CN109214501A - The method and apparatus of information for identification - Google Patents
The method and apparatus of information for identification Download PDFInfo
- Publication number
- CN109214501A CN109214501A CN201710513137.0A CN201710513137A CN109214501A CN 109214501 A CN109214501 A CN 109214501A CN 201710513137 A CN201710513137 A CN 201710513137A CN 109214501 A CN109214501 A CN 109214501A
- Authority
- CN
- China
- Prior art keywords
- type
- information
- probability
- neural network
- model
- 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.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/047—Probabilistic or stochastic networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Abstract
This application discloses the method and apparatus of information for identification.One specific embodiment of this method includes: to obtain information to be identified；The information to be identified is imported into neural network model, generate various types of probability that the information to be identified is belonging respectively at least three kinds of preset kinds, and according to the probability of generation identify the information to be identified belonging to type, wherein, the neural network model is obtained according to updated initial neural network model, initial neural network model is updated based on following steps: the first training sample being imported initial neural network model, first training sample is generated and belongs to various types of probability；It obtains at least one and distinguishes difficulty weight；Based on this, at least one distinguishes difficulty weight and first training sample belongs to various types of probability, determines the model error of initial neural network model；Initial neural network model is updated according to the model error.This embodiment improves the accuracys rate of information identification.
Description
Technical field
This application involves field of computer technology, and in particular to Internet technical field, more particularly, to identification information
Method and apparatus.
Background technique
With the development of computer technology, it can use computer and information identified, such as to pictorial information, text
Information, acoustic information etc. are identified.In the prior art, usually information is identified using the model pre-established.
However, existing information identification method, is usually difficult to differentiate between more similar information, accordingly, there are information
The lower problem of recognition accuracy.
Summary of the invention
The purpose of the application is the method and apparatus for proposing a kind of improved information for identification, to solve background above
The technical issues of technology segment is mentioned.
In a first aspect, the embodiment of the present application provides a kind of method of information for identification, the above method include: obtain to
Identification information；Above-mentioned information to be identified is imported into neural network model, above-mentioned information to be identified is generated and is belonging respectively at least three kinds
Various types of probability in preset kind, and according to the probability of generation identify above-mentioned information to be identified belonging to type,
In, above-mentioned neural network model is obtained according to updated initial neural network model, and initial neural network model is based on following
Step updates: the first training sample being imported initial neural network model, above-mentioned first training sample of generation belongs to above-mentioned various
The probability of type, wherein above-mentioned initial neural network model belongs to above-mentioned various types of probability for characterization information and information
Between corresponding relationship；It obtains at least one and distinguishes difficulty weight, wherein distinguish difficulty weight for characterizing Distinguish centering
The difficulty of two kinds of information, type are formed to by the two types in above-mentioned at least three kinds of preset kinds；Based on it is above-mentioned extremely
A few differentiation difficulty weight and above-mentioned first training sample belong to above-mentioned various types of probability, determine initial neural network mould
The model error of type；Initial neural network model is updated according to above-mentioned model error.
Second aspect, the embodiment of the present application provide a kind of device of information for identification, and above-mentioned apparatus includes: to obtain list
Member, for obtaining information to be identified；Recognition unit generates above-mentioned for above-mentioned information to be identified to be imported neural network model
Information to be identified is belonging respectively to various types of probability at least three kinds of preset kinds, and according in the identification of the probability of generation
State type belonging to information to be identified, wherein above-mentioned neural network model is obtained according to updated initial neural network model,
What initial neural network model was updated by updating unit, above-mentioned updating unit includes: generation module, for training sample for first
The initial neural network model of this importing generates above-mentioned first training sample and belongs to above-mentioned various types of probability, wherein above-mentioned first
Beginning neural network model belongs to the corresponding relationship between above-mentioned various types of probability for characterization information and information；Obtain mould
Block distinguishes difficulty weight for obtaining at least one, wherein distinguish difficulty weight for characterizing Distinguish centering two types
Information difficulty, type forms to by the two types in above-mentioned at least three kinds of preset kinds；Determining module, for being based on
It states at least one differentiation difficulty weight and above-mentioned first training sample belongs to above-mentioned various types of probability, determine initial nerve net
The model error of network model；Update module, for updating initial neural network model according to above-mentioned model error.
The third aspect, the embodiment of the present application provide a kind of server, and above-mentioned server includes: one or more processing
Device；Storage device, for storing one or more programs, when said one or multiple programs are by said one or multiple processors
It executes, so that said one or multiple processors realize the method such as first aspect.
Fourth aspect, the embodiment of the present application provide a kind of computer readable storage medium, are stored thereon with computer journey
Sequence realizes the method such as first aspect when the program is executed by processor.
The method and apparatus of information for identification provided by the embodiments of the present application, by using carry out information identification
Neural network model, by determining model error according to above-mentioned differentiation difficulty weight, model error can be made and distinguishing difficulty
Weight is related.For popular, it is difficult to which the type of differentiation is larger to corresponding differentiation difficulty weight, then is difficult to differentiate between type to mould
The model error contribution of type is larger, if the separating capacity of type pair of the model for being difficult to differentiate between is poor, model error is also
Being can be larger.During reducing model error, model can be made to improve the separating capacity for being difficult to differentiate between type pair, from
And it can be with the accuracy rate of identification information.
Detailed description of the invention
By reading a detailed description of non-restrictive embodiments in the light of the attached drawings below, the application's is other
Feature, objects and advantages will become more apparent upon:
Fig. 1 is that this application can be applied to exemplary system architecture figures therein；
Fig. 2 is the flow chart according to one embodiment of the method for the information for identification of the application；
Fig. 3 a is the flow chart according to a kind of implementation of the method for the information for identification of the application；
Fig. 3 b is the flow chart for realizing the step 303 in process shown in Fig. 3 a；
Fig. 4 is the schematic diagram according to an application scenarios of the method for the information for identification of the application；
Fig. 5 is the flow chart according to another implementation of the method for the information for identification of the application；
Fig. 6 is the structural schematic diagram according to one embodiment of the device of the information for identification of the application；
Fig. 7 is adapted for the structural schematic diagram for the computer system for realizing the server of the embodiment of the present application.
Specific embodiment
The application is described in further detail with reference to the accompanying drawings and examples.It is understood that this place is retouched
The specific embodiment stated is used only for explaining related invention, rather than the restriction to the invention.It also should be noted that in order to
Convenient for description, part relevant to related invention is illustrated only in attached drawing.
It should be noted that in the absence of conflict, the features in the embodiments and the embodiments of the present application can phase
Mutually combination.The application is described in detail below with reference to the accompanying drawings and in conjunction with the embodiments.
Fig. 1 is shown can be using the method or the implementation of the device of information for identification of the information for identification of the application
The exemplary system architecture 100 of example.
As shown in Figure 1, system architecture 100 may include terminal device 101,102,103, network 104 and server 105.
Network 104 between terminal device 101,102,103 and server 105 to provide the medium of communication link.Network 104 can be with
Including various connection types, such as wired, wireless communication link or fiber optic cables etc..
User can be used terminal device 101,102,103 and be interacted by network 104 with server 105, to receive or send out
Send message etc..Various telecommunication customer end applications can be installed, such as Image Acquisition class is answered on terminal device 101,102,103
With the application of, video playback class, picture display editor class application, instant messaging tools, mailbox client, social platform software etc..
Terminal device 101,102,103 can be with display screen and support picture or video to upload, browsing it is various
Electronic equipment, including but not limited to smart phone, tablet computer, E-book reader, MP3 player (Moving Picture
Experts Group Audio Layer III, dynamic image expert's compression standard audio level 3), MP4 (Moving
Picture Experts Group Audio Layer IV, dynamic image expert's compression standard audio level 4) player, knee
Mo(u)ld top half portable computer and desktop computer etc..
Server 105 can be to provide the server of various services, such as terminal device 101,102,103 is uploaded or shown
The information (such as picture or video) shown provides the background server supported.Background server can be to the information (example received
Such as picture or video) data carry out the processing such as analyzing, it is alternatively possible to by processing result (such as the picture or view identified
Type belonging to frequency) feed back to terminal device.
It should be noted that the method for information is generally held by server 105 for identification provided by the embodiment of the present application
Row, correspondingly, the device of information is generally positioned in server 105 for identification.
It should be understood that the number of terminal device, network and server in Fig. 1 is only schematical.According to realization need
It wants, can have any number of terminal device, network and server.It should be noted that in some implementations, Fig. 1
Shown in can not also include terminal device and network in exemplary system architecture.
With continued reference to Fig. 2, it illustrates the processes according to one embodiment of the method for the information for identification of the application
200.The method of above-mentioned information for identification, comprising the following steps:
Step 201, information to be identified is obtained.
In the present embodiment, the first electronic equipment of the method operation of information thereon is (such as shown in FIG. 1 for identification
Server) locally or from other electronic equipments information to be identified can be obtained from above-mentioned first electronic equipment.
As an example, information to be identified can be text information, pictorial information, video information or acoustic information etc..
Step 202, information to be identified is imported into neural network model, generates information to be identified and is belonging respectively at least three kinds in advance
If various types of probability in type, and according to the probability of generation identify information to be identified belonging to type.
In the present embodiment, the first electronic equipment of the method operation of information thereon is (such as shown in FIG. 1 for identification
Server) information to be identified that can first obtain step 201 imports neural network model, then generate above-mentioned letter to be identified
Breath belongs to various types of probability at least three kinds of preset kinds, finally identifies above-mentioned information to be identified according to the probability of generation
Affiliated type.
As an example, information to be identified can be animal picture, at least three kinds of preset kinds can be respectively dog class picture,
Tiger picture and cat class picture.Animal picture is imported into neural network model, available above-mentioned animal picture is belonging respectively to dog
The probability 80% of the probability 10% of class picture, the probability 20% of tiger picture and cat class picture.It can be by the 80% of maximum probability
Indicated this type of cat class picture, is determined as type belonging to above-mentioned animal picture.
In the present embodiment, above-mentioned neural network model can be obtained according to updated initial neural network model, be asked
With reference to Fig. 3 a, it illustrates the exemplary flow 300 for updating initial neural network model, initial neural network model can be based on
Step 301, step 302, step 303 and step 304 in process 300 update:
Step 301, the first training sample is imported into initial neural network model, the first training sample of generation belongs to various types of
The probability of type.
In the present embodiment, the second electronic equipment (such as server shown in FIG. 1) can import the first training sample
Initial neural network model generates above-mentioned first training sample and belongs to above-mentioned various types of probability.Herein, above-mentioned initial mind
Belong to the corresponding relationship between above-mentioned various types of probability for characterization information and information through network model.
In the present embodiment, the first training sample can be one, be also possible to multiple.
It should be noted that the second electronic equipment can be identical as the first electronic equipment, can also be different.Step 301- step
Rapid 304 executing subject can pre-establish above-mentioned neural network model before step 201, then by the neural network of foundation
Model is used for the executing subject of step 201 and step 202.
In the present embodiment, above-mentioned initial neural network model can be the neural network model built in advance.About such as
What build initial neural network model be it is known in those skilled in the art, details are not described herein.
Step 302, it obtains at least one and distinguishes difficulty weight.
In the present embodiment, at least one available differentiation difficulty weight of above-mentioned electronic equipment.Herein, difficulty is distinguished
Weight is used to characterize the difficulty of the two kinds of information of Distinguish centering.Herein, type is to default by above-mentioned at least three kinds
Two types composition in type.
As an example, at least three kinds of preset kinds can be dog class picture, tiger picture and cat class picture respectively.Type pair
It can be made of dog class picture and tiger picture both types, type dog class picture and cat class picture to can be made of, class
Type tiger picture and cat cat figure to can also be made of.The differentiation difficulty weight of acquisition can be some one or more: be used for
Characterization distinguishes the weight of the difficulty of dog class picture and tiger picture, difficulty for characterize differentiation dog class picture and cat class picture
Weight, for characterizing the weight for distinguishing the difficulty of tiger picture and cat class picture.
In some optional implementations of the present embodiment, each differentiation difficulty weight can be pre-set default
Value.
In some optional implementations of the present embodiment, each differentiation difficulty weight can be through the following steps that obtain
To: obtain various types of information.Determine the similarity between various types of information.According between various types of information
Similarity determines above-mentioned differentiation difficulty weight.
As an example, at least three kinds of preset kinds can be dog class picture type, tiger picture type and cat class figure respectively
Sheet type.The picture of the picture of dog class picture type, the picture of tiger picture type and cat class picture type can be obtained respectively.
Calculate separately to obtain similarity 10% between the picture of dog class picture type and the picture of tiger picture type, dog class picture category
The picture and cat class picture of similarity 20% and tiger picture type between the picture of type and the picture of cat class picture type
Similarity 60% between the picture of type.Above-mentioned similarity 10% can be determined as dog class picture type and tiger picture category
Above-mentioned similarity 20% can be determined as between dog class picture type and cat class picture type by the differentiation difficulty weight between type
Differentiation difficulty weight, the differentiation that above-mentioned similarity 60% can be determined as between tiger picture type and cat class picture type
Difficulty weight.
Step 303, difficulty weight is distinguished based at least one and the first training sample belongs to various types of probability, determined
The model error of initial neural network model.
In the present embodiment, above-mentioned second electronic equipment can distinguish difficulty based at least one acquired in step 302
Above-mentioned first training sample that weight and step 301 generate belongs to above-mentioned various types of probability, determines above-mentioned initial nerve net
The model error of network model.
It will be understood by those skilled in the art that model error can be used for characterization model reality output and desired output it
Between gap.It should be noted that existing training method is by the way of to each type fair play for disaggregated model
Training pattern.But may be different for the differentiation difficulty between different type pair, and the traditional training method of this point is not
Consider.
In some optional implementations of the present embodiment, the first training sample is identified with affiliated type in advance, the
One training sample has multiple.Step 303 can be accomplished by the following way: selecting the first training sample and belong to each type
Maximum probability in probability.The type that the corresponding type of maximum probability identifies in advance with the first training sample is compared.
If the result compared is identical, it is determined that above-mentioned neural network model identification is correct；If compare result be it is not identical,
Determine above-mentioned neural network model identification mistake.For the various types at least three types, above-mentioned neural network mould is determined
The type ratio that the first training sample identification of the type is wrong.For each type pair of at least one type centering, determine
The type determines that the differentiation of the sum and the type pair of the corresponding ratio of any two types is difficult to the sum of corresponding ratio
Spend the product of weight.Identified product is summed it up, as above-mentioned model error.
In some optional implementations of the present embodiment, step 303 can also be by process 303 shown in Fig. 3 b
Step 3031, step 3032 and step 3033 realize:
Step 3031, various types of probability are belonged to according to the first training sample, determined for characterizing the first training sample
Belong to the probable error of various probability.
In this implementation, above-mentioned second electronic equipment can belong to various types of general according to the first training sample
Rate is determined for characterizing the probable error that the first training sample belongs to various types of probability.
As an example, can use following formula determines probable error:
zi=xi-max(x1, x2..., xN) (1)
Wherein: xiIt is the probability that the first training sample belongs to the i-th seed type, the value range of i is 1,2,3 ..., N；x1,
x1..., xNIn x1It is the probability that the first training sample belongs to the 1st seed type, x2It is that the first training sample belongs to the 2nd seed type
Probability, xNIt is the probability that the first training sample belongs to N seed type, N is the number of species of preset kind；Max () expression takes
Maximum value；ZjIn j value range be 1,2,3 ..., N；As it can be seen that when the value of j is 1, Z1Value can basis
Formula (1) is determining, and so on, it can determine each Zj；liBelong to the probability of the probability of the i-th seed type for the first training sample
Error, the Z that formula (1) is acquirediIt brings into formula (2), available each li。
Step 3032, for each type pair of at least one type centering, according to the two kinds of of the type centering
The sum for distinguishing difficulty weight and corresponding probable error, determines the weight probable error of the type pair.
In this implementation, above-mentioned second electronic equipment can for each type pair of at least one type centering,
According to the sum of two kinds of the differentiation difficulty weight and corresponding probable error of the type centering, the weight of the type pair is determined
Probable error.
As an example, can be by for each type pair of at least one type centering, by two types of the type centering
The product of the sum of the differentiation difficulty weight and corresponding probable error of type, is determined as the weight probable error of the type pair.
Step 3033, according to identified weighted error, model error is determined.
In this implementation, above-mentioned second electronic equipment can determine model according to identified weight probable error
Error.
As an example, the sum of at least one identified weight probable error can be determined as above-mentioned model error.
Step 304, initial neural network model is updated according to model error.
In the present embodiment, the model error that above-mentioned second electronic equipment can be determined according to step 303 updates above-mentioned first
Beginning neural network model.
It should be noted that being provided in the prior art more how according to model error more this technology of new model itself
Kind method, those skilled in the art can realize step 304 based on the prior art, and details are not described herein.
It should be noted that initial neural network model is updated based on step 301- step 304, execution step can be
301- step 304 is once to once be updated, alternatively, executing step 301- step 304 more times to repeatedly be updated.
Step 301- step 304 circulation execute it is multiple when, can when time recycle in the updated initial neural network model of step 304
As the initial neural network model in step 301 in circulation next time.Above-mentioned neural network model can be updated mind
Through network model.
With continued reference to the signal that Fig. 4, Fig. 4 are according to the application scenarios of the method for the information for identification of the present embodiment
Figure.In the application scenarios of Fig. 4, firstly, server 401 can be to the initial neural network model 402 built in advance, based on more
New technological process 403 is updated, according to updated initial neural network neural network model 404, it should be noted that more
New technological process can be with reference to the explanation in step 301- step 304, and details are not described herein.Then, user can use terminal 405 to
Server 401 sends information 406 to be identified, as an example, information to be identified is the picture with cat image.After again, server
Available above-mentioned information to be identified.Finally, above-mentioned information 406 to be identified can be imported neural network model by server 401
404, it generates above-mentioned information to be identified and is belonging respectively to various types of probability at least three kinds of preset kinds, and according to generation
Probability identify type belonging to above-mentioned information to be identified, as an example, output " cat class picture " this type.
It is the main source of model error, the above-mentioned implementation of the application by the data separation mistake for the type being difficult to differentiate between
Example provide method, using carries out information identification neural network model, by according to above-mentioned differentiation difficulty weight determination
Model error can make model error related to difficulty weight is distinguished.For popular, it is difficult to which the type of differentiation is to corresponding area
Point difficulty weight is larger, then it is larger to the model error contribution to model to be difficult to differentiate between type, if model is for being difficult to differentiate between
Type pair separating capacity it is poor, then model error still can be larger.During reducing model error, mould can be made
Type improves the separating capacity for being difficult to differentiate between type pair, it is thus possible to the accuracy rate of identification information.Contrastingly, existing mould
Type misses the source that method for determining difference does not distinguish model error, and various types of probability indifferences are treated, are caused according to mould
When type error update model, even if model error very little, but the type pair for being difficult to differentiate between, the area of neural network model
Point ability or poor.
In some optional implementations of the present embodiment, referring to FIG. 5, it illustrates update initial neural network mould
A kind of exemplary flow 500 of type, initial neural network model are also based on step 501, step 502, step in process 500
Rapid 503, step 504, step 505 and step 506 update:
Step 501, the first training sample is imported into initial neural network model, the first training sample of generation belongs to various types of
The probability of type.
In the present embodiment, the second electronic equipment (such as server shown in FIG. 1) can import the first training sample
Initial neural network model generates above-mentioned first training sample and belongs to above-mentioned various types of probability.
It should be noted that the realization details of step 501 can be no longer superfluous herein respectively with reference to the explanation in step 301
It states.
Step 502, for each type pair of at least one type centering, two classification moulds of the training for the type pair
Type.
In the present embodiment, the second electronic equipment (such as server shown in FIG. 1) can be at least one type pair
In each type pair, training be directed to the type pair two disaggregated models.Herein, two disaggregated model for characterization information with
Indicate the corresponding relationship between information, instruction information is used to indicate two kinds of one kind that information belongs to the type centering.
In some optional implementations of the present embodiment, step 502 can be accomplished by the following way: for above-mentioned
Each type of each type centering of at least one type centering obtains and is labeled with the type in advance and belongs to the general of the type
Second sample set of rate.Probability in response to belonging to the type is less than the second training sample of predetermined probabilities threshold value in the set
In ratio be less than preset ratio threshold value, based on belong to the type probability be less than predetermined probabilities threshold value the second training sample,
Amplification obtains amplified sample.Amplified sample and the second training sample in the set are merged, obtain merging sample set.
For each type pair of at least one type centering, the corresponding merging sample set of the two types of the type centering is utilized
It closes, training is directed to two disaggregated models of the type pair.
In this implementation, the probability that the second training sample belongs to the type is less than predetermined probabilities threshold value, then can incite somebody to action
This second training sample is determined as being difficult to sample (hard example).If being difficult to ratio of the sample in the set
Example is less than preset ratio threshold value, can be expanded based on sample is difficult to, obtain one or more amplified samples.
It in this implementation, is expanded based on sample mode is difficult to, can be and utilize one or more of
Mode is handled sample is difficult to: up-sampling, down-sampling, mirror image etc..Particularly, it for pictorial information, can use
One or more of mode is expanded: being cut out, small rotation, mirror image, is changed channel color, and slight fuzzy.
In this implementation, it should be noted that amplified sample is also to be difficult to sample, and amplified sample is marked
It is identical that the type of note and amplification were based on is difficult to sample.It can be set to merge and be difficult to sample (packet in sample set
Include amplified sample and belong to the probability of the type less than the second training sample of predetermined probabilities threshold value) (belong to such with normal sample
The probability of type be not less than the second training sample of predetermined probabilities threshold value) ratio.
It should be noted that under normal circumstances, in the second training sample set of the type, it is difficult to identify sample
Ratio may be smaller, two disaggregated models trained in this way, it is more likely that for being difficult to sample without preferably identifying energy
Power.Merge in sample set and is difficult to two classification that sample proportion is higher, trains using such merging sample set
Model, can preferably identify the sample that is difficult to of the type, and then can better discriminate between the targeted class of two disaggregated models
The two types of type centering.
In some optional implementations of the present embodiment, obtains and be labeled with the type in advance and belong to the general of the type
Second sample set of rate, can be accomplished by the following way: can be by staff previously according to experience marking types and general
Rate.
In some optional implementations of the present embodiment, obtains and be labeled with the type in advance and belong to the general of the type
Second sample set of rate, can be accomplished by the following way: obtain the second training sample for being labeled with the type in advance.It will be pre-
The second training sample for first marking the type imports the generative probabilistic model pre-established, determines the second of mark the type in advance
Training sample belongs to the probability of the type.Herein, above-mentioned generative probabilistic model is for characterizing the second training sample and the second instruction
Practice sample and is belonging respectively to the corresponding relationship between above-mentioned various types of probability.
In this implementation, generative probabilistic model can be the model that recognition capability is told somebody what one's real intentions are, for the second training sample
Originally the probability for belonging to the type marked in advance carries out preresearch estimates, so that determination is difficult to sample later.
Step 503, using the test sample for being labeled with type in advance, the model error of each two disaggregated model is determined.
In the present embodiment, the second electronic equipment can use the test sample for being labeled with type in advance, determine each two
The model error of disaggregated model.
As an example, can first choose two kinds of test sample, the test selected for each two disaggregated model
Sample is labeled with the type of the targeted type centering of two disaggregated model.The test sample selected is imported into the two classification mould
Type obtains the probability for belonging to both types, and the biggish type of probability is determined as to the type of test sample.Judgement using this two
Whether the type that disaggregated model determines and the type that test sample marks are consistent.Inconsistent ratio is counted, to determine this two points
The error of class model.
Step 504, for each of being trained two disaggregated models, according to the model error of two disaggregated model, determining should
The differentiation difficulty weight of the targeted type pair of two disaggregated models.
In the present embodiment, the second electronic equipment can be for each of being trained two disaggregated models, according to two classification
The model error of model determines the differentiation difficulty weight of the targeted type pair of two disaggregated model.
As an example, the model error of two disaggregated model can be determined as the targeted type of two disaggregated model
Pair differentiation difficulty weight.
Step 505, difficulty weight is distinguished based at least one and the first training sample belongs to various types of probability, determined
The model error of initial neural network model.
In the present embodiment, the second electronic equipment can distinguish difficulty weight and the first training sample category based at least one
In various types of probability, the model error of initial neural network model is determined.
Step 506, initial neural network model is updated according to model error.
In the present embodiment, the model error that above-mentioned second electronic equipment can be determined according to step 505 updates above-mentioned first
Beginning neural network model.
It should be noted that the realization details of step 505 and step 506 can be respectively with reference in step 303 and step 304
Explanation, details are not described herein.
It should be noted that initial neural network model is updated based on step 501- step 505, execution step can be
501- step 506 is once to once being updated, alternatively, step 501- step 506 more times to repeatedly be updated.Step
501- step 506 circulation execute it is multiple when, can when time recycle in the updated initial neural network model conduct of step 506
Initial neural network model in circulation in step 501 next time.Above-mentioned neural network model can be updated nerve net
Network model.
From figure 5 it can be seen that Fig. 5 is corresponding just compared with the more new technological process of the corresponding initial neural network model of Fig. 3
The more new technological process of beginning neural network model is highlighted to the step of difficulty weight is determined is distinguished, thus, it is possible to introduce more
Difficulty weight is accurately distinguished, to realize more effective information identification.
With further reference to Fig. 6, as the realization to method shown in above-mentioned each figure, this application provides one kind to believe for identification
One embodiment of the device of breath, the Installation practice is corresponding with embodiment of the method shown in Fig. 2, which can specifically answer
For in various electronic equipments.
As shown in fig. 6, the device 600 of the above-mentioned information for identification of the present embodiment includes: that acquiring unit 601, identification are single
Member 602 and updating unit 603, also, updating unit includes generation module 6031, obtains module 6032,6033 and of determining module
Update module 6034.Wherein, acquiring unit, for obtaining information to be identified；Recognition unit, for leading above-mentioned information to be identified
Enter neural network model, generate above-mentioned information to be identified and be belonging respectively to various types of probability at least three kinds of preset kinds,
And according to the probability of generation identify above-mentioned information to be identified belonging to type, wherein above-mentioned neural network model is according to update
Initial neural network model afterwards obtains, and initial neural network model is updated by updating unit, and above-mentioned updating unit includes:
Generation module, for the first training sample to be imported initial neural network model, above-mentioned first training sample of generation belongs to above-mentioned
Various types of probability, wherein above-mentioned initial neural network model belongs to for characterization information with information above-mentioned various types of
Corresponding relationship between probability；Module is obtained, distinguishes difficulty weight for obtaining at least one, wherein distinguish difficulty weight and use
In the difficulty of the characterization two kinds of information of Distinguish centering, type is to by two types in above-mentioned at least three kinds of preset kinds
Type composition；Determining module, it is above-mentioned each for being belonged to based at least one above-mentioned differentiation difficulty weight and above-mentioned first training sample
The probability of seed type determines the model error of initial neural network model；Update module, for being updated according to above-mentioned model error
Initial neural network model.
In the present embodiment, the specific of acquiring unit 601, recognition unit 602 and updating unit 603 handles and its is brought
Technical effect can be respectively with reference to the related description of step 201, step 202 and step 203 in Fig. 2 corresponding embodiment, herein
It repeats no more.
In some optional implementations of the present embodiment, above-mentioned determining module is also used to: according to above-mentioned first training
Sample belongs to above-mentioned various types of probability, determines and belongs to above-mentioned various types of probability for characterizing above-mentioned first training sample
Probable error；It is difficult according to the two kinds of differentiation of the type centering for each type pair of at least one type centering
The sum for spending weight and corresponding probable error, determines the weight probable error of the type pair；It is missed according to identified weight probability
Difference determines above-mentioned model error.
In some optional implementations of the present embodiment, above-mentioned acquisition module is also used to: at least one type
Each type pair of centering, training be directed to the type pair two disaggregated models, wherein two disaggregated model for characterization information with
Indicate the corresponding relationship between information, instruction information is used to indicate one of the two types that information belongs to the type centering；
Using the test sample for being labeled with type in advance, the model error of each two disaggregated model is determined；For each of being trained two
Disaggregated model determines the differentiation difficulty of the targeted type pair of two disaggregated model according to the model error of two disaggregated model
Weight.
In some optional implementations of the present embodiment, above-mentioned acquisition module is also used to: for it is above-mentioned at least one
Each type of each type centering of type centering obtains and is labeled with the type and the second of the probability for belonging to the type in advance
The set of training sample；Probability in response to belonging to the type is less than the second training sample of predetermined probabilities threshold value in the set
Ratio be less than preset ratio threshold value, based on belong to the type probability be less than predetermined probabilities threshold value the second training sample, expand
Increasing obtains amplified sample；Amplified sample and the second training sample in the set are merged, obtain merging sample set；It is right
In each type pair of at least one above-mentioned type centering, the corresponding merging sample of the two types of the type centering is utilized
Set, training are directed to two disaggregated models of the type pair.
In some optional implementations of the present embodiment, above-mentioned acquisition module is also used to: for it is above-mentioned at least one
Each type of each type centering of type centering obtains the second training sample for being labeled with the type in advance；It will mark in advance
Second training sample of note the type imports the generative probabilistic model pre-established, determines second training of mark the type in advance
Sample belongs to the probability of the type, wherein above-mentioned generative probabilistic model is for characterizing the second training sample and the second training sample
The corresponding relationship being belonging respectively between above-mentioned various types of probability.
It should be noted that in the device of information for identification provided in this embodiment each unit realization details and technology
Effect can be with reference to the explanation of other embodiments in the application, and details are not described herein.
Below with reference to Fig. 7, it illustrates the computer systems 700 for the server for being suitable for being used to realize the embodiment of the present application
Structural schematic diagram.Server shown in Fig. 7 is only an example, should not function and use scope band to the embodiment of the present application
Carry out any restrictions.
As shown in fig. 7, computer system 700 includes central processing unit (CPU) 701, it can be read-only according to being stored in
Program in memory (ROM) 702 or be loaded into the program in random access storage device (RAM) 703 from storage section 708 and
Execute various movements appropriate and processing.In RAM 703, also it is stored with system 700 and operates required various programs and data.
CPU 701, ROM 702 and RAM 703 are connected with each other by bus 704.Input/output (I/O) interface 705 is also connected to always
Line 704.
I/O interface 705 is connected to lower component: the importation 706 including keyboard, mouse etc.；It is penetrated including such as cathode
The output par, c 707 of spool (CRT), liquid crystal display (LCD) etc. and loudspeaker etc.；Storage section 708 including hard disk etc.；
And the communications portion 709 of the network interface card including LAN card, modem etc..Communications portion 709 via such as because
The network of spy's net executes communication process.Driver 710 is also connected to I/O interface 705 as needed.Detachable media 711, such as
Disk, CD, magneto-optic disk, semiconductor memory etc. are mounted on as needed on driver 710, in order to read from thereon
Computer program be mounted into storage section 708 as needed.
Particularly, in accordance with an embodiment of the present disclosure, it may be implemented as computer above with reference to the process of flow chart description
Software program.For example, embodiment of the disclosure includes a kind of computer program product comprising be carried on computer-readable medium
On computer program, which includes the program code for method shown in execution flow chart.In such reality
It applies in example, which can be downloaded and installed from network by communications portion 709, and/or from detachable media
711 are mounted.When the computer program is executed by central processing unit (CPU) 701, limited in execution the present processes
Above-mentioned function.
It should be noted that the above-mentioned computer-readable medium of the application can be computer-readable signal media or meter
Calculation machine readable storage medium storing program for executing either the two any combination.Computer readable storage medium for example can be --- but not
Be limited to --- electricity, magnetic, optical, electromagnetic, infrared ray or semiconductor system, device or device, or any above combination.Meter
The more specific example of calculation machine readable storage medium storing program for executing can include but is not limited to: have the electrical connection, just of one or more conducting wires
Taking formula computer disk, hard disk, random access storage device (RAM), read-only memory (ROM), erasable type may be programmed read-only storage
Device (EPROM or flash memory), optical fiber, portable compact disc read-only memory (CD-ROM), light storage device, magnetic memory device,
Or above-mentioned any appropriate combination.In this application, computer readable storage medium can be it is any include or storage journey
The tangible medium of sequence, the program can be commanded execution system, device or device use or in connection.And at this
In application, computer-readable signal media may include in a base band or as carrier wave a part propagate data-signal,
Wherein carry computer-readable program code.The data-signal of this propagation can take various forms, including but unlimited
In electromagnetic signal, optical signal or above-mentioned any appropriate combination.Computer-readable signal media can also be that computer can
Any computer-readable medium other than storage medium is read, which can send, propagates or transmit and be used for
By the use of instruction execution system, device or device or program in connection.Include on computer-readable medium
Program code can transmit with any suitable medium, including but not limited to: wireless, electric wire, optical cable, RF etc. are above-mentioned
Any appropriate combination.
Flow chart and block diagram in attached drawing are illustrated according to the system of the various embodiments of the application, method and computer journey
The architecture, function and operation in the cards of sequence product.In this regard, each box in flowchart or block diagram can generation
A part of one module, program segment or code of table, a part of the module, program segment or code include one or more use
The executable instruction of the logic function as defined in realizing.It should also be noted that in some implementations as replacements, being marked in box
The function of note can also occur in a different order than that indicated in the drawings.For example, two boxes succeedingly indicated are actually
It can be basically executed in parallel, they can also be executed in the opposite order sometimes, and this depends on the function involved.Also it to infuse
Meaning, the combination of each box in block diagram and or flow chart and the box in block diagram and or flow chart can be with holding
The dedicated hardware based system of functions or operations as defined in row is realized, or can use specialized hardware and computer instruction
Combination realize.
Being described in the embodiment of the present application involved unit or module can be realized by way of software, can also be with
It is realized by way of hardware.Described unit or module also can be set in the processor, for example, can be described as:
A kind of processor includes acquiring unit, recognition unit and updating unit, wherein updating unit include generation module, obtain module,
Determining module and update module.Wherein, the title of these units or module is not constituted to the unit itself under certain conditions
Restriction, for example, receiving unit is also described as " obtaining the unit of information to be identified ".
As on the other hand, present invention also provides a kind of computer-readable medium, which be can be
Included in device described in above-described embodiment；It is also possible to individualism, and without in the supplying device.Above-mentioned calculating
Machine readable medium carries one or more program, when said one or multiple programs are executed by the device, so that should
Device: information to be identified is obtained；Above-mentioned information to be identified is imported into neural network model, above-mentioned information to be identified is generated and belongs to respectively
Various types of probability at least three kinds of preset kinds, and according to belonging to the above-mentioned information to be identified of the probability of generation identification
Type, wherein above-mentioned neural network model is obtained according to updated initial neural network model, initial neural network model
It is updated based on following steps: the first training sample is imported into initial neural network model, generated above-mentioned first training sample and belong to
Above-mentioned various types of probability, wherein above-mentioned initial neural network model belongs to for characterization information with information above-mentioned various types of
Corresponding relationship between the probability of type；It obtains at least one and distinguishes difficulty weight, wherein distinguish difficulty weight and distinguished for characterizing
The difficulty of the two kinds of information of type centering, type are formed to by the two types in above-mentioned at least three kinds of preset kinds；Base
Belong to above-mentioned various types of probability at least one above-mentioned differentiation difficulty weight and above-mentioned first training sample, determines initial mind
Model error through network model；Initial neural network model is updated according to above-mentioned model error.
Above description is only the preferred embodiment of the application and the explanation to institute's application technology principle.Those skilled in the art
Member is it should be appreciated that invention scope involved in the application, however it is not limited to technology made of the specific combination of above-mentioned technical characteristic
Scheme, while should also cover in the case where not departing from foregoing invention design, it is carried out by above-mentioned technical characteristic or its equivalent feature
Any combination and the other technical solutions formed.Such as features described above has similar function with (but being not limited to) disclosed herein
Can technical characteristic replaced mutually and the technical solution that is formed.
Claims (10)
1. a kind of method of information for identification, which is characterized in that the described method includes:
Obtain information to be identified；
The information to be identified is imported into neural network model, the information to be identified is generated and is belonging respectively at least three kinds of default classes
Various types of probability in type, and according to the probability of generation identify the information to be identified belonging to type, wherein it is described
Neural network model is obtained according to updated initial neural network model, and initial neural network model is based on following steps more
It is new:
First training sample is imported into initial neural network model, generation first training sample belongs to described various types of
Probability, wherein the initial neural network model belongs between various types of probability for characterization information and information
Corresponding relationship；
It obtains at least one and distinguishes difficulty weight, wherein it is two kinds of for characterizing Distinguish centering to distinguish difficulty weight
The difficulty of information, type are formed to by the two types in above-mentioned at least three kinds of preset kinds；
Belong to various types of probability based at least one described differentiation difficulty weight and first training sample, determines
The model error of initial neural network model；
Initial neural network model is updated according to the model error.
2. the method according to claim 1, wherein described based at least one described differentiation difficulty weight and institute
It states the first training sample and belongs to various types of probability, determine the model error of initial neural network model, comprising:
Belong to various types of probability according to first training sample, determines for characterizing the first training sample category
In the probable error of various types of probability；
For each type pair of at least one type centering, according to the two kinds of differentiation difficulty weight of the type centering and
The sum of corresponding probable error determines the weight probable error of the type pair；
According to identified weight probable error, above-mentioned model error is determined.
3. method according to claim 1 or 2, which is characterized in that described to obtain at least one differentiation difficulty weight, packet
It includes:
For each type pair of at least one type centering, training is directed to two disaggregated models of the type pair, wherein this two points
Class model is for characterization information and indicates the corresponding relationship between information, and instruction information is used to indicate information and belongs to the type centering
One of two types；
Using the test sample for being labeled with type in advance, the model error of each two disaggregated model is determined；
For each of being trained two disaggregated models to determine the two disaggregated models institute according to the model error of two disaggregated model
For type pair differentiation difficulty weight.
4. according to the method described in claim 3, it is characterized in that, each type at least one type centering
Right, training is directed to two disaggregated models of the type pair, comprising:
For each type of each type centering of at least one type centering, obtains and be labeled with the type and category in advance
In the set of the second training sample of the probability of the type；Probability in response to belonging to the type is less than the of predetermined probabilities threshold value
Ratio of two training samples in the set is less than preset ratio threshold value, is less than predetermined probabilities threshold based on the probability for belonging to the type
Second training sample of value, amplification obtain amplified sample；Amplified sample and the second training sample in the set are merged,
It obtains merging sample set；
For each type pair of at least one type centering, the corresponding conjunction of the two types of the type centering is utilized
And sample set, training are directed to two disaggregated models of the type pair.
5. according to the method described in claim 4, it is characterized in that, each class at least one type centering
Each type of type centering obtains the set for being labeled with the type with the second training sample of the probability for belonging to the type in advance,
Include:
For each type of each type centering of at least one type centering, obtains and be labeled with the of the type in advance
Two training samples；Second training sample of preparatory mark the type is imported into the generative probabilistic model pre-established, is determined preparatory
Second training sample of mark the type belongs to the probability of the type, wherein the generative probabilistic model is for characterizing the second instruction
Practice sample and the second training sample is belonging respectively to the corresponding relationship between various types of probability.
6. a kind of device of information for identification, which is characterized in that described device includes:
Acquiring unit, for obtaining information to be identified；
Recognition unit generates the information to be identified and is belonging respectively to for the information to be identified to be imported neural network model
Various types of probability at least three kinds of preset kinds, and belonging to identifying the information to be identified according to the probability of generation
Type, wherein the neural network model is obtained according to updated initial neural network model, and initial neural network model is logical
Updating unit update is crossed, the updating unit includes:
Generation module generates first training sample and belongs to for the first training sample to be imported initial neural network model
Various types of probability, wherein the initial neural network model belongs to for characterization information with information described various types of
Corresponding relationship between the probability of type；
Module is obtained, distinguishes difficulty weight for obtaining at least one, wherein distinguish difficulty weight for characterizing Distinguish pair
In two kinds of information difficulty, type forms to by the two types in above-mentioned at least three kinds of preset kinds；
Determining module, it is described various types of for being belonged to based at least one described differentiation difficulty weight and first training sample
The probability of type determines the model error of initial neural network model；
Update module, for updating initial neural network model according to the model error.
7. device according to claim 6, which is characterized in that the determining module is also used to:
Belong to various types of probability according to first training sample, determines for characterizing the first training sample category
In the probable error of various types of probability；
For each type pair of at least one type centering, according to the two kinds of differentiation difficulty weight of the type centering and
The sum of corresponding probable error determines the weight probable error of the type pair；
According to identified weight probable error, above-mentioned model error is determined.
8. device according to claim 6 or 7, which is characterized in that the acquisition module is also used to:
For each type pair of at least one type centering, training is directed to two disaggregated models of the type pair, wherein this two points
Class model is for characterization information and indicates the corresponding relationship between information, and instruction information is used to indicate information and belongs to the type centering
One of two types；
Using the test sample for being labeled with type in advance, the model error of each two disaggregated model is determined；
For each of being trained two disaggregated models to determine the two disaggregated models institute according to the model error of two disaggregated model
For type pair differentiation difficulty weight.
9. a kind of server, which is characterized in that the server includes:
One or more processors；
Storage device, for storing one or more programs,
When one or more of programs are executed by one or more of processors, so that one or more of processors
Realize such as method as claimed in any one of claims 1 to 5.
10. a kind of computer readable storage medium, is stored thereon with computer program, which is characterized in that the program is by processor
Such as method as claimed in any one of claims 1 to 5 is realized when execution.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710513137.0A CN109214501B (en) | 2017-06-29 | 2017-06-29 | Method and apparatus for identifying information |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710513137.0A CN109214501B (en) | 2017-06-29 | 2017-06-29 | Method and apparatus for identifying information |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109214501A true CN109214501A (en) | 2019-01-15 |
| CN109214501B CN109214501B (en) | 2021-11-30 |
Family
ID=64960452
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710513137.0A Active CN109214501B (en) | 2017-06-29 | 2017-06-29 | Method and apparatus for identifying information |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109214501B (en) |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109785072A (en) * | 2019-01-23 | 2019-05-21 | 北京京东尚科信息技术有限公司 | Method and apparatus for generating information |
| CN110674414A (en) * | 2019-09-20 | 2020-01-10 | 北京字节跳动网络技术有限公司 | Target information identification method, device, equipment and storage medium |
| CN110796366A (en) * | 2019-10-28 | 2020-02-14 | 中国联合网络通信集团有限公司 | Quality difference cell identification method and device |
| CN113093967A (en) * | 2020-01-08 | 2021-07-09 | 富泰华工业（深圳）有限公司 | Data generation method, data generation device, computer device, and storage medium |
| CN113642592A (en) * | 2020-04-27 | 2021-11-12 | 武汉Tcl集团工业研究院有限公司 | Training method of training model, scene recognition method and computer equipment |
| CN115114467A (en) * | 2021-03-17 | 2022-09-27 | 腾讯科技（深圳）有限公司 | Training method and device of picture neural network model |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| KR101591626B1 (en) * | 2015-02-04 | 2016-02-04 | 한양대학교 산학협력단 | Method and apparatus for conceal packet loss based on the deep neural network |
| CN106778820A (en) * | 2016-11-25 | 2017-05-31 | 北京小米移动软件有限公司 | Identification model determines method and device |
| CN106845330A (en) * | 2016-11-17 | 2017-06-13 | 北京品恩科技股份有限公司 | A kind of training method of the two-dimension human face identification model based on depth convolutional neural networks |
-
2017
- 2017-06-29 CN CN201710513137.0A patent/CN109214501B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| KR101591626B1 (en) * | 2015-02-04 | 2016-02-04 | 한양대학교 산학협력단 | Method and apparatus for conceal packet loss based on the deep neural network |
| CN106845330A (en) * | 2016-11-17 | 2017-06-13 | 北京品恩科技股份有限公司 | A kind of training method of the two-dimension human face identification model based on depth convolutional neural networks |
| CN106778820A (en) * | 2016-11-25 | 2017-05-31 | 北京小米移动软件有限公司 | Identification model determines method and device |
Non-Patent Citations (2)
| Title |
|---|
| YAJUN QU ET AL: "Neural-network based regression model with prior from ranking information", 《THE 2010 INTERNATIONAL JOINT CONFERENCE ON NEURAL NETWORKS (IJCNN)》 * |
| 高晨旭: "监控视频中的行人检测技术及其应用", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109785072A (en) * | 2019-01-23 | 2019-05-21 | 北京京东尚科信息技术有限公司 | Method and apparatus for generating information |
| CN110674414A (en) * | 2019-09-20 | 2020-01-10 | 北京字节跳动网络技术有限公司 | Target information identification method, device, equipment and storage medium |
| CN110796366A (en) * | 2019-10-28 | 2020-02-14 | 中国联合网络通信集团有限公司 | Quality difference cell identification method and device |
| CN113093967A (en) * | 2020-01-08 | 2021-07-09 | 富泰华工业（深圳）有限公司 | Data generation method, data generation device, computer device, and storage medium |
| US11527058B2 (en) | 2020-01-08 | 2022-12-13 | Hon Hai Precision Industry Co., Ltd. | Data generating method, and computing device and non-transitory medium implementing same |
| CN113642592A (en) * | 2020-04-27 | 2021-11-12 | 武汉Tcl集团工业研究院有限公司 | Training method of training model, scene recognition method and computer equipment |
| CN115114467A (en) * | 2021-03-17 | 2022-09-27 | 腾讯科技（深圳）有限公司 | Training method and device of picture neural network model |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109214501B (en) | 2021-11-30 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109214501A (en) | The method and apparatus of information for identification | |
| CN108898185A (en) | Method and apparatus for generating image recognition model | |
| CN107578017A (en) | Method and apparatus for generating image | |
| CN108446387A (en) | Method and apparatus for updating face registration library | |
| CN108805091A (en) | Method and apparatus for generating model | |
| CN108960316A (en) | Method and apparatus for generating model | |
| CN108830235A (en) | Method and apparatus for generating information | |
| CN109034069A (en) | Method and apparatus for generating information | |
| CN108388674A (en) | Method and apparatus for pushed information | |
| CN108595628A (en) | Method and apparatus for pushed information | |
| CN109492160A (en) | Method and apparatus for pushed information | |
| CN109086719A (en) | Method and apparatus for output data | |
| CN108345387A (en) | Method and apparatus for output information | |
| CN109086780A (en) | Method and apparatus for detecting electrode piece burr | |
| CN108831505A (en) | The method and apparatus for the usage scenario applied for identification | |
| CN109299477A (en) | Method and apparatus for generating text header | |
| CN109815365A (en) | Method and apparatus for handling video | |
| CN109947989A (en) | Method and apparatus for handling video | |
| CN108897853A (en) | The method and apparatus for generating pushed information | |
| CN109961032A (en) | Method and apparatus for generating disaggregated model | |
| CN108446659A (en) | Method and apparatus for detecting facial image | |
| CN108509921A (en) | Method and apparatus for generating information | |
| CN108510084A (en) | Method and apparatus for generating information | |
| CN108509994A (en) | character image clustering method and device | |
| CN109389182A (en) | Method and apparatus for generating information |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |