CN116542136A - Universal method and device for searching and multiplexing learning objects - Google Patents

Universal method and device for searching and multiplexing learning objects Download PDF

Info

Publication number
CN116542136A
CN116542136A CN202310396671.3A CN202310396671A CN116542136A CN 116542136 A CN116542136 A CN 116542136A CN 202310396671 A CN202310396671 A CN 202310396671A CN 116542136 A CN116542136 A CN 116542136A
Authority
CN
China
Prior art keywords
learning
model
user
statistical
semantic
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202310396671.3A
Other languages
Chinese (zh)
Inventor
周志华
谭志豪
詹德川
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Nanjing University
Original Assignee
Nanjing University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Nanjing University filed Critical Nanjing University
Priority to CN202310396671.3A priority Critical patent/CN116542136A/en
Publication of CN116542136A publication Critical patent/CN116542136A/en
Pending legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/20—Design optimisation, verification or simulation
    • G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/20—Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
    • G06F16/24—Querying
    • G06F16/245—Query processing
    • G06F16/2458—Special types of queries, e.g. statistical queries, fuzzy queries or distributed queries
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/22—Matching criteria, e.g. proximity measures
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
    • Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • General Physics & Mathematics (AREA)
  • Software Systems (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Artificial Intelligence (AREA)
  • Probability & Statistics with Applications (AREA)
  • Medical Informatics (AREA)
  • Fuzzy Systems (AREA)
  • Mathematical Physics (AREA)
  • Geometry (AREA)
  • Computational Linguistics (AREA)
  • Computer Hardware Design (AREA)
  • Databases & Information Systems (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Evolutionary Biology (AREA)
  • Machine Translation (AREA)

Abstract

The invention aims to provide a method and a device for searching and multiplexing general learning objects. The method comprises the following steps: based on query information corresponding to the task demands of the current user, a plurality of candidate learning pieces are screened from the learning piece library by matching based on semantic conventions; selecting one or more learning objects from the plurality of candidate learning objects by matching based on a statistical specification; and returning the selected one or more learning objects to the user for multiplexing the corresponding models by the user. The embodiment of the application has the following advantages: the method has the advantages that the models of the developers are collected, semantic conventions and statistical conventions are generated for the models, the semantic conventions and the statistical conventions are combined into learning pieces to be stored in a learning piece library, and according to the machine learning task requirements of different users, the learning pieces which are helpful to the users are searched and matched based on the semantic conventions and the statistical conventions to be further multiplexed to the user tasks, so that the users can obtain a good-quality model which is helpful to the users to be multiplexed without training a new model from the beginning, and the efficiency is improved.

Description

Universal method and device for searching and multiplexing learning objects
Technical Field
The invention relates to the technical field of computers, in particular to a method and a device for searching and multiplexing general learning objects.
Background
Machine learning methods have found successful application in a number of fields including natural language processing, image recognition, video speech recognition, and the like. However, current machine learning still faces unavoidable problems: the desire to build an excellent machine learning model from scratch is still costly, often requires a large amount of conceptually tagged historical data (referred to as tagged data), computational resources, and machine learning expertise; the existing high-quality model is difficult to cope with environmental changes, and the model with excellent performance in the old environment can lose the original performance in the face of the new environment, so that the model has to be abandoned; the model easily forgets the learned knowledge in the history environment in the process of adapting to the new environment; because of the problems of data privacy and ownership, high-quality data cannot be shared and disclosed, and the acquisition of the data mark on the new task requires a great deal of manpower and material resources. How to use a large number of high-quality models obtained by training on the existing tasks to assist in solving the new tasks, saving the cost of data collection and model training, has become an important subject in machine learning research.
In addition, conventional methods such as transfer learning and field adaptation have been widely used in many scenarios; however, these methods still have many limitations in practice. First, the number of models they consider is very limited in scale, and it is often assumed that source domain tasks for knowledge migration are potentially useful for the user's target tasks, i.e., the user can only be aided by a small number of source domain models closely associated therewith; second, they generally assume that the original data of the source domain task can be accessed, which assumption severely affects the application of both types of methods in the practical task of emphasizing data privacy protection.
Disclosure of Invention
The invention aims to provide a method and a device for searching and multiplexing general learning objects.
According to an embodiment of the application, a method for searching and multiplexing models based on a learning model is provided, wherein the method comprises the following steps:
based on query information corresponding to the task demands of the current user, a plurality of candidate learning pieces are screened from the learning piece library by matching based on semantic conventions;
selecting one or more learning objects from the plurality of candidate learning objects by matching based on a statistical specification;
and returning the selected one or more learning objects to the user for multiplexing the corresponding models by the user.
According to an embodiment of the present application, there is provided a device for performing model searching and multiplexing based on a learning model, where the device includes:
means for screening a plurality of candidate learning objects from the learning object library by matching based on semantic conventions based on query information corresponding to a current user task demand;
means for selecting one or more learning objects from the plurality of candidate learning objects by matching based on a statistical specification;
and means for returning the selected one or more learning objects to the user for multiplexing by the user the corresponding models.
According to an embodiment of the present application, there is provided a computer device including a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor implements the method of the embodiments of the present application when executing the program.
According to an embodiment of the present application, there is provided a computer-readable storage medium having stored thereon a computer program, characterized in that the program, when executed by a processor, implements the method of the embodiments of the present application.
Compared with the prior art, the embodiment of the application has the following advantages: the method has the advantages that the models of the developers are collected, semantic conventions and statistical conventions are generated for the models, the semantic conventions and the statistical conventions are combined into learning pieces to be stored in a learning piece library, and aiming at the machine learning task requirements of different users, the learning pieces which are helpful to the users are searched and matched based on the semantic conventions and the statistical conventions to be further multiplexed in the user tasks, so that the users can obtain a good-quality model which is helpful to the users without training a new model from the beginning to be multiplexed, and the efficiency is improved; according to the embodiment of the application, the model is further searched through the statistical protocol on the basis of the semantic protocol, so that the model helpful to the user can be accurately identified; the embodiment of the application does not contact the original data of the user and all model developers, so that the data privacy is strictly protected.
Drawings
Other features, objects and advantages of the present invention will become more apparent upon reading of the detailed description of non-limiting embodiments, made with reference to the accompanying drawings in which:
FIG. 1 shows a flow chart of a method for model searching and multiplexing based on a learning model form according to an embodiment of the application;
FIG. 2 illustrates a schematic diagram of an exemplary learning-piece management system, according to an embodiment of the present application;
fig. 3 shows a schematic structural diagram of a device for performing model searching and multiplexing based on a learning model according to an embodiment of the application.
The same or similar reference numbers in the drawings refer to the same or similar parts.
Detailed Description
Before discussing exemplary embodiments in more detail, it should be mentioned that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although a flowchart depicts operations as a sequential process, many of the operations can be performed in parallel, concurrently, or at the same time. Furthermore, the order of the operations may be rearranged. The process may be terminated when its operations are completed, but may have additional steps not included in the figures. The processes may correspond to methods, functions, procedures, subroutines, and the like.
In this context, the term "computer device", also called a "computer", refers to an intelligent electronic device that can execute a predetermined process such as numerical computation and/or logic computation by executing a predetermined program or instruction, and may include a processor and a memory, the predetermined process being executed by the processor executing a stored instruction stored in the memory, or the predetermined process being executed by hardware such as ASIC, FPGA, DSP, or a combination of both. Computer devices include, but are not limited to, servers, personal computers, notebook computers, tablet computers, smart phones, and the like.
The computer device includes a user device and a network device. Wherein the user equipment includes, but is not limited to, a computer, a smart phone, a PDA, etc.; the network device includes, but is not limited to, a single network server, a server group of multiple network servers, or a Cloud based Cloud Computing (Cloud Computing) consisting of a large number of computers or network servers, where Cloud Computing is one of distributed Computing, and is a super virtual computer consisting of a group of loosely coupled computer sets. The computer device can be independently operated to realize the application, and can also be accessed to a network and realize the application through interaction with other computer devices in the network. Wherein the network where the computer device is located includes, but is not limited to, the internet, a wide area network, a metropolitan area network, a local area network, a VPN network, and the like.
It should be noted that the user device, the network, etc. are only examples, and other computer devices or networks that may be present in the present application or in the future are applicable to the present application, and are also included in the scope of the present application and are incorporated herein by reference.
The methods discussed below (some of which are illustrated by flowcharts) may be implemented by hardware, software, firmware, middleware, microcode, hardware description languages, or any combination thereof. When implemented in software, firmware, middleware or microcode, the program code or code segments to perform the necessary tasks may be stored in a machine or computer readable medium such as a storage medium. The processor(s) may perform the necessary tasks.
Specific structural and functional details disclosed herein are merely representative and are for purposes of describing example embodiments of the present application. This application may be embodied in many alternate forms and should not be construed as limited to only the embodiments set forth herein.
It will be understood that, although the terms "first," "second," etc. may be used herein to describe various elements, these elements should not be limited by these terms. These terms are only used to distinguish one element from another element. For example, a first element could be termed a second element, and, similarly, a second element could be termed a first element, without departing from the scope of example embodiments. The term "and/or" as used herein includes any and all combinations of one or more of the associated listed items.
It will be understood that when an element is referred to as being "connected" or "coupled" to another element, it can be directly connected or coupled to the other element or intervening elements may be present. In contrast, when an element is referred to as being "directly connected" or "directly coupled" to another element, there are no intervening elements present. Other words used to describe relationships between units (e.g., "between" versus "directly between," "adjacent to" versus "directly adjacent to," etc.) should be interpreted in a similar manner.
The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of example embodiments. As used herein, the singular forms "a", "an" and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will be further understood that the terms "comprises" and/or "comprising," when used herein, specify the presence of stated features, integers, steps, operations, elements, and/or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and/or groups thereof.
It should also be noted that, in some alternative implementations, the functions/acts noted may occur out of the order noted in the figures. For example, two figures shown in succession may in fact be executed substantially concurrently or the figures may sometimes be executed in the reverse order, depending upon the functionality/acts involved.
The invention is described in further detail below with reference to the accompanying drawings.
Fig. 1 shows a flowchart of a method for model searching and multiplexing based on a learning model according to an embodiment of the application. The method comprises the steps of S1, S2 and S3.
The present embodiments employ a learning model to store and manage models from different developers. Wherein the study includes a model and a specification describing the model. A study library containing numerous studies may be referred to as a study market.
The method of the embodiment of the application is applicable to various models, such as a linear model, a forest model, a support vector machine, a neural network model and the like.
The method of the embodiment of the application is executed by equipment where the learning-oriented library is located.
Prior to step S1, the method receives models from various developers and builds corresponding school pieces through steps S4 and S6.
In step S4, based on the model uploaded by the developer, a semantic specification corresponding to the model is generated.
Wherein the semantic conventions include one or more tag information that is used to describe the model. The semantic conventions may also include textual descriptive information, such as, for example, textual descriptive information of model processing tasks entered by the model uploader.
According to one embodiment, the semantic conventions include semantic tag sets based on a preset multi-level tag library. The multi-level tag library is a multi-level semantic tag describing the attribute of the learning object, such as task type, data information, model type and size, and the like. The semantic tags can be custom designed according to a model library.
For example, in the sales prediction application scenario, the multi-level tag library contains three levels of tags related to the traffic type: wholesale retail industry-retail-department, wholesale retail industry-wholesale-fresh, manufacturing industry-machinery-engine; secondary labels on task type: supervised learning-regression, supervised learning-classification, unsupervised learning-clustering; primary tags on data type: forms, text, images; primary labels on model type: linear model, depth model, integrated model. One specific semantic specification is a set of semantic tags in a tag library, for example, the user's semantic specification may be [' wholesale retail-department ',' supervised learning-regression ',' form data ','. The user does not need to fill all the labels, and the more detailed the labels are filled, smaller protocol islands can be selected, so that the matching of statistical protocols can be performed.
According to one embodiment, quality detection is performed on the model uploaded by the developer, and whether the model is received or not is judged based on the detection result and is stored in a learning-class library. If the quality detection result meets the requirement, the model is received and stored in a learning-class library; if the quality detection result does not meet the requirement, corresponding prompt is carried out to the user.
The person skilled in the art should be familiar with the method, for example, the method can calculate the index reflecting the performance of the model, and detect whether the index value is better than the predetermined threshold value, etc., and the person skilled in the art can select a suitable method to perform the quality detection based on the actual requirement, which is not described herein.
Next, in step S5, a statistical specification corresponding to the model is obtained according to the semantic specification corresponding to the model.
Wherein the statistical conventions include information describing statistical distributions of the data sets. Preferably, the forms of statistical conventions include, but are not limited to, abbreviated kernel mean embedding, statistical thumbnail sets, cluster center points, distribution thumbnail sets.
Specifically, a statistical protocol construction interface and parameters corresponding to the semantic protocol corresponding to the model are sent to equipment where a developer is located, so that the developer can construct a statistical protocol for describing statistical information of a data set according to the statistical protocol construction interface and parameters and an original data set used by the model, and obtain the statistical protocol uploaded by the developer.
According to one embodiment, the corresponding protocol islands are determined based on the semantic protocols corresponding to the models, and then the statistical protocol construction interfaces and parameters corresponding to the determined protocol islands are sent to the device where the developer is located. Wherein the parameters include, but are not limited to, kernel functions. And then, the equipment where the developer is located constructs an interface and parameters according to the statistical protocol, constructs a statistical protocol for describing statistical information of the data set according to the original data set used by the model, and uploads the statistical protocol to a learning library.
In step S6, based on the model and the semantic and statistical conventions corresponding to the model, a corresponding learning object of the model is obtained.
According to one embodiment, the method comprises step S7 and step S8.
In step S7, a plurality of protocol islands for storing the study are generated.
Wherein the models in the same protocol island address the same or similar and their learning task requirements.
In step S8, for the models received from the respective developers, based on the semantic conventions corresponding to the models, they are placed in the corresponding protocol islands for storage, so that the models of the chemicals contained in each protocol island have the same or similar semantic conventions.
And after receiving a new model, obtaining a learning piece corresponding to the model based on the semantic protocol and the statistical protocol corresponding to the model, and correspondingly updating the protocol island corresponding to the model to finish the addition of the new learning piece.
In step S1, a plurality of candidate learning objects are screened from the learning object library by matching based on semantic conventions based on the query information corresponding to the current user task requirements, as described below with reference to fig. 1.
The query information comprises text information for querying a model in a learning-oriented library.
The method for acquiring the query information includes, but is not limited to:
1) Generating one or more keywords corresponding to the task demands of the current user based on preset label information, and taking the keywords as corresponding query information; for example, based on a multi-level tag library preset by a learning component library, acquiring multi-level tags corresponding to tasks selected or input by a user through a visual interface, and taking corresponding keyword sets as query information;
2) Taking other text information input by a user as query information; for example, if the user does not provide the completed multi-level label, the text description information for the task uploaded by the user is used as query information.
According to one embodiment, the method stores models from various developers in a plurality of protocol islands, each protocol island containing models of chemicals having the same or similar semantic protocols, and step S1 includes step S101.
In step S101, based on the query information, by calculating semantic similarity, a protocol island with the highest semantic similarity to the query information is determined, and a learning object in the protocol island is used as a candidate learning object.
The semantic similarity calculation includes various text similarity calculation modes, for example, cosine similarity calculation of two labels or optimal transportation distance calculation.
For example, in an application scene of sales volume prediction, based on a multi-level tag library preset by a learning library, a multi-level tag corresponding to a task of a user is obtained through a visual interface, and a keyword set 'regression task and sales volume prediction' corresponding to the selected tag is used as query information. And then, based on the query information, positioning to a learning piece protocol island corresponding to the sales volume prediction task through calculating semantic similarity calculation, and taking the learning pieces in the protocol island as candidate learning pieces, so that a large number of learning models for solving the sales volume prediction task are screened out from a learning piece library.
Continuing with the description of FIG. 1, in step S2, one or more learning objects are selected from the plurality of candidate learning objects by matching based on statistical conventions.
Wherein the method provides the selected one or more school components to the user as the school component most likely to be helpful to the user's task.
Wherein, the step S2 includes a step S201 and a step S202.
In step S201, a plurality of candidate learning objects are sent to a user to obtain corresponding test feedback information.
Wherein the feedback information includes, but is not limited to, model accuracy, precision, recall, etc.
In step S202, matching is performed among a plurality of candidate learning objects according to the test feedback information and the statistical rules of the user task, so as to obtain one or more matched learning objects.
Wherein, matching of statistical conventions refers to metric computation between statistical conventions in combination with measurement indicators on representative learning pieces, including but not limited to distance computation in regenerated kernel hilbert space, and the like.
According to one embodiment, the method further screens the candidate learning objects to obtain a representative learning object, sends the representative learning object to a user to obtain corresponding test feedback information, and matches the test feedback information with a statistical rule of a user task in the representative learning objects to obtain one or more matched learning objects.
The method according to the embodiment screens the representative learning item includes, but is not limited to, calculating a distance matrix between statistical conventions of each learning item in the conventions island, using a clustering algorithm based on the distance matrix, and taking the learning item corresponding to the cluster center as the representative learning item.
According to one embodiment, the method constructs the statistical specification of the user task through the preset statistical specification construction interface and parameters and the original data set collected by the user task under the condition of obtaining the user consent.
If the user does not wish to go to the traditional metering protocol, a learning match is made based on the semantic protocol and the user's test feedback information.
In step S3, the selected one or more learning objects are returned to the user for multiplexing of the corresponding models by the user.
According to one embodiment, the method guides the user to perform learning-part multiplexing including integrating predicted results of individual learning-parts on user tasks, fine-tuning the learning-part model using user task data, expanding the user training data with the predicted results of the learning-part model as additional features, and the like.
For example, in the scenario of sales prediction, the method completes matching of semantic conventions based on computation of semantic similarity of labels. Assuming the user fills out multi-level tags [ 'wholesale retail-department', 'supervised learning-regression', 'text data', 'integrated model','s complete identity, a number of models remain, assuming the method searches the library of learning components for a number of learning components that may be helpful to the user's task. And then, according to the statistical specifications uploaded by the user, searching the statistical specifications based on the specification matching function, for example, when the abbreviated kernel mean value is adopted for embedding, selecting one or more of the closest statistical specifications with the user, and returning the selected statistical specifications to the user for the user to solve own learning task.
The method of the embodiments of the present application is described below in connection with an example.
FIG. 2 illustrates a schematic diagram of an exemplary learning-related management system, according to an embodiment of the present application.
Referring to FIG. 2, the parts management system employs parts to store and manage models from different developers. The flow of this example mainly includes an upload phase and a deployment phase. In the upload phase, tens of thousands of models (denoted as model Y) from different developers (denoted as developer 1 through developer N) 1 To Y N ) Is submitted to a study store, known as the study market. In the deployment stage, when a user faces a new machine learning task, the learning management system performs matching through the semantic rule matching module and the statistical rule matching module according to the actual requirement of the user corresponding to the new task, returns the learning most likely to be helpful to the user and guides the user to apply the learning on the new task.
Specifically, the uploading stage mainly includes:
p1: the developer generates a multi-level label set for the model based on a multi-level label library preset in a learning-oriented market according to the task and the characteristics of the model to be uploaded, and the multi-level label set is used as a semantic specification; given development The person has a machine learning model Y to be uploaded 0 And the original data set D 0 According to a multi-level label library M preset in a learning part market, the model Y is obtained 0 Generating semantic conventions, denoted S 0 ＝{s 1 ,…,s n }；
P2: model Y uploaded to developer by server side of learning management system 0 And performing quality detection to judge whether the model is received. If the quality detection is passed, executing the next step P3; if the quality detection is not passed, prompting an error for the user to upload the model again;
p3: based on semantic conventions S 0 Determining a corresponding protocol islandThe protocol island->In which all semantic rules are placed about S 0 Is to learn the parts of the model;
p4 will reduce islandThe corresponding statistical protocol construction interface and parameters are sent to the equipment where the developer is located; wherein the parameters include a kernel function->
P5: the device where the developer is located constructs an interface and a kernel function through the statistical protocolRaw data set D for use with a model 0 Construction of a statistical protocol Φ for describing statistical information of a dataset 0 And uploading to a system;
p6: the system constructs a learning object L corresponding to the model based on the model of the received developer, the semantic conventions and the statistical conventions of the model 0 ＝(Y 0 ,S 0 ,Φ 0 )；
P7: is tied up withWill learn the part L 0 Placed in the protocol island corresponding to its model Is stored in a protocol islandTo complete the reception of the new learning object, thereby making about island +.>Updated accordingly as +.>
The process of the deployment phase mainly comprises the following steps:
q1: based on a multi-level tag library M preset by the learning-oriented library, a multi-level tag corresponding to a task of the user is obtained through a visual interface, so that the user requirement is converted into a keyword set U= { U which is convenient for subsequent matching of learning-oriented items 1 ,…,u n }；
Q2: the system performs semantic specification matching according to the keyword set provided by the user, and performs semantic similarity calculation to match to a learning piece specification island most relevant to the user requirement so as to complete pre-screening of learning pieces in the learning piece market; for example, for arbitrary semantic tag setsAnd-> (assume |S 1 |＝|S 2 |=n), a semantic similarity function is defined:
for cosine similarity, the semantic similarity function is defined as:
then, based on the keyword set U of the user, minimizing the semantic similarity function to obtain the optimization result which is the protocol island most relevant to the user requirementThus, the prescreening of the learning objects in the market is completed, and the solving process is expressed as follows:
min Sim(S′,U)，
I S′ ＝{(Y,S,Φ)|S＝S′} (3)
q3: slave protocol islandSelecting one or more representative learning pieces, transmitting a statistical protocol construction interface and parameters corresponding to a protocol island to a user for testing, and designating a series of information to be uploaded, such as model expression indexes and the like, to the user; for example, from the protocol island- >M representative learning objects are selected and denoted as +.>Statistical protocol construction interface and kernel function k corresponding to the protocol island S* Returning to the user;
q4: and the user tests the representative learning object on the self task and uploads the required test result to the server side. The user may voluntarily select whether to apply traditional metering protocols to help the system better identify the learning object that is helpful to the task. If agreeing, the user builds an interface and parameters through the received statistical protocol and builds an original data set collected by the user task, and the statistical protocol describing the user task distribution is uploaded to the system; wherein the userTesting models corresponding to representative learning objects on self tasksObtaining a test index set->Wherein-> Representing one or more evaluation indices, e.g. for model accuracy, the test index set is defined as +.>If the user agrees to the conventional rule, the user can pass through the local data set D u Kernel function k provided by AND system S* Construction of the corresponding statistical protocol Φ u ；
Q5: the system is based on the test index of the representative learning object on the user task and the statistical specification of the user dataCarrying out fine matching of statistical protocols in the protocol island, searching learning pieces which are most likely to be helpful to user tasks from the data distribution level; specifically, the system pair is for a series of test indicators t= { T 1 ,…,t m ' and any statistical conventions phi } 1 And phi is phi 2 Will define a protocol matching function F (T, phi) 1 ,Φ 2 ) For example, for statistical conventions, using abbreviated kernel mean embedding and without regard to the assistance of test indicators, the convention matching function is defined as:
next, the system determines the statistical specification phi according to the user u Index set returned by userT combination u In the protocol islandInternal minimization protocol matching function min F (T u ,Φ,Φ u ) Obtaining a matched result learning piece, wherein the solving process is expressed as follows:
min F(T u ,Φ,Φ u )
q6: the system returns the matched learning objects to the user and guides the user to multiplex the learning objects on own tasks; specifically, based on the final matching result, one or more learning objects L are obtained * ＝(Y * ,S * ,Φ * ) The learning item which is the most similar to the user task in terms of model semantics and data distribution level and is most likely to be helpful to the user task is returned to the user, and the user is guided to multiplex the learning item L on own task * 。
It should be noted that the foregoing examples are only for better illustrating the technical solution of the present invention, and not for limiting the present invention, and those skilled in the art should understand that any implementation of searching and multiplexing based on the embodiments of the present application should be included in the scope of the present invention.
According to the method, the semantic conventions and the statistical conventions are generated for the models of the developers through collecting the models of the developers and combined into the learning parts to be stored in the learning part library, and the learning parts which are helpful to the users are searched and matched based on the semantic conventions and the statistical conventions according to the machine learning task requirements of different users to be further multiplexed to the user tasks, so that the users can obtain the good-quality models which are helpful to the users to be multiplexed without training new models from scratch, and the efficiency is improved; according to the method, the model is further searched through the statistical protocol on the basis of the semantic protocol, so that the model helpful to the user can be accurately identified; the embodiment of the application does not contact the original data of the user and all model developers, so that the data privacy is strictly protected.
Fig. 3 shows a schematic structural diagram of a device for performing model searching and multiplexing based on a learning model according to an embodiment of the application. The device comprises: means for screening a plurality of candidate learning objects from a learning object library by matching based on semantic rules based on query information corresponding to a task requirement of a current user (hereinafter referred to as "semantic rule matching means 1"), means for selecting one or more learning objects from the plurality of candidate learning objects by matching based on statistical rules (hereinafter referred to as "statistical rule matching means 2"), and means for returning the selected one or more learning objects to the user for multiplexing of the corresponding model by the user (hereinafter referred to as "learning object returning and multiplexing means 3").
The present embodiments employ a learning model to store and manage models from different developers. Wherein the study includes a model and a specification describing the model. A study library containing numerous studies may be referred to as a study market.
The method of the embodiment of the application is applicable to various models, such as a linear model, a forest model, a support vector machine, a neural network model and the like.
Prior to the operation shown in fig. 3, the device receives models from various developers and builds corresponding learning objects. The device comprises a device for generating a semantic specification corresponding to a model based on the model uploaded by a developer (hereinafter referred to as a semantic specification generating device), a device for acquiring a statistical specification corresponding to the model according to the semantic specification corresponding to the model (hereinafter referred to as a statistical specification acquiring device), and a device for acquiring a learning object corresponding to the model based on the model and the semantic specification and the statistical specification corresponding to the model (hereinafter referred to as a learning object generating device).
The semantic specification generating device generates a semantic specification corresponding to the model based on the model uploaded by the developer.
Wherein the semantic conventions include one or more tag information that is used to describe the model. The semantic conventions may also include textual descriptive information, such as, for example, textual descriptive information of model processing tasks entered by the model uploader.
According to one embodiment, the semantic conventions include semantic tag sets based on a preset multi-level tag library. The multi-level tag library is a multi-level semantic tag describing the attribute of the learning object, such as task type, data information, model type and size, and the like. The semantic tags can be custom designed according to a model library.
According to one embodiment, the device detects the quality of the model uploaded by the developer, and judges whether to receive the model or not based on the detection result and stores the model in the learning-oriented library. If the quality detection result meets the requirement, the device receives the model and stores the model into a learning-class library; if the quality detection result does not meet the requirement, the device gives a corresponding prompt to the user.
The person skilled in the art should be familiar with the method, for example, the method can calculate the index reflecting the performance of the model, and detect whether the index value is better than the predetermined threshold value, etc., and the person skilled in the art can select a suitable method to perform the quality detection based on the actual requirement, which is not described herein.
Then, the statistical protocol obtaining device obtains the statistical protocol corresponding to the model according to the semantic protocol corresponding to the model.
Wherein the statistical conventions include information describing statistical distributions of the data sets. Preferably, the forms of statistical conventions include, but are not limited to, abbreviated kernel mean embedding, statistical thumbnail sets, cluster center points, distribution thumbnail sets.
Specifically, the statistical protocol obtaining device sends a statistical protocol construction interface and parameters corresponding to the semantic protocol corresponding to the model to equipment where a developer is located, so that the developer can construct a statistical protocol for describing statistical information of a data set according to the statistical protocol construction interface and parameters and an original data set used by the model, and obtain the statistical protocol uploaded by the developer.
According to one embodiment, the statistical protocol acquisition device determines a corresponding protocol island based on the semantic protocol corresponding to the model, and then sends the statistical protocol construction interface and parameters corresponding to the determined protocol island to the device where the developer is located. Wherein the parameters include, but are not limited to, kernel functions. And then, the equipment where the developer is located constructs an interface and parameters according to the statistical protocol, constructs a statistical protocol for describing statistical information of the data set according to the original data set used by the model, and uploads the statistical protocol to a learning library.
And the learning-oriented generating device obtains the corresponding learning-oriented of the model based on the model and the corresponding semantic and statistical conventions thereof.
According to one embodiment, the apparatus includes means for generating a plurality of protocol islands for storing a study (hereinafter referred to as "protocol island generating means"), and means for storing models received from respective developers based on semantic protocols corresponding to the models, placed in the respective protocol islands (hereinafter referred to as "study storing means").
The protocol island generating means generates a plurality of protocol islands for storing the study.
Wherein the models in the same protocol island address the same or similar and their learning task requirements.
The study storage device stores models received from various developers in corresponding protocol islands based on semantic protocols corresponding to the models, so that the models of the study contained in each protocol island have the same or similar semantic protocols.
And after receiving a new model, the learning-part storage device obtains the learning-part corresponding to the model based on the semantic protocol and the statistical protocol corresponding to the model, and correspondingly updates the protocol island corresponding to the model to finish adding the new learning-part.
The following description will refer to fig. 3, where the semantic rule matching apparatus 1 screens a plurality of candidate learning objects from the learning object library by matching based on the semantic rule based on the query information corresponding to the task requirement of the current user.
The query information comprises text information for querying a model in a learning-oriented library.
The means for obtaining the query information by the device includes, but is not limited to:
1) Generating one or more keywords corresponding to the task demands of the current user based on preset label information, and taking the keywords as corresponding query information; for example, based on a multi-level tag library preset by a learning component library, acquiring multi-level tags corresponding to tasks selected or input by a user through a visual interface, and taking corresponding keyword sets as query information;
2) Taking other text information input by a user as query information; for example, if the user does not provide the completed multi-level label, the text description information for the task uploaded by the user is used as query information.
According to one embodiment, the apparatus stores models from various developers in a plurality of protocol islands, each protocol island containing models of chemicals having the same or similar semantic protocols. The semantic specification matching device 1 determines a specification island with the highest semantic similarity with the query information by calculating the semantic similarity based on the query information, and takes the learning pieces in the specification island as candidate learning pieces.
The semantic similarity calculation includes various text similarity calculation modes, for example, cosine similarity calculation of two labels or optimal transportation distance calculation.
For example, in an application scene of sales volume prediction, based on a multi-level tag library preset by a learning library, a multi-level tag corresponding to a task of a user is obtained through a visual interface, and a keyword set 'regression task and sales volume prediction' corresponding to the selected tag is used as query information. Then, the semantic rule matching device 1 calculates the semantic similarity based on the query information, locates the learning rule island corresponding to the sales volume prediction task, and takes the learning rule in the rule island as a candidate learning rule, thereby screening a large number of learning models for solving the sales volume prediction task from the learning rule library.
Continuing with the description of fig. 3, the statistical specification matching device 2 selects one or more learning objects from the plurality of candidate learning objects by matching based on a statistical specification.
Wherein the device provides the selected one or more school components to the user as the school component most likely to be helpful to the user's task.
The statistical specification matching device 2 includes a device for sending a plurality of candidate learning objects to a user to obtain corresponding test feedback information (hereinafter referred to as "candidate learning object sending device"), and a device for matching among the plurality of candidate learning objects according to the test feedback information and the statistical specification of the user task to obtain one or more matched learning objects (hereinafter referred to as "sub-statistical specification matching device").
The candidate learning aid sending device sends the candidate learning aids to the user so as to obtain corresponding test feedback information.
Wherein the feedback information includes, but is not limited to, model accuracy, precision, recall, etc.
And the sub-statistical protocol matching device is used for matching among a plurality of candidate learning objects according to the test feedback information and the statistical rules of the user tasks to obtain one or more matched learning objects.
Wherein, matching of statistical conventions refers to metric computation between statistical conventions in combination with measurement indicators on representative learning pieces, including but not limited to distance computation in regenerated kernel hilbert space, and the like.
According to one embodiment, the device further screens the candidate learning objects to obtain a representative learning object, sends the representative learning object to the user to obtain corresponding test feedback information, and matches the test feedback information with the statistical rules of the user task in the representative learning objects to obtain one or more matched learning objects.
The method for screening the representative learning objects by the device comprises, but is not limited to, calculating a distance matrix among statistical conventions of the learning objects in the conventions island, using a clustering algorithm based on the distance matrix, taking the learning objects corresponding to the cluster center as the representative learning objects, and the like.
According to one embodiment, the device constructs the statistical specification of the user task through the preset statistical specification construction interface and parameters and the original data set collected by the user task under the condition of obtaining the user consent.
If the user does not wish to go to the traditional metering protocol, the device performs a learning fit based on the semantic protocol and the user's test feedback information.
The learning-part returning and multiplexing device 3 returns the selected one or more learning-parts to the user for the user to multiplex the corresponding models.
According to one embodiment, the learning-part returning and multiplexing device 3 guides the user to perform learning-part multiplexing, and the learning-part multiplexing includes integrating prediction results of the learning-parts on user tasks, fine-tuning the learning-part model by using user task data, expanding the prediction results of the learning-part model as additional features to user training data, and the like.
For example, in the case of sales prediction, the semantic specification matching apparatus 1 completes matching of semantic specifications based on calculation of semantic similarity of tags. Assuming the user fills out multi-level tags [ 'wholesale retail-department', 'supervised learning-regression', 'text data', 'integrated model','s complete identity, a number of models remain, assuming the method searches the library of learning components for a number of learning components that may be helpful to the user's task. Then, the statistical protocol matching device 2 performs statistical protocol searching on the learning components based on the protocol matching function according to the statistical protocol uploaded by the user, for example, when the abbreviated kernel mean value embedding is adopted, the learning component returning and multiplexing device 3 returns the selected one or more learning components closest to the statistical protocol of the user to the user for the user to solve own learning task.
According to the device provided by the embodiment of the application, the models of the developers are collected, semantic conventions and statistical conventions are generated for the models and combined into the learning parts to be stored in the learning part library, and the learning parts which are helpful to the users are searched and matched based on the semantic conventions and the statistical conventions according to the machine learning task requirements of different users to be further multiplexed in the user tasks, so that the users can obtain the good-quality models which are helpful to the users without training the new models from the beginning to multiplex, and the efficiency is improved; the device of the embodiment of the application further searches the model through the statistical protocol on the basis of the semantic protocol, so that the model helpful to the user can be accurately identified; the embodiment of the application does not contact the original data of the user and all model developers, so that the data privacy is strictly protected.
The software program of the present invention may be executed by a processor to perform the steps or functions described above. Likewise, the software programs of the present invention (including associated data structures) may be stored on a computer readable recording medium, such as RAM memory, magnetic or optical drive or diskette and the like. In addition, some of the steps or functions of the present invention may be implemented in hardware, for example, as circuitry that cooperates with the processor to perform various functions or steps.
Furthermore, portions of the present invention may be implemented as a computer program product, such as computer program instructions, which when executed by a computer, may invoke or provide methods and/or techniques in accordance with the present invention by way of operation of the computer. Program instructions for invoking the inventive methods may be stored in fixed or removable recording media and/or transmitted via a data stream in a broadcast or other signal bearing medium and/or stored within a working memory of a computer device operating according to the program instructions. An embodiment according to the invention comprises an apparatus comprising a memory for storing computer program instructions and a processor for executing the program instructions, wherein the computer program instructions, when executed by the processor, trigger the apparatus to operate a method and/or a solution according to the embodiments of the invention as described above.
It will be evident to those skilled in the art that the invention is not limited to the details of the foregoing illustrative embodiments, and that the present invention may be embodied in other specific forms without departing from the spirit or essential characteristics thereof. The present embodiments are, therefore, to be considered in all respects as illustrative and not restrictive, the scope of the invention being indicated by the appended claims rather than by the foregoing description, and all changes which come within the meaning and range of equivalency of the claims are therefore intended to be embraced therein. Any reference sign in a claim should not be construed as limiting the claim concerned. Furthermore, it is evident that the word "comprising" does not exclude other elements or steps, and that the singular does not exclude a plurality. A plurality of units or means recited in the system claims can also be implemented by means of software or hardware by means of one unit or means. The terms first, second, etc. are used to denote a name, but not any particular order.

Claims (11)

1. A method for model searching and multiplexing based on a learning model, wherein the method comprises:
based on query information corresponding to the task demands of the current user, a plurality of candidate learning pieces are screened from the learning piece library by matching based on semantic conventions;
selecting one or more learning objects from the plurality of candidate learning objects by matching based on a statistical specification;
and returning the selected one or more learning objects to the user for multiplexing the corresponding models by the user.
2. The method of claim 1, wherein the study includes a model and a specification describing the model, and the library of studies includes studies from individual developers.
3. The method of claim 1 or 2, wherein the semantic conventions include one or more tag information describing a model, the method comprising:
and generating one or more keywords corresponding to the task demands of the current user based on the preset label information, and taking the keywords as corresponding query information.
4. A method according to any one of claims 1 to 3, wherein the method stores models from individual developers in a plurality of protocol islands, each protocol island containing models of chemicals having the same or similar semantic protocols;
The step of screening a plurality of candidate learning objects from the learning object library by matching based on the query information corresponding to the current user task requirement and based on the semantic rules comprises the following steps:
and determining a protocol island with the highest semantic similarity with the query information by calculating the semantic similarity based on the query information, and taking the learning item in the protocol island as a candidate learning item.
5. The method of claim 1 or 2, wherein the statistical conventions include information describing a statistical distribution of the dataset, the step of selecting one or more learning elements from the plurality of candidate learning elements by matching based on the statistical conventions comprising:
sending the candidate learning articles to a user to obtain corresponding test feedback information;
and matching among a plurality of candidate learning pieces according to the test feedback information and the statistical rule of the user task to obtain one or more matched learning pieces.
6. The method according to claim 5, wherein the method comprises:
under the condition of obtaining user consent, constructing an interface and parameters through a preset statistical protocol, and constructing a statistical protocol of the user task through an original data set collected by the user task.
7. The method according to claim 1 or 2, wherein the method comprises:
based on the model uploaded by the developer, generating a semantic specification corresponding to the model;
according to the semantic protocol corresponding to the model, acquiring a statistical protocol corresponding to the model;
and obtaining a corresponding learning piece of the model based on the model and the corresponding semantic and statistical conventions thereof.
8. The method of claim 7, wherein the method comprises:
generating a plurality of protocol islands for storing the study;
for models received from individual developers, the models are stored in respective protocol islands based on their corresponding semantic protocols, such that the models of the chemicals contained in each protocol island have the same or similar semantic protocols.
9. A device for model searching and multiplexing based on a learning model form, wherein the device comprises:
means for screening a plurality of candidate learning objects from the learning object library by matching based on semantic conventions based on query information corresponding to a current user task demand;
means for selecting one or more learning objects from the plurality of candidate learning objects by matching based on a statistical specification;
and means for returning the selected one or more learning objects to the user for multiplexing by the user the corresponding models.
10. A computer device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor implements the method of any one of claims 1 to 8 when the program is executed by the processor.
11. A computer-readable storage medium, on which a computer program is stored, characterized in that the program, when being executed by a processor, implements the method according to any one of claims 1 to 8.
CN202310396671.3A 2023-04-13 2023-04-13 Universal method and device for searching and multiplexing learning objects Pending CN116542136A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202310396671.3A CN116542136A (en) 2023-04-13 2023-04-13 Universal method and device for searching and multiplexing learning objects

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202310396671.3A CN116542136A (en) 2023-04-13 2023-04-13 Universal method and device for searching and multiplexing learning objects

Publications (1)

Publication Number Publication Date
CN116542136A true CN116542136A (en) 2023-08-04

Family

ID=87444406

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202310396671.3A Pending CN116542136A (en) 2023-04-13 2023-04-13 Universal method and device for searching and multiplexing learning objects

Country Status (1)

Country Link
CN (1) CN116542136A (en)

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20040078190A1 (en) * 2000-09-29 2004-04-22 Fass Daniel C Method and system for describing and identifying concepts in natural language text for information retrieval and processing
CN108418833A (en) * 2018-03-23 2018-08-17 中科创达软件股份有限公司 A kind of management method of software, cloud server and terminal
CN109063221A (en) * 2018-11-02 2018-12-21 北京百度网讯科技有限公司 Query intention recognition methods and device based on mixed strategy
CN115437627A (en) * 2021-06-04 2022-12-06 中国移动通信集团浙江有限公司 Method, device and equipment for developing learning object and storage medium

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20040078190A1 (en) * 2000-09-29 2004-04-22 Fass Daniel C Method and system for describing and identifying concepts in natural language text for information retrieval and processing
CN108418833A (en) * 2018-03-23 2018-08-17 中科创达软件股份有限公司 A kind of management method of software, cloud server and terminal
CN109063221A (en) * 2018-11-02 2018-12-21 北京百度网讯科技有限公司 Query intention recognition methods and device based on mixed strategy
CN115437627A (en) * 2021-06-04 2022-12-06 中国移动通信集团浙江有限公司 Method, device and equipment for developing learning object and storage medium

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
"Learnware: Small Models Do Big", 《ARXIV:2210.03647V1》, pages 1 - 19 *
张沪寅;李鑫;陆春涛;吴笛;: "LMOD:一种基于本体描述的电子学习资源库模型", 计算机应用研究, no. 04 *

Similar Documents

Publication Publication Date Title
US11631234B2 (en) Automatically detecting user-requested objects in images
US9910910B2 (en) Syntactic graph modeling in a functional information system
CN108292310A (en) For the relevant technology of digital entities
US11768869B2 (en) Knowledge-derived search suggestion
US20170220589A1 (en) Item recommendation method, device, and system
CN102402593A (en) Multi-modal approach to search query input
CN104160390A (en) Entity augmentation service from latent relational data
CN109684625A (en) Entity handles method, apparatus and storage medium
US11915129B2 (en) Method and system for table retrieval using multimodal deep co-learning with helper query-dependent and query-independent relevance labels
US20220100963A1 (en) Event extraction from documents with co-reference
US10755332B2 (en) Multi-perceptual similarity detection and resolution
CN111291765A (en) Method and device for determining similar pictures
US20220100772A1 (en) Context-sensitive linking of entities to private databases
Chen et al. Integrating information theory and adversarial learning for cross-modal retrieval
CN116601626A (en) Personal knowledge graph construction method and device and related equipment
WO2024099037A1 (en) Data processing method and apparatus, entity linking method and apparatus, and computer device
CN113761219A (en) Knowledge graph-based retrieval method and device, electronic equipment and storage medium
US20230028266A1 (en) Product recommendation to promote asset recycling
US20220121823A1 (en) System and method for artificial intelligence driven document analysis, including searching, indexing, comparing or associating datasets based on learned representations
CN113987274A (en) Video semantic representation method and device, electronic equipment and storage medium
CN114492601A (en) Resource classification model training method and device, electronic equipment and storage medium
CN109635004B (en) Object description providing method, device and equipment of database
US8577814B1 (en) System and method for genetic creation of a rule set for duplicate detection
CN116796027A (en) Commodity picture label generation method and device, equipment, medium and product thereof
CN116340635A (en) Article recommendation method, model training method, device and equipment

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination