a model data processing module;

the automatic characteristic engineering construction module comprises a data acquisition unit, an automatic case hypergraph network construction unit and an automatic network structure updating unit;

the data acquisition unit is configured to acquire fragmented, multi-modal and dynamically-evolved big data as input data; the input data comprises text, image, audio and video data of different modes in a plurality of fields;

the automatic construction unit of the affair hypergraph network is configured to extract the concept symbols of affairs in different fields contained in the input data through a self-coding neural network, and construct a affair concept symbol space; based on the matter concept symbol space, automatically aggregating the super edges of different modes by a multi-mode feature automatic conformal representation method, and performing conformal calculation to generate a matter concept hypergraph network;

the network structure automatic updating unit is configured to extract concept symbols of affairs from new input data and obtain newly added hypergraph vertexes; performing increase and decrease alignment convolution calculation on newly added hypergraph vertexes and nodes covered by the affair hypergraph network through the hypercide multi-mode dynamic evolution calculation, and further realizing automatic updating of the affair hypergraph network;

the mechanism model automatic construction module comprises a field task definition unit, a model game design search unit and a search grid acceleration optimization unit;

the domain task definition unit is configured to divide concept symbols and hyper-parameter distribution conditions of the affairs of a set domain by combining the prior of nerve + symbol feature engineering aiming at an input data set of the domain, and pre-construct a model structure and a hyper-parameter search space; the hyper-parameters include: learning rate, the number of neural network layers, network structure, convolution kernel selection, iteration times, the number of hidden layer layers, neuron scale, sliding window and popular commonality index, activation function, the number of clusters and the number of topics;

the model game design searching unit is configured to distill a series of model structure candidate models from nodes of a rational hypergraph network through a game tree searching strategy in a pre-constructed model structure and a hyper-parameter searching space, wherein the model structure candidate models comprise a neural network prediction algorithm and a group of hyper-parameters, and the model game design searching unit is used for carrying out iterative high-fidelity evaluation on the series of candidate model structures and the hyper-parameter set and automatically searching out an optimal mechanism model most suitable for a current input data set; the mechanism model comprises a neural network model for description, prediction and early warning;

the search grid acceleration optimization unit is configured to obtain a learning curve corresponding to a model structure through a neural network structure prediction technology based on reinforcement learning, fit a model trained from a training sample set and automatically generate a model error minimum variance mean weight from a test sample set according to model structure distribution and a hyper-parameter process weight predicted by the learning curve, and globally share the weight in each model structure; the training sample set is a sample set constructed based on concept symbols of affairs in the affairs hypergraph network; the test sample set is an input data set of the set field;

the hyper-parameter optimizing module comprises a hyper-parameter space construction unit, a hyper-parameter self-adaptive selection strategy unit, a self-adaptive optimizing reasoning unit and a hyper-parameter automatic migration unit;

the hyper-parameter space construction unit is configured to divide the hyper-parameter data of the existing automatic machine learning algorithm into hyper-parameter populations of different automatic machine learning algorithms based on the hyper-parameter data of the existing automatic machine learning algorithm, and further construct a hyper-parameter space;

the hyper-parameter adaptive selection strategy unit is configured to select a multi-type candidate hyper-parameter set which meets the learning target task from the hyper-parameter space by taking concept symbols of affairs of each field contained in the updated affair hyper-graph network as the learning target task and combining the prior knowledge of each field and a predefined hyper-parameter adaptive selection strategy function;

the self-adaptive optimization searching and reasoning unit is configured to iteratively search a super-parameter combination of an optimal structure of a candidate algorithm and learning rate, regularization and network structure depth by adopting a parallel and sequence combined method according to a learning target task through a self-adaptive optimization searching and reasoning algorithm based on the candidate super-parameter set, search the super-parameter combination once to generate a super-parameter optimal curve, automatically compare the variation of the super-parameter optimal curve generated for many times, increase interference information until the variation exceeds a threshold value, terminate the self-adaptive optimization searching, and obtain an optimal super-parameter combination;

the automatic migration unit is configured to perform similar matching on the newly added learning target task and the existing learning target task, migrate the hyper-parameter combination corresponding to the existing learning target task with the type similarity higher than the preset threshold value to the hyper-parameter space of the newly added learning target task, and configure the optimal hyper-parameter for the newly added learning target task;

the model data processing module is configured to process the input data sets of all the set fields by combining the optimal mechanism model screened by the mechanism model automatic construction module and the optimal hyper-parameter combination obtained by the hyper-parameter optimization module; the processing comprises description, early warning and prediction.

2. The system according to claim 1, wherein the method comprises the steps of automatically aggregating the super edges of different modes based on the concept-concept symbol space by a multi-modal automatic conformal feature representation method to perform conformal computation to generate a concept-hypergraph network, and the method comprises the following steps:

a10, adopting multi-mode feature selection calculation to solve the matter concept symbol space, and extracting a finite node set of the matter concept symbol in the input data;

a20, taking a finite node set of concept symbols of the affairs as a vertex set of a hypergraph, representing learning calculation and solving through the hypergraph to generate a hyperedge set, giving a weight and a time stamp to each hyperedge, and generating Laplace matrixes of the hyperedges of a plurality of modes to obtain a hyperedge structure group of the plurality of modes;

and A30, performing high-order correlation conformal entropy solving calculation among hypergraphs of different modes on the Laplacian matrix of the hypergraphs of the multiple modes to generate a case hypergraph network.

3. The system according to claim 2, wherein the finite node set of concept symbols of the affairs in the input data is extracted by multi-modal feature selection computation, and the method comprises:

wherein the content of the first and second substances,the symbol space of the concept of the matter is represented,a characteristic D dimension tensor representing the ith concept symbol, m represents the number of modes in the input data, N represents the total N sample sequences of the input,a corresponding a priori vector representing the input data,a tensor coefficient matrix representing the ith mode, and storing all the multi-mode tensor coefficient vectors obtained currently in；

Performing objective function of multi-modal feature selection calculationNormalized generalization calculation of norm to obtainMultimodal rank conversion, forming a finite set of nodes of m modal feature vectors, i.e. things in the input dataFinite set of nodes of concept symbols of a theory。

4. The system according to claim 3, wherein the method comprises the following steps of calculating and solving a computation through hypergraph representation to generate a set of hyperedges, and assigning a weight and a time stamp to each hyperedge to generate a Laplace matrix of a plurality of modal hypergraphs:

in a finite set of nodesIn the method, a k-means clustering algorithm is adopted, and one characteristic vertex is selected at willAs a central node, continuously calculating the central node and other nodesThe central node is connected with other adjacent m-1 vertexes to construct N super edges, and each super edge is given weightAnd time stampAnd defining a hypergraph incidence matrix:

wherein the content of the first and second substances,to representNode pointTo the central nodeThe distance of (a) to (b),indicates the currentThe Euclidean distance between the corresponding vertexes;

hypergraph-based vertex angle diagonal matrixGenerating Laplace half-angle matrix of hypergraphWherein, in the step (A),is a matrix of the vertex degrees of the hypergraph,the degree matrix of the super edge of the hypergraph, H is the incidence matrix of the hypergraph, and W is the system matrix of the hypergraph super edge weight.

5. The automatic learning system of multi-modal big data machine based on nerve and symbol as claimed in claim 2, wherein the "Laplace matrix of multi-modal hypergraphs is subjected to high-order correlation conformal entropy solving calculation among hypergraphs of different modalities to generate a case hypergraph network", and the method comprises:

solving the first conformal entropy and the second conformal entropy; the first conformal entropy is hypergraph conformal entropy of m modes; the second conformal entropy is a conformal entropy when m = 2;

if the first conformal entropy is larger than the second conformal entropy, combining hypergraphs of different modalities into a theme; if the first conformal entropy is less than the second conformal entropy, the hypergraph stands alone as a topic;

continuously carrying out iterative computation for multiple times to generate a matter hypergraph network;

the conformal entropy calculation method comprises the following steps:

wherein the content of the first and second substances,a higher order correlation conformal entropy function between hypergraphs representing m modes,is a curved surface modal joint distribution entropy used for representing the conformal probability of each hyper-edge between hyper-graphs,is shown asA hypergraph of the characteristics of the individual modes,to representJoint probability distribution of individual modal feature matrices.

6. The automatic learning system of multimodality big data machine based on nerve and symbol as claimed in claim 2, wherein the automatic updating of the hypergraph network is realized by performing addition and subtraction alignment convolution calculation on newly added hypergraph vertices and nodes covered by the hypergraph network through the dynamic evolution calculation of the hyperedge multimodality, and the method comprises:

for the newly added input data stream, if newly added isolated nodes are obtained through A10, the complementary entropy of the isolated nodes and the nodes covered by the event hypergraph network obtained by the event hypergraph network automatic construction unit is calculated, and the isolated nodes are merged to the hypergraph corresponding to the nodes of which the complementary entropy is smaller than the set threshold;

if a new peak and a new super edge are obtained from the newly added data stream training sample, carrying out the convolution calculation of nodes of the super graph on the newly added super graph peak and the existing super graph peak according to the step A20 to complete the increase and decrease processing of the super graph peak; and C, carrying out hypergraph edge convolution calculation on the newly-added hypergraph edge and the existing hypergraph edge according to the step A30, and updating the matter hypergraph network.

7. The neural and symbolic based multimodal big data machine automatic learning system of claim 1, wherein the hyper-parameter adaptive selection policy functionComprises the following steps:

wherein the content of the first and second substances,for measuring having candidate hyper-parametersIs calculated byIn the hyper-parameter spaceAnd learning a target task datasetA denotes an algorithm in the hyper-parameter space.

8. The neural and symbolic based multimodal big data machine automatic learning system of claim 7, wherein the adaptive evolutionary algorithm objective function y is:

wherein the content of the first and second substances,representing a hyper-parameter selection policy functionThe combined optimal curve function of (a) is,representing selection of a set of candidate hyperparameters according to a hyperparameter selection policyAnd the medium screen is selected from an adaptive optimization-searching training function, the adaptive optimization-searching training function is combined with data samples obtained in real time, optimal hyper-parameters are automatically set for each algorithm, the data samples are derived from learning target tasks, and c represents the number of hyper-parameter combinations of the ith historical data sample.

9. The system according to claim 8, wherein the objective function of the auto-migration unit in auto-migration learning is:

wherein the content of the first and second substances,for newly adding a hyper-parameter variable of a learning target task domain,representing the target prediction function corresponding to the algorithm in the hyperparametric space,representing hyper-parametersThe combined optimal curve function of (a) is,the weight of the migration is represented as,representing selection strategies with optimal hyper-parametersThe set of hyper-parameters of (a),the number of sets of source algorithms is represented,a loop iteration counter is shown and is,represent iteration ofA secondary source algorithm; the source algorithm is an existing hyper-parameter algorithm corresponding to the learning target task.

Background

Big data has become a core element of resource allocation and optimization in the fields of global industrial production, circulation, distribution, consumption activity, economic operation and the like. The method for exploring the cognitive big data is an important research direction in the field of artificial intelligence, and the independent operation mechanism of the real-world complex system and the motion tracks of dependence, competition, association and the like between the complex system and the environment are recorded by the big data at the end. The knowledge has very important research value for scientifically, timely and accurately mastering national economic development, optimizing industrial structure, promoting social scientific treatment and the like. However, the conventional machine learning method requires collection of data by a large amount of professional data analysis scientists.

However, data collected by data analysis scientists are native to complex systems, and characteristics such as nonlinearity, emergence, spontaneous order, adaptability and feedback loops are derived, so that the existing artificial intelligence reduction theory (such as statistical machine learning, bayesian network, neural network and the like) is difficult to effectively explain the universality law of a big data implicit system. Therefore, obtaining the system operation universality law from big data becomes an important direction for artificial intelligence cognition research, specifically: not only can a multi-modal data distribution rule of complex system operation be learned from a perception and observation angle, but also the potential risk or development trend of system operation needs to be deduced from a newly-added data stream from a system operation mechanism cognition angle, namely: big data cognition modeling analysis becomes one of the development leading subjects in the field of artificial intelligence 3.0.

The existing big data driving system modeling analysis mainly focuses on three aspects: 1. big data modeling analysis based on unsupervised generation learning, the method focuses on timely obtaining system data characteristic attributes from a complex system, and vividly describes a system operation mechanism of a specific time region through time-space correlation among learning system operation data; 2. recognizing a data space attraction rule based on space-time structure measurement, adopting a basic rule of the space-time structure measurement in the attribute motion of a specific scene, and carrying out macroscopic and microscopic combined super-geometric description and behavior trend prediction, wherein the basic rule is concise and rich in insights, and the basic rule of system mechanism dynamic evolution is disclosed by means of prior knowledge such as common knowledge; 3. the model automatic construction method based on the automatic machine learning type continuously samples the data stream generated by the system, and searches a model adaptive to a specific service scene from a high-quality training sample space through multiple iterations, thereby realizing the automatic modeling of the system. However, the three methods are still based on a system reduction theory, and a system mechanism is reduced from big data by means of a machine learning tool, so that the idea cannot simply understand the whole system operation mechanism as the sum of individuals, cannot effectively infer the determination property of a system operation mechanism model from the overall behavior, and increases the model training calculation complexity. Based on the neural and symbolic based multi-modal big data machine automatic learning system, the invention provides the neural and symbolic based multi-modal big data machine automatic learning system.

Disclosure of Invention

In order to solve the above problems in the prior art, that is, to solve the problem that the existing machine learning method cannot automatically fragment dynamic evolution data to obtain a high fidelity mechanism model, the first aspect of the present invention provides a neural and symbolic based multi-modal big data machine automatic learning system, which includes: the system comprises a characteristic engineering automatic construction module, a mechanism model automatic construction module, a hyper-parameter optimizing and optimizing module and a model data processing module;

the automatic characteristic engineering construction module comprises a data acquisition unit, an automatic case hypergraph network construction unit and an automatic network structure updating unit;

the data acquisition unit is configured to acquire fragmented, multi-modal and dynamically-evolved big data as input data; the input data comprises text, image, audio and video data of different modes in a plurality of fields;

the automatic construction unit of the affair hypergraph network is configured to extract the concept symbols of affairs in different fields contained in the input data through a self-coding neural network, and construct a affair concept symbol space; based on the matter concept symbol space, automatically aggregating the super edges of different modes by a multi-mode feature automatic conformal representation method, and performing conformal calculation to generate a matter concept hypergraph network;

the network structure automatic updating unit is configured to extract concept symbols of affairs from new input data and obtain newly added hypergraph vertexes; performing increase and decrease alignment convolution calculation on newly added hypergraph vertexes and nodes covered by the affair hypergraph network through the hypercide multi-mode dynamic evolution calculation, and further realizing automatic updating of the affair hypergraph network;

the mechanism model automatic construction module comprises a field task definition unit, a model game design search unit and a search grid acceleration optimization unit;

the domain task definition unit is configured to divide concept symbols and hyper-parameter distribution conditions of the affairs of a set domain by combining the prior of nerve + symbol feature engineering aiming at an input data set of the domain, and pre-construct a model structure and a hyper-parameter search space; the hyper-parameters include: learning rate, the number of neural network layers, network structure, convolution kernel selection, iteration times, the number of hidden layer layers, neuron scale, sliding window and popular commonality index, activation function, the number of clusters and the number of topics;

the model game design searching unit is configured to distill a series of model structure candidate models from nodes of a rational hypergraph network through a game tree searching strategy in a pre-constructed model structure and a hyper-parameter searching space, wherein the model structure candidate models comprise a neural network prediction algorithm and a group of hyper-parameters, and the model game design searching unit is used for carrying out iterative high-fidelity evaluation on the series of candidate model structures and the hyper-parameter set and automatically searching out an optimal mechanism model most suitable for a current input data set; the mechanism model comprises a neural network model for description, prediction and early warning;

the search grid acceleration optimization unit is configured to obtain a learning curve corresponding to a model structure through a neural network structure prediction technology based on reinforcement learning; according to the model structure distribution and the hyper-parameter process weight predicted by the learning curve, fitting a model trained from a training sample set and a model error minimum variance mean weight automatically generated from a test sample set, and globally sharing the weights in each model structure; the training sample set is a sample set constructed based on concept symbols of affairs in the affairs hypergraph network; the test sample set is an input data set of the set field;

the hyper-parameter optimizing module comprises a hyper-parameter space construction unit, a hyper-parameter self-adaptive selection strategy unit, a self-adaptive optimizing reasoning unit and a hyper-parameter automatic migration unit;

the hyper-parameter space construction unit is configured to divide the hyper-parameter data of the existing automatic machine learning algorithm into hyper-parameter populations of different automatic machine learning algorithms based on the hyper-parameter data of the existing automatic machine learning algorithm, and further construct a hyper-parameter space;

the hyper-parameter adaptive selection strategy unit is configured to select a multi-type candidate hyper-parameter set which meets the learning target task from the hyper-parameter space by taking concept symbols of affairs of each field contained in the updated affair hyper-graph network as the learning target task and combining the prior knowledge of each field and a predefined hyper-parameter adaptive selection strategy function;

the self-adaptive optimization searching and reasoning unit is configured to iteratively search a super-parameter combination of an optimal structure of a candidate algorithm and learning rate, regularization and network structure depth by adopting a parallel and sequence combined method according to a learning target task through a self-adaptive optimization searching and reasoning algorithm based on the candidate super-parameter set, search the super-parameter combination once to generate a super-parameter optimal curve, automatically compare the variation of the super-parameter optimal curve generated for many times, increase interference information until the variation exceeds a threshold value, terminate the self-adaptive optimization searching, and obtain an optimal super-parameter combination;

the automatic migration unit is configured to perform similar matching on the newly added learning target task and the existing learning target task, migrate the hyper-parameter combination corresponding to the existing learning target task with the type similarity higher than the preset threshold value to the hyper-parameter space of the newly added learning target task, and configure the optimal hyper-parameter for the newly added learning target task;

the model data processing module is configured to process input data of each set field by combining an optimal mechanism model screened by the mechanism model automatic construction module and an optimal hyper-parameter combination obtained by the hyper-parameter optimization module; the processing comprises description, early warning and prediction.

In some preferred embodiments, the method for generating the event hypergraph network by automatically aggregating the hyperedges of different modalities and performing conformal calculation based on the event conceptual symbol space by using a multi-modal feature automatic conformal representation method includes:

a10, adopting multi-mode feature selection calculation to solve the matter concept symbol space, and extracting a finite node set of the matter concept symbol in the input data;

a20, taking a finite node set of concept symbols of the affairs as a vertex set of a hypergraph, representing learning calculation and solving through the hypergraph to generate a hyperedge set, giving a weight and a time stamp to each hyperedge, and generating Laplace matrixes of the hyperedges of a plurality of modes to obtain a hyperedge structure group of the plurality of modes;

and A30, performing high-order correlation conformal entropy solving calculation among hypergraphs of different modes on the Laplacian matrix of the hypergraphs of the multiple modes to generate a case hypergraph network.

In some preferred embodiments, "extracting a finite set of nodes of concept symbols of a case in the input data by using a multi-modal feature selection computation solution" is performed by:

wherein the content of the first and second substances,the symbol space of the concept of the matter is represented,a characteristic D dimension tensor representing the ith concept symbol, m represents the number of modes in the input data, N represents the total N sample sequences of the input,a corresponding a priori vector representing the input data,a tensor coefficient matrix representing the ith mode, and storing all the multi-mode tensor coefficient vectors obtained currently in；

Performing objective function of multi-modal feature selection calculationNormalized generalization calculation of norm to obtainMulti-modal rank conversion, forming a finite set of nodes of m modal feature vectors, i.e., a finite set of nodes of conceptual symbols of the events in the input data。

In some preferred embodiments, "solve by hypergraph representation learning calculation, generate a set of hyperedges, and assign a weight and a timestamp to each hyperedge, generate a laplacian matrix of multiple modal hypergraphs", by:

in a finite set of nodesIn the method, a k-means clustering algorithm is adopted, and one characteristic vertex is selected at willAs a central node, continuously calculateCenter node and other nodesThe central node is connected with other adjacent m-1 vertexes to construct N super edges, and each super edge is given weightAnd time stampAnd defining a hypergraph incidence matrix:

wherein the content of the first and second substances,representing nodesTo the central nodeThe distance of (a) to (b),indicates the currentThe Euclidean distance between the corresponding vertexes;

hypergraph-based vertex angle diagonal matrixGenerating Laplace half-angle matrix of hypergraphWherein, in the step (A),is a matrix of the vertex degrees of the hypergraph,the degree matrix of the super edge of the hypergraph, H is the incidence matrix of the hypergraph, and W is the system matrix of the hypergraph super edge weight.

In some preferred embodiments, for laplacian matrices of hypergraphs in multiple modalities, performing high-order correlation conformal entropy solving calculation between hypergraphs in different modalities to generate a case hypergraph network, the method includes:

solving the first conformal entropy and the second conformal entropy; the first conformal entropy is hypergraph conformal entropy of m modes; the second conformal entropy is a conformal entropy when m = 2;

if the first conformal entropy is larger than the second conformal entropy, combining hypergraphs of different modalities into a theme; if the first conformal entropy is less than the second conformal entropy, the hypergraph stands alone as a topic;

continuously carrying out iterative computation for multiple times to generate a matter hypergraph network;

the conformal entropy calculation method comprises the following steps:

，

wherein the content of the first and second substances,a higher order correlation conformal entropy function between hypergraphs representing m modes,is a curved surface modal joint distribution entropy used for representing the conformal probability of each hyper-edge between hyper-graphs,is shown asA hypergraph of the characteristics of the individual modes,to representJoint probability distribution of individual modal feature matrices.

In some preferred embodiments, the method includes performing increase and decrease alignment convolution calculation on newly added hypergraph vertices and nodes covered by the event hypergraph network through the hyper-edge multi-modal dynamic evolution calculation, and further implementing automatic update of the event hypergraph network, and includes:

for the newly added input data stream, if newly added isolated nodes are obtained through A10, the complementary entropy of the isolated nodes and the nodes covered by the event hypergraph network obtained by the event hypergraph network automatic construction unit is calculated, and the isolated nodes are merged to the hypergraph corresponding to the nodes of which the complementary entropy is smaller than the set threshold;

if a new peak and a new super edge are obtained from the newly added data stream training sample, carrying out the convolution calculation of nodes of the super graph on the newly added super graph peak and the existing super graph peak according to the step A20 to complete the increase and decrease processing of the super graph peak; and C, carrying out hypergraph edge convolution calculation on the newly-added hypergraph edge and the existing hypergraph edge according to the step A30, and updating the matter hypergraph network.

In some preferred embodiments, the hyper-parametric adaptive selection policy functionComprises the following steps:

wherein the content of the first and second substances,is used for measuringThe measure has a candidate hyperparameterIs calculated byIn the hyper-parameter spaceAnd learning a target task datasetA denotes an algorithm in the hyper-parameter space.

In some preferred embodiments, the objective function y of the adaptive optimizing reasoning algorithm is:

wherein the content of the first and second substances,representing a hyper-parameter selection policy functionThe combined optimal curve function of (a) is,representing selection of a set of candidate hyperparameters according to a hyperparameter selection policyAnd the medium screen is selected from an adaptive optimization-searching training function, the adaptive optimization-searching training function is combined with data samples obtained in real time, optimal hyper-parameters are automatically set for each algorithm, the data samples are derived from learning target tasks, and c represents the number of hyper-parameter combinations of the ith historical data sample.

In some preferred embodiments, the objective function in the automatic migration unit in the automatic migration learning is:

wherein the content of the first and second substances,for newly adding a hyper-parameter variable of a learning target task domain,representing the target prediction function corresponding to the algorithm in the hyperparametric space,representing hyper-parametersThe combined optimal curve function of (a) is,the weight of the migration is represented as,representing selection strategies with optimal hyper-parametersThe set of hyper-parameters of (a),the number of sets of source algorithms is represented,a loop iteration counter is shown and is,represent iteration ofA secondary source algorithm; the source algorithm is an existing theoryAnd learning the hyper-parameter algorithm corresponding to the target task.

The invention has the beneficial effects that:

according to the invention, through a big data machine automatic learning method, fidelity evaluation reasoning is continuously iterated, a high fidelity mechanism model is automatically obtained, and the precision of behavior cognition prediction of a complex system is improved.

The invention simulates the cognitive learning behavior of nerves and symbols of the brain, combines a hypergraph representation model with automatic game interactive learning, automatically designs an industry mechanism model meeting specific task planning from massive fragmented multimodal dynamic data under limited time and computational complexity, performs predictive analysis on the logic relationship of the system mechanism model through continuous iteration model fidelity evaluation and reasoning, and improves the cognitive accuracy of complex system behaviors, thereby solving the problem that the existing machine learning method is difficult to obtain a high-fidelity mechanism model from dynamically evolving data, and establishing a big data machine learning system with real interpretability and robustness.

Drawings

Other features, objects and advantages of the present application will become more apparent upon reading of the following detailed description of non-limiting embodiments thereof, made with reference to the accompanying drawings.

FIG. 1 is an exemplary diagram of the overall architecture of a neural and symbolic based multimodal big data machine auto-learning system according to one embodiment of the present invention;

FIG. 2 is a diagram of an exemplary architecture of a feature engineering auto-construction module of the neural and symbolic based multi-modal big data machine auto-learning system according to an embodiment of the present invention;

FIG. 3 is a diagram of an example of a mechanism model automatic construction module of a neural and symbolic based multi-modal big data machine automatic learning system according to an embodiment of the present invention;

FIG. 4 is an exemplary diagram of an architecture of a hyper-parameter optimization module of the neural and symbolic based multi-modal big data machine auto-learning system according to an embodiment of the present invention.

Detailed Description

In order to make the objects, technical solutions and advantages of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings, and it is apparent that the described embodiments are some, but not all embodiments of the present invention. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.

The present application will be described in further detail with reference to the following drawings and examples. It is to be understood that the specific embodiments described herein are merely illustrative of the relevant invention and not restrictive of the invention. It should be noted that, for convenience of description, only the portions related to the related invention are shown in the drawings.

It should be noted that the embodiments and features of the embodiments in the present application may be combined with each other without conflict.

The invention relates to a multi-modal big data machine automatic learning system based on nerves and symbols, which comprises: the system comprises a characteristic engineering automatic construction module, a mechanism model automatic construction module, a hyper-parameter optimizing and optimizing module and a model data processing module;

the automatic characteristic engineering construction module comprises a data acquisition unit, an automatic case hypergraph network construction unit, an automatic network structure updating unit and a model data processing module;

the data acquisition unit is configured to acquire fragmented, multi-modal and dynamically-evolved big data as input data; the input data comprises text, image, audio and video data of different modes in a plurality of fields;

the automatic construction unit of the affair hypergraph network is configured to extract the concept symbols of affairs in different fields contained in the input data through a self-coding neural network, and construct a affair concept symbol space; based on the matter concept symbol space, automatically aggregating the super edges of different modes by a multi-mode feature automatic conformal representation method, and performing conformal calculation to generate a matter concept hypergraph network;

the network structure automatic updating unit is configured to extract concept symbols of affairs from new input data and obtain newly added hypergraph vertexes; performing increase and decrease alignment convolution calculation on newly added hypergraph vertexes and nodes covered by the affair hypergraph network through the hypercide multi-mode dynamic evolution calculation, and further realizing automatic updating of the affair hypergraph network;

the mechanism model automatic construction module comprises a field task definition unit, a model game design search unit and a search grid acceleration optimization unit;

the domain task definition unit is configured to divide concept symbols and hyper-parameter distribution conditions of the affairs of a set domain by combining the prior of nerve + symbol feature engineering aiming at an input data set of the domain, and pre-construct a model structure and a hyper-parameter search space; the hyper-parameters include: learning rate, the number of neural network layers, network structure, convolution kernel selection, iteration times, the number of hidden layer layers, neuron scale, sliding window and popular commonality index, activation function, the number of clusters and the number of topics;

the model game design searching unit is configured to distill a series of model structure candidate models from nodes of a rational hypergraph network through a game tree searching strategy in a pre-constructed model structure and a hyper-parameter searching space, wherein the model structure candidate models comprise a neural network prediction algorithm and a group of hyper-parameters, and the model game design searching unit is used for carrying out iterative high-fidelity evaluation on the series of candidate model structures and the hyper-parameter set and automatically searching out an optimal mechanism model most suitable for a current input data set; the mechanism model comprises a neural network model for description, prediction and early warning;

the search grid acceleration optimization unit is configured to obtain a learning curve corresponding to a model structure through a neural network structure prediction technology based on reinforcement learning, to fit a model trained from a training sample set and a model error minimum variance mean weight automatically generated from a test sample set according to model structure distribution predicted by the learning curve and hyper-parameter process weight sharing, and to share the weight globally in each model structure; the training sample set is a sample set constructed based on concept symbols of affairs in the affairs hypergraph network; the test sample set is an input data set of the set field;

the hyper-parameter optimizing module comprises a hyper-parameter space construction unit, a hyper-parameter self-adaptive selection strategy unit, a self-adaptive optimizing reasoning unit and a hyper-parameter automatic migration unit;

the hyper-parameter space construction unit is configured to divide the hyper-parameter data of the existing automatic machine learning algorithm into hyper-parameter populations of different automatic machine learning algorithms based on the hyper-parameter data of the existing automatic machine learning algorithm, and further construct a hyper-parameter space;

the hyper-parameter adaptive selection strategy unit is configured to select a multi-type candidate hyper-parameter set which meets the learning target task from the hyper-parameter space by taking concept symbols of affairs of each field contained in the updated affair hyper-graph network as the learning target task and combining the prior knowledge of each field and a predefined hyper-parameter adaptive selection strategy function;

the self-adaptive optimization searching and reasoning unit is configured to iteratively search a super-parameter combination of an optimal structure of a candidate algorithm and learning rate, regularization and network structure depth by adopting a parallel and sequence combined method according to a learning target task through a self-adaptive optimization searching and reasoning algorithm based on the candidate super-parameter set, search the super-parameter combination once to generate a super-parameter optimal curve, automatically compare the variation of the super-parameter optimal curve generated for many times, increase interference information until the variation exceeds a threshold value, terminate the self-adaptive optimization searching, and obtain an optimal super-parameter combination;

the automatic migration unit is configured to perform similar matching on the newly added learning target task and the existing learning target task, migrate the hyper-parameter combination corresponding to the existing learning target task with the type similarity higher than the preset threshold value to the hyper-parameter space of the newly added learning target task, and configure the optimal hyper-parameter for the newly added learning target task.

The model data processing module is configured to process the input data sets of all the set fields by combining the optimal mechanism model screened by the mechanism model automatic construction module and the optimal hyper-parameter combination obtained by the hyper-parameter optimization module; the processing comprises description, early warning and prediction.

In order to more clearly explain the multi-modal big data machine automatic learning system based on nerves and symbols, the following develops and details the modules of the system in various embodiments with reference to the attached drawings.

The invention discloses a multi-modal big data machine automatic learning system based on nerves and symbols, which is divided into a characteristic engineering automatic construction module 100, a mechanism model automatic construction module 200, a hyper-parameter optimization module 300 and a model data processing module 400 as shown in fig. 1;

1. automatic construction module for feature engineering

The automatic feature engineering construction module comprises a data acquisition unit 101, an automatic case-by-case hypergraph network construction unit 102 and an automatic network structure updating unit 103, which are shown in fig. 2; the method is mainly used for automatically inducing and expressing system affair symbol space from massive fragmented, multi-modal and dynamically-evolved large data streams, and logic association, time sequence and development relation among system affair concept symbols are described in real time through a hypergraph neural network, so that deep fusion of symbol learning reasoning advantages and neural network data perception advantages is realized, namely: neural + sign-event hypergraph networks. By continuously acquiring new observation sample data, mapping the observation sample to the event hypergraph network, further carrying out induction, deduction and statistical analysis, and finding out the rules of the sample in different areas of the event hypergraph network, the task goal to be realized by the algorithm is achieved, the event symbols and the associated structures are updated by self, and the event hypergraph network with the self-evolution nerve + symbol characteristics is automatically constructed. The method comprises the following specific steps:

the data acquisition unit 101 is configured to acquire fragmented, multi-modal and dynamically evolved big data as input data; the input data comprises text, image, audio and video data of different modes in a plurality of fields;

in this embodiment, massive fragmented, multimodal and dynamically evolved big data, including text, image, audio and video data of different modalities in multiple fields (or industries), i.e., multimodal data, is obtained as input data; the multi-modal data comprises numerical values, model structure parameters, hyper-parameters and the like.

The automatic construction unit 102 of the event hypergraph network is configured to extract concept symbols of events in different fields contained in the input data through a self-coding neural network, and construct a concept symbol space of events; based on the matter concept symbol space, automatically aggregating the super edges of different modes by a multi-mode feature automatic conformal representation method, and performing conformal calculation to generate a matter concept hypergraph network;

in this embodiment, entity symbols, i.e., entities, such as devices, materials, processes, and process rules, in each field are automatically sensed from the obtained multimodal data through a self-coding neural network, thousands of event concept symbols, i.e., event features, in different fields are extracted from the entities, and a physical concept symbol space, i.e., a tensor domain of the event features, is constructed.

Based on the extracted event features, by a preset multi-mode feature automatic conformal representation method, different modal super-edge structure groups are aggregated and conformal calculation is carried out, and a case hypergraph network is generated.

A10, adopting multi-mode feature selection calculation to solve the matter concept symbol space, and extracting a finite node set of the matter concept symbol in the input data;

the multi-modal feature selection calculates the corresponding objective function as:

wherein the content of the first and second substances,the symbol space of the concept of the matter is represented,a characteristic D dimension tensor representing the ith concept symbol, m represents the number of modes in the input data, N represents the total N sample sequences of the input,representing input numberBased on the corresponding a-priori vectors,a tensor coefficient matrix representing the ith mode, and storing all the multi-mode tensor coefficient vectors obtained currently in；

Performing objective function of multi-modal feature selection calculationNormalized generalization calculation of norm to obtainMulti-modal rank conversion, forming a finite set of nodes of m modal feature vectors, i.e., a finite set of nodes of conceptual symbols of the events in the input data。

A20, taking a finite node set of concept symbols of the affairs as a vertex set of a hypergraph, representing learning calculation and solving through the hypergraph to generate a hyperedge set, giving a weight and a time stamp to each hyperedge, and generating Laplace matrixes of the hyperedges of a plurality of modes to obtain a hyperedge structure group of the plurality of modes;

in a finite set of nodesIn the method, a k-means clustering algorithm is adopted, and one characteristic vertex is selected at willAs a central node, continuously calculating the central node and other nodesThe central node is connected with other adjacent m-1 vertexes to construct N stripsSuper edges, each super edge being given a weightAnd time stampAnd defining a hypergraph incidence matrix:

wherein the content of the first and second substances,representing nodesTo the central nodeThe distance of (a) to (b),indicates the currentThe Euclidean distance between the corresponding vertexes;

hypergraph-based vertex angle diagonal matrixGenerating Laplace half-angle matrix of hypergraphWherein, in the step (A),is a matrix of the vertex degrees of the hypergraph,degree matrix of superedges for hypergraphs, H being hypergraphsW is a system matrix of the hypergraph hyper-edge weights.

By adopting a Fourier transform method, the feature tensor of the Laplace half-angle matrix of the hypergraph is subjected to increase and decrease dynamic decomposition, and a convolution expression of the hypergraph can be obtainedWherein, in the step (A),representing the rank of the decomposition of the characteristics of the hypergraph Laplace half-angle matrix,representing hypergraph convolution kernel，Is a product of the Hadamard and the Hadamard,and (3) constructing a multi-mode dynamic hypergraph convolutional network, namely an initial affairs hypergraph network, for multi-mode data of m modes.

And A30, performing high-order correlation conformal entropy solving calculation among hypergraphs of different modes on the Laplacian matrix of the hypergraphs of the multiple modes to generate a case hypergraph network.

The conformal entropy calculation method comprises the following steps:

，

wherein the content of the first and second substances,a higher order correlation conformal entropy function between hypergraphs representing m modes,is a curved surface modal joint distribution entropy used for representing the conformal probability of each hyper-edge between hyper-graphs,is shown asA hypergraph of the characteristics of the individual modes,to representJoint probability distribution of individual modal feature matrices.

The method comprises the following steps of performing high-order related conformal entropy solving calculation on the Laplace matrix of the hypergraphs in the multiple modes to generate a case hypergraph network, wherein the high-order related conformal entropy solving calculation is performed on the hypergraphs in the different modes:

solving the first conformal entropy and the second conformal entropy; the first conformal entropy is hypergraph conformal entropy of m modes; the second conformal entropy is a conformal entropy when m = 2;

if the first conformal entropy is larger than the second conformal entropy, combining hypergraphs of different modalities into a theme; if the first conformal entropy is less than the second conformal entropy, the hypergraph stands alone as a topic;

and continuously carrying out iterative computation for many times, and updating the affair hypergraph network.

The network structure automatic updating unit 103 is configured to extract concept symbols of a matter from new input data and obtain a newly added hypergraph vertex; performing increase and decrease alignment convolution calculation on newly added hypergraph vertexes and nodes covered by the affair hypergraph network through the hypercide multi-mode dynamic evolution calculation, and further realizing automatic updating of the affair hypergraph network;

in this embodiment, a new supernode is obtained from newly added multi-modal data through a Self-induced game (Self-induced Play) algorithm, a concept symbol is convolutely aggregated to serve as a subject vertex of a supergraph network, a multi-element logic relationship among the concept symbol is output through a supergraph neural convolution method, a superedge of the supergraph network is generated, and logical relationships such as dependency, association and the like among group features are represented, and the specific process is as follows:

for the newly added input data stream, if newly added isolated nodes are obtained through A10, the complementary entropy of the isolated nodes and the nodes covered by the event hypergraph network obtained by the event hypergraph network automatic construction unit is calculated, and the isolated nodes are merged to the hypergraph corresponding to the nodes of which the complementary entropy is smaller than the set threshold;

if a new peak and a new super edge are obtained from the newly added data stream training sample, carrying out the convolution calculation of nodes of the super graph on the newly added super graph peak and the existing super graph peak according to the step A20 to complete the increase and decrease processing of the super graph peak; and C, carrying out hypergraph edge convolution calculation on the newly-added hypergraph edge and the existing hypergraph edge according to the step A30, and updating the matter hypergraph network.

Continuously adding source data as test samples, performing iterative graph convolution on the nodes and edges of the newly added hypergraph network without human experience and common sense intervention, mining the new features at the updated position of the dynamic hypergraph network, and automatically constructing a dynamic hypergraph convolution network with self-evolution capability, namely a neural and symbolic feature rational hypergraph network.

The hypergraph construction module supports the deep perception and automatic synthesis of a multi-modal heterogeneous data source, single-mode representation is projected to a multi-mode space, symbolic coding of various modal data is automatically completed, the problem of geological redundancy of the data source is solved, thousands of different-field concept symbol cooperation association modes are realized by calculating a common hypergeometric association modeling mode, interaction dynamics expansion relations among various modal matters are described in real time, and the coexistence of pivot nodes and interaction existing in different-scale groups is disclosed, so that the cooperation evolution is influenced. And a three-dimensional visual hypergraph inductive representation network is established, so that redundancy and correlation of mass data characteristics are eliminated, and the problem of simultaneous, different and different representation of mass training data characteristics is fundamentally solved.

2. Mechanism model automatic construction module

The mechanism model automatic construction module comprises a field task definition unit 201, a model game design search unit 202 and a search grid acceleration optimization unit 203, as shown in fig. 3; the method is mainly used for automatically carrying out zero and/or non-zero sum, complete information/incomplete information and other interactive game analysis and comparison on a super-large scale candidate model space and a specific data set by a multi-modal interactive learning algorithm, then selecting a model search strategy meeting specific system mechanism cognition, and carrying out fidelity evaluation on a trained model, thereby quickly and accurately searching out a high fidelity mechanism model suitable for the data set. The method comprises the following specific steps:

the domain task defining unit 201 is configured to divide concept symbols and hyper-parameter distribution conditions of the affairs of a set domain by combining the prior of nerve + symbol feature engineering according to an input data set of the domain, and pre-construct a model structure and a hyper-parameter search space; the hyper-parameters include: learning rate, the number of neural network layers, network structure, convolution kernel selection, iteration times, the number of hidden layer layers, neuron scale, sliding window and popular commonality index, activation function, the number of clusters and the number of topics;

in this embodiment, for input data sets in different fields, that is, multi-modal data, a priori of the neural + symbolic feature engineering, that is, a case hypergraph network constructed by an automatic feature engineering construction module is combined, model structure parameter nodes (that is, concept symbols of a case) and a hyper-parameter distribution condition corresponding to the input data sets in the set field in the case hypergraph network are partitioned, and a model structure and a hyper-parameter search space are further constructed.

The model game design search unit 202 is configured to distill a series of model structure candidate models including a neural network prediction algorithm and a group of hyper-parameters from nodes of a rational hyper-graph network through a game tree search strategy in a pre-constructed model structure and hyper-parameter search space, and perform iterative high-fidelity evaluation on the series of candidate model structures and hyper-parameter sets to automatically search out an optimal mechanism model most suitable for a current input data set; the mechanism model comprises a neural network model for description, prediction and early warning;

in the embodiment, in a pre-constructed model structure and hyper-parameter search space, a series of model structure candidate models are distilled from nodes of a rational hypergraph network by supporting model hyper-parameter automatic search strategies such as zero and/non-zero sum, complete information/incomplete information and the like, namely a game tree search algorithm, wherein the model structure candidate models comprise a neural network prediction algorithm and a group of hyper-parameters.

After distillation, various candidate model subnet structures are continuously cut, combined and decomposed, a model prediction search algorithm is selected, a high-fidelity system mechanism model is predicted from data like human brain, evaluation operators and cross operators are continuously corrected by combining source continuous data flow, the prediction result is iteratively evaluated according to the fitness, iterative evolution is carried out, and the model with the optimal structure and parameters can be obtained until the model fitness meets the existing test data set. Performing iterative high-fidelity evaluation on a series of candidate model structures and a hyper-parameter set, and automatically searching out an optimal mechanism model most suitable for a current input data set; the mechanism model comprises a neural network model for description, prediction and early warning. Namely, the mechanism model with the highest fidelity evaluation value is used as the mechanism model which is most suitable for the current input data set.

The search grid acceleration optimizing unit 203 is configured to obtain a learning curve corresponding to a model structure through a neural network structure prediction technology based on reinforcement learning; according to the model structure distribution and the hyper-parameter process weight predicted by the learning curve, fitting a model trained from a training sample set and a model error minimum variance mean weight automatically generated from a test sample set, and globally sharing the weights in each model structure; the training sample set is a sample set constructed based on concept symbols of affairs in the affairs hypergraph network; the test sample set is an input data set of the set field.

The search grid acceleration optimization is responsible for optimizing a network structure designer, and the search space scale and the search time complexity are reduced in the modes of continuously iterative search for predicting the model structure and parameter process weight sharing from a huge search space, sharing the network structure and the like according to global and local cooperative strategies.

In the embodiment, a learning curve corresponding to a model structure is obtained through a neural network structure prediction technology based on reinforcement learning; and according to the model structure distribution and the hyper-parameter process weight predicted by the learning curve, fitting a model trained from a training sample set and a model error minimum variance mean weight automatically generated from a test sample set, and globally sharing the weight in each model structure.

And when the similarity of the newly added input data set and the historical input data set is higher than a set threshold, sharing the model structure distribution and the hyper-parameter process weight corresponding to the historical input data set to the newly added input data set.

On one hand, a series of prediction neuron network structure candidate models are distilled from a hypergraph network structure of the hypergraph network according to a specific task training sample from a psycho-hypergraph network of a nerve + symbol characteristic engineering, and the hypergraph network structure candidate models comprise a neural network prediction algorithm and a group of hyper-parameters. And on the other hand, iterative high-fidelity evaluation is carried out on a series of candidate model structures and parameter sets to obtain the minimum variance mean value of errors of the training samples and the testing samples, so that a neural network model which is most suitable for behavior cognition such as optimal description, diagnosis, prediction, early warning and the like of the current data set is actively searched.

By adopting an automatic 'interaction-trial-and-error' game interactive learning mechanism, various automatic operations such as cutting, decomposition, recombination and the like of the interior and the super edge of a super node of a network are summarized and expressed for a three-dimensional visual hypergraph, a multi-type task cognitive model is generated in a limited time, the high-fidelity model structure and parameters meeting the current data set are automatically and iteratively evaluated and searched out without depending on field expert experience, the model search space is greatly shortened, the problem of high calculation complexity of the automatic modeling of the current big data environment based on an automatic machine learning NAS mechanism is solved, and the success rate of model structure search is accelerated.

In addition, the data-driven automatic mechanism learning method for the complex system introduces a disaster entropy mutation cause and effect inference technology, realizes real-time active monitoring of the nonlinear, emergent, balance step, adaptability and special property instability, periodic oscillation and mutation cause and effect elements of a feedback loop of the complex system, and realizes active pre-warning of major faults in the operation process of the system by means of a case library, an expert library and a pre-plan library. The method comprises the following specific steps:

the differential residual learning module is configured to complete calculation of profits generated by population adjustment in the operation process through a copied differential equation simulating biological evolution and a copied differential equation simulating biological evolution based on the searched optimal mechanism model, and automatically capture causal factors of system instability and periodic oscillation in the system operation process through a differential residual learning function in the population fitness; obtaining a continuous instability Hamilton equation of the system through a differential residual error learning function in the fitness of the whole system dynamics control rule;

in this embodiment, the specific process is as follows:

(1) based on the optimal mechanism modelTo do so byRepresenting a set of N different individual behaviors, i.e. nodes, toRepresenting the causal game relation of the behaviors between every two groups, namely the overedge;

(2) based on the optimal mechanism modelCalculating the first game income generated by population adjustment in the running process by simulating the replication differential equation of biological evolutionAutomatically capturing the first game income through a differential residual learning function in the population fitness based on the first game incomeAcquiring a Hamilton equation of continuous system instability by causal factors of system instability and periodic oscillation in the system operation process;representing behaviorAndcausal relative utility between;

the replication differential equation for simulating the biological evolution is as follows:

，

wherein the content of the first and second substances,the fitness of the individual behavior is represented,which represents the degree of average of the images,which is indicative of the behavior of the individual,representing the behavior of the individual after evolutionary replication;

if the individual behavior fitness degreeGreater than average degreeRepresenting individual behaviorThe population number rapidly increases; if the individual behavior fitness degreeLess than average degreeRepresenting individual behaviorThe population number is rapidly reduced; if the individual behavior fitness degreeEqual to the mean degreeRepresenting individual behaviorThe number of the groups is not changed; obtaining a continuous instability Hamilton equation of the system through a differential residual error learning function in the fitness of the whole system dynamics control rule;

a mutation causal model construction module which is configured to complete calculation of gains generated by population adjustment in the operation process by simulating a biological evolution mutation differential equation based on causal factors inducing system instability, periodic oscillation and mutation in the system operation processCompleting training of dynamic microstructures of different-layer neuron parametric hidden units of the dynamic system by differential residual learning functions in the fitness of the dynamic control rule of the whole system to obtain a Hamilton equation of continuous mutation of the system; and constructing mutation causal models of all parameters and attributes in the system, such as completely homogeneous symmetrical behavior networks, completely heterogeneous symmetrical behavior networks, asymmetrical behavior networks and the like.

In the present embodiment, (1) is based on the aboveThe causal factors of system instability and periodic oscillation calculate the second game income generated by population adjustment by simulating the differential equation of biological evolution mutationCompleting dynamic microstructure training of different-layer neuron parameterization hidden units of the dynamic system through a differential residual learning function in the population fitness based on the second game income; the differential equation of the biological evolution mutation is as follows:

，

(2) based on a trained power system, a continuous mutation Hamiltonian equation of the power system is obtained, and further mutation causal models of all parameters and attributes in the system, such as a completely homogeneous symmetric behavior network, a completely heterogeneous symmetric behavior network and an asymmetric behavior network, are obtained.

For the construction of the mutation causal model building module, see the literature: "Tianli, Dong xi Wang, Zhao Xiong, Li Qingdong, Lu jin Hu, ren Chapter. heterogeneous cluster system distributed adaptive output time-varying formation tracking control. automation newspapers, 2020". and "Zheng Zhiming, Lu jin Hu, Weiwei, sugar Ting" precise intelligent theory: artificial intelligence facing complex dynamic objects "China science, 2021".

And the prediction acquisition module is configured to perform Hamiltonian equation solution calculation on the Glandum causal weight of each node-neighbor pair of the sequence long-range association hypergraph model based on the completely homogeneous symmetrical behavior network, the completely heterogeneous symmetrical behavior network, the asymmetrical behavior network and other mutation causal models, construct a Hamiltonian function, perform intervention and counter-fact calculation of causal random gradients, maximize the prediction causal weight error between the historical time sequence and the future time sequence, and obtain the complex system behavior prediction result.

In the embodiment, the risk active early warning constructed by the data-driven automatic mechanism learning method for the complex system can automatically analyze the possible fault risk and fault position of the complex system after data input, and improve the reliability of the complex system. And the accuracy of the prediction is improved.

3. Super-parameter optimization module

The hyper-parameter optimizing module comprises a hyper-parameter initial space construction unit 301, a hyper-parameter adaptive selection strategy unit 302, an adaptive optimizing reasoning unit 303 and a hyper-parameter automatic migration unit 304, as shown in fig. 4; the method is mainly used for automatically completing the configuration of the hyper-parameters under the condition of limited time constraint, performing various automatic search, combination, fitting, evaluation, experience migration and other optimization reasoning on the hyper-parameters, reducing predefined loss functions, accurately obtaining the mechanism evolution state of the multi-target entity, generating various high-value semantic symbols and improving the prediction or classification accuracy of given independent data.

The hyper-parameter initial space construction unit 301 is configured to divide the hyper-parameter data of the existing automatic machine learning algorithm into hyper-parameter populations of different automatic machine learning algorithms based on the hyper-parameter data of the existing automatic machine learning algorithm, and further construct a hyper-parameter initial space;

in this embodiment, the hyper-parameter space is used as a hyper-parameter configuration space for multiple types of machine learning model algorithms (i.e. the neural network model described above), and is mainly used for defining the parameters to be configuredThe instantiated various cognitive model algorithms, real variables, integer variables, binary variables, learning PipeLine Pipeline and the like of each machine learning algorithm are convenient for an optimizer to traverse all hyper-parameter problem definitions and descriptions and correlation configuration.

Classifying the hyper-parameters, generating hyper-parameter populations of various automatic machine learning algorithms, calculating hyper-parameter Euclidean similarity in the hyper-parameter populations, setting a sharing edge between the hyper-parameters with the Euclidean similarity being larger than or equal to a preset sharing edge threshold, generating hyper-parameter distribution subgraphs related to different population algorithm nodes by the hyper-parameters with the Euclidean similarity being smaller than the preset sharing edge threshold, and constructing a hyper-parameter space.

The hyper-parameter adaptive selection policy unit 302 is configured to select a multi-type candidate hyper-parameter set satisfying the learning target task from the hyper-parameter space by taking concept symbols of affairs of each field included in the updated affair hyper-graph network as the learning target task and combining the prior knowledge of each field and a predefined hyper-parameter adaptive selection policy function;

in this embodiment, the hyper-parametric adaptive selection policy functionComprises the following steps:

wherein the content of the first and second substances,for measuring having candidate hyper-parametersIs calculated byIn the hyper-parameter spaceAnd learning a target task datasetA denotes an algorithm in the hyper-parameter space.

Alternative strategies include the following: learning type super-parameter optimization based on reinforcement learningThe method automatically trains and evaluates the characteristics of the method through single or combined hyperparameter search strategies, such as search type hyperparameter optimization based on evolutionary algorithm, probability type hyperparameter optimization based on Bayesian optimization, and the likeAlgorithm of parametersIn experimental validation of the data setAnd existing super parameter spaceTo be lost.

The adaptive optimal reasoning unit 303 is configured to base the candidate hyper-parameter set onIteratively exploring the hyperparametric combination of the optimal structure of the candidate algorithm and the learning rate, regularization and network structure depth by adopting a parallel and sequence combined method through a self-adaptive optimization-searching reasoning algorithm according to a learning target task, exploring the hyperparametric combination once every time to generate a hyperparametric optimal curve, automatically comparing the variation of the hyperparametric optimal curve generated for multiple times, increasing interference information until the variation exceeds a threshold value, terminating the self-adaptive optimization, and obtaining the optimal hyperparametric combination;

in this embodiment, the objective function y of the adaptive optimizing reasoning algorithm is:

wherein the content of the first and second substances,representing a hyper-parameter selection policy functionThe combined optimal curve function of (a) is,representing selection of a set of candidate hyperparameters according to a hyperparameter selection policyAnd the medium screen is selected from an adaptive optimization-searching training function, the adaptive optimization-searching training function is combined with data samples obtained in real time, optimal hyper-parameters are automatically set for each algorithm, the data samples are derived from learning target tasks, and c represents the number of hyper-parameter combinations of the ith historical data sample.

The automatic migration unit 304 is configured to perform similar matching between the newly added learning target task and the existing learning target task, and migrate the hyper-parameter combination corresponding to the existing learning target task with the type similarity higher than the preset threshold value to the hyper-parameter space of the newly added learning target task, so as to configure the optimal hyper-parameter for the newly added learning target task.

In this embodiment, the objective function in the automatic transfer learning is:

wherein the content of the first and second substances,for newly adding a hyper-parameter variable of a learning target task domain,representing the target prediction function corresponding to the algorithm in the hyperparametric space,representing hyper-parametersThe combined optimal curve function of (a) is,the weight of the migration is represented as,representing selection strategies with optimal hyper-parametersThe set of hyper-parameters of (a),the number of sets of source algorithms is represented,a loop iteration counter is shown and is,represent iteration ofA secondary source algorithm; the source algorithm is an existing hyper-parameter algorithm corresponding to the learning target task.

According to the method, the automatically set hyper-parameters are carried out under the condition of limited time constraint, and the hyper-parameters are subjected to optimization such as various searching, combination, fitting, evaluation, experience migration and the like, so that predefined loss functions are reduced, the performance of the cognitive neuron network algorithm in dynamic data flow is improved, and the prediction or classification precision of given independent data is improved.

In addition, according to the automatic assembly line hyper-parameter closed-loop optimization mechanism such as problem perception, combination optimization, self evaluation and the like, the invention integrates the methods such as learning type hyper-parameter optimization, probability type hyper-parameter optimization, search type hyper-parameter optimization and the like, and solves the problems of infinite hyper-parameter traversal space, difficulty in convergence, small information gain and the like of the current hyper-parameter optimization such as nonlinearity, non-convexity, combination optimization, hybrid optimization and the like.

4. Model data processing module

The model data processing module is configured to process the input data sets of all the set fields by combining the optimal mechanism model screened by the mechanism model automatic construction module and the optimal hyper-parameter combination obtained by the hyper-parameter optimization module; the processing comprises description, early warning and prediction.

It should be noted that, the neural and symbolic based multi-modal big data machine automatic learning system provided in the above embodiment is only exemplified by the division of the above functional modules, and in practical applications, the above functions may be allocated to different functional modules according to needs, that is, the modules or steps in the embodiment of the present invention are further decomposed or combined, for example, the modules in the above embodiment may be combined into one module, or may be further split into multiple sub-modules, so as to complete all or part of the above described functions. The names of the modules and steps involved in the embodiments of the present invention are only for distinguishing the modules or steps, and are not to be construed as unduly limiting the present invention.

The terms "comprises," "comprising," or any other similar term are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus.

So far, the technical solutions of the present invention have been described in connection with the preferred embodiments shown in the drawings, but it is easily understood by those skilled in the art that the scope of the present invention is obviously not limited to these specific embodiments. Equivalent changes or substitutions of related technical features can be made by those skilled in the art without departing from the principle of the invention, and the technical scheme after the changes or substitutions can fall into the protection scope of the invention.