CN110110845A - Learning method based on parallel multi-level width neural network - Google Patents
Learning method based on parallel multi-level width neural network Download PDFInfo
- Publication number
- CN110110845A CN110110845A CN201910331708.8A CN201910331708A CN110110845A CN 110110845 A CN110110845 A CN 110110845A CN 201910331708 A CN201910331708 A CN 201910331708A CN 110110845 A CN110110845 A CN 110110845A
- Authority
- CN
- China
- Prior art keywords
- neural network
- level width
- test
- width neural
- sample
- 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
- 238000013528 artificial neural network Methods 0.000 title claims abstract description 201
- 238000000034 method Methods 0.000 title claims abstract description 43
- 238000012360 testing method Methods 0.000 claims abstract description 96
- 238000012549 training Methods 0.000 claims abstract description 75
- 230000006870 function Effects 0.000 claims description 43
- 239000011159 matrix material Substances 0.000 claims description 37
- 230000009466 transformation Effects 0.000 claims description 5
- 239000012141 concentrate Substances 0.000 claims description 4
- 238000013501 data transformation Methods 0.000 claims description 4
- PXFBZOLANLWPMH-UHFFFAOYSA-N 16-Epiaffinine Natural products C1C(C2=CC=CC=C2N2)=C2C(=O)CC2C(=CC)CN(C)C1C2CO PXFBZOLANLWPMH-UHFFFAOYSA-N 0.000 claims description 3
- 238000010606 normalization Methods 0.000 claims 1
- 238000012795 verification Methods 0.000 abstract description 6
- 238000004364 calculation method Methods 0.000 abstract 1
- 230000010354 integration Effects 0.000 abstract 1
- 230000008569 process Effects 0.000 description 15
- 238000005070 sampling Methods 0.000 description 6
- 238000013527 convolutional neural network Methods 0.000 description 5
- 238000010586 diagram Methods 0.000 description 5
- 210000004218 nerve net Anatomy 0.000 description 4
- 238000007637 random forest analysis Methods 0.000 description 4
- 238000012706 support-vector machine Methods 0.000 description 4
- 238000013136 deep learning model Methods 0.000 description 3
- 238000012986 modification Methods 0.000 description 3
- 230000004048 modification Effects 0.000 description 3
- 230000008901 benefit Effects 0.000 description 2
- 238000013480 data collection Methods 0.000 description 2
- 238000013135 deep learning Methods 0.000 description 2
- 239000000203 mixture Substances 0.000 description 2
- 210000005036 nerve Anatomy 0.000 description 2
- 230000001537 neural effect Effects 0.000 description 2
- 238000005457 optimization Methods 0.000 description 2
- 230000002159 abnormal effect Effects 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 230000000052 comparative effect Effects 0.000 description 1
- 230000006835 compression Effects 0.000 description 1
- 238000007906 compression Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 230000004927 fusion Effects 0.000 description 1
- 210000004209 hair Anatomy 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
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
- G06N3/045—Combinations of networks
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Computing Systems (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Molecular Biology (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Image Analysis (AREA)
- Investigating Or Analysing Biological Materials (AREA)
Abstract
The invention discloses a learning method based on a parallel multi-level width neural network, which comprises the following steps: obtaining a verification set and constructing a base classifier; training and verifying each level of the parallel M-level width neural network to obtain a trained parallel M-level width neural network and verification output corresponding to each level of the width neural network; obtaining a decision threshold value of each level of width neural network through statistical calculation; and testing the verified parallel multi-level width neural network through a test set. The neural network has a multi-stage structure, each stage learns different parts of data, and parallel training and testing can be realized. Each level adopts a width neural network to carry out feature learning in the width direction; realizing the integration of the classifiers in two width directions by using a plurality of width neural networks as the reconnection of the base classifiers in the width directions; incremental learning of the network is realized by adding a new level of width neural network; and can realize parallelization test.
Description
Technical field
The invention belongs to artificial intelligence and machine learning techniques fields, and in particular to one kind is based on parallel multi-level width mind
Learning method through network.
Background technique
As the learning model based on deep learning network obtains in the fields such as large-scale image processing and machine vision
The complexity of immense success, learning model also quicklys increase, they need a large amount of high dimensional data to go to train, to increase
Add required computing resource and calculates the time.In addition, actual data are frequently not homogeneity, some samples are very easy to
Classification, but also have many sample classifications relatively difficult.Most of errors in classification occur when input sample is difficult classification
Wait, for example, sample uneven distribution, abnormal obtain sample and close to classification boundaries or the sample of linearly inseparable etc..
In existing deep learning model, simple sample and complex samples are handled in a like fashion, reduce meter
Calculate the service efficiency of resource.Meanwhile existing deep learning network such as convolutional neural networks often have plurality of layers, all samples
Originally it will pass through all network layers, it can be very time-consuming when carrying out extensive or test to network.And early stage is parallel
Organizing network only receives the sample by nonlinear transformation refused by upper level in every level-one, these samples are become
The other spaces for being easy to classify are changed to, to classify again.But how to realize high dimensional data for different difficulty
Data sample carry out adjustment and the distribution of computing resource with improve the speed of learning classification and efficiency this problem there is no
To very good solution.
Summary of the invention
In view of the foregoing drawbacks, the present invention provides a kind of learning method based on parallel multi-level width neural network, this hairs
Bright neural network has multilevel structure, and every level-one is learnt for the different piece in data, and can realize that parallelization is instructed
Practice and tests.Every level-one carries out feature learning in width direction using a kind of width neural network；Pass through multiple width nerve nets
Network is again coupled to as base classifier in width direction, realizes the combining classifiers in two width directions；It is new by increasing
The incremental learning of the width neural fusion network of level-one；And can realize that parallelization is tested, substantially reduce complex samples
The learning classification time improves network operation efficiency.
In order to achieve the above object, the present invention is resolved using following technical scheme.
(2) a kind of learning method based on parallel multi-level width neural network, parallel multi-level width neural network include
Multistage width neural network, wherein every level width neural network includes sequentially connected input layer, hidden layer, decision-making level and defeated
Layer out, whether the decision-making level is for determining each test sample by when prime exports, the learning method includes following step
It is rapid:
Step 1, original training sample collection is obtained, parallel M level width neural network Net is constructed1... Netm..., NetM(m
=1,2 ..., M), base classifier of every level width neural network as respective stages；By being carried out M times to original training sample collection
Data transformation, it is corresponding to obtain M verifying collection xv_1... xv_m... xv_M；
Wherein, the total sample number that original training sample integrates is Ntr。
Step 2, x is collected using original training sample collection and M verifyingv_1... xv_m... xv_MRespectively to parallel M level width
Every grade of neural network is trained and verifies, parallel M level width neural network and every level width nerve net after being trained
The corresponding verifying of network exports yv_m(m=1,2 ..., M)；Each verifying output y is obtained using minimum error methodv_mCorresponding label
yv_ind_m, and then correct point of the verifying collection of every level width neural network of the parallel M level width neural network after being trained
Class sample set yvc_mWith wrong classified sample set yvw_m；
Step 3, to the correct of the verifying collection of every level width neural network of the parallel M level width neural network after training
Classified sample set yvc_mWith wrong classified sample set yvw_mStatistics calculating is carried out respectively, corresponds to every level width mind after being trained
Decision-making value T through networkm；By the decision-making value T of every level width neural networkmDecision as corresponding level width neural network
Foundation obtains the parallel M level width neural network that decision-making value determines；
Step 4, test set is obtained, the input for the parallel M level width neural network that test set is determined as decision-making value
Data parallel inputs to every level width neural network that decision-making value determines and is tested, and obtains every grade that decision-making value determines
The output of width neural network；The error vector for obtaining every level width neural network, every level width mind that decision-making value is determined
Output through network is judged, so that the test output for obtaining every level width neural network that decision-making value determines is corresponding
Label ytest_ind_m。
The characteristics of technical solution of the present invention and further improvement are as follows:
(1) in step 1, the data be transformed to the sample that original sample is concentrated by elastic registration (Elastic) into
Row compression or deformation；Or the data are transformed to revolve by the sample that affine transformation (Affine) concentrates original sample
Turn, overturn, zoom in or out.
(2) described that x is collected using original training sample collection and M verifying in step 2v_1... xv_m... xv_MRespectively to parallel
Every grade of M level width neural network is trained and verifies, and it includes following sub-steps:
Sub-step 2.1, using original training sample collection as the 1st level width neural network Net1Input sample, to the 1st
Level width neural network Net1It is trained, the first level width neural network after being trained.
Sub-step 2.2 collects x using the first verifyingv_1The 1st level width neural network after training is verified, is obtained
The wrong classified sample set y of the verifying collection of 1st level width neural networkvw_1。
Sub-step 2.3, by the wrong classified sample set y of the first level width neural networkvw_1As the 2nd level width nerve
The input sample A of networkv_1；Training sample set A is randomly selected from original training sample concentration againv_2, make total input sample collection
{Av_1+Av_2In sample number be equal to the sample number that original training sample is concentrated, and by total input sample collection { Av_1+Av_2Make
For the input sample of the 2nd level width neural network.
Sub-step 2.4, using total input sample collection { Av_1+Av_22nd level width neural network is trained, it obtains
The 2nd level width neural network after training；X is collected using the second verifyingv_2The 2nd level width neural network after training is tested
Card obtains the wrong classified sample set y of the verifying collection of the 2nd level width neural networkvw_2。
And so on, 3rd level is trained respectively to M level width neural network, parallel M grades after being trained
The corresponding verifying output y of width neural network and every level width neural networkv_m(m=1,2 ..., M).
(3) in step 2, the minimum error method are as follows:
Firstly, setting total classification number that original training sample integrates as C, R-matrix R is constructedj(1≤j≤C)。
Wherein, R-matrix RjThe element of jth row be all 1, remaining element is all 0, each R-matrix RjDimension
For C × Ntr。
Secondly, exporting y according to the verifying of every level width neural network after trainingv_m, obtain verifying output yv_mWith it is corresponding
The R-matrix R of gradejBetween error vector:
Jv_mj=| | softmax (yv_m)-Rj||2, 1≤j≤C；
Wherein, Jv_mjDimension be 1 × Ntr；yv_mDimension be C × Ntr。
Finally, exporting y to verifyingv_mWith the R-matrix R of respective stagesjBetween error vector Jv_mjIt minimizes, obtains
The corresponding class label y of every level width neural network after trainingv_ind_m:
Wherein, yv_ind_mDimension be 1 × Ntr。
(4) in step 3, it includes following sub-step that the statistics, which calculates:
Sub-step 3.1, correct point of the m level width neural network of the parallel M level width neural network after setting training
Class sample set and wrong classified sample set are respectively as follows: yvc_mAnd yvw_m, what correct classified sample set and wrong classification samples were concentrated
Total sample number is respectively as follows: Nvc_mAnd Nvw_m, and Nvc_m+Nvw_m=Ntr, then correct classified sample set and wrong classified sample set
Error is respectively as follows:
evc_m=| | softmax (yvc_m)-tvc_m||2；
evw_m=| | softmax (yvw_m)-tvw_m||2；
Wherein, tvc_mIt is correct classification samples y in m level width neural networkvc_mCorresponding true tag, tvw_mIt is m grades
Wrong classification samples y in width neural networkvw_mCorresponding true tag.
Sub-step 3.2, according to correct classified sample set yvc_mWith wrong classified sample set yvw_m, calculate separately out correct point
Class sample set yvc_mMean value and variance be respectively μcAnd σc；Mistake classified sample set yvw_mMean value and variance be respectively: uwWith
σw；Then correct classified sample set yvc_mWith wrong classified sample set yvw_mCorresponding Gaussian Profile is respectively:
Correct classified sample set yvc_mWith wrong classified sample set yvw_mCorresponding Gaussian probability-density function is respectively:
Sub-step 3.3, according to wrong classified sample set yvw_mError evw_mAnd variances sigmaw, obtain m level width nerve net
The decision-making value T of networkm=min (evw_m)-ασw。
Wherein, α is a constant, for providing allowance, so that the wrong classification samples y of institutevw_mIt is refused when prime
Absolutely.
(5) in step 4, the acquisition test set are as follows: obtain original test sample collection xtest；By M data extending,
It is corresponding to obtain M group test sample collection xtest_1..., xtest_m..., xtest_M, as test set.
Further, the data extending are as follows: to the original test sample collection xtestIn each sample carry out respectively
NtestDThe secondary data transformation, correspondence obtain NtestDA test sample collection, the parallel M level width mind determined as decision-making value
The test set x of m level width neural network through networktest_m。
Wherein, original test sample collection XtestMiddle test sample sum is Ntest_saples。
(6) in step 4, the error vector for obtaining every level width neural network includes following sub-step:
Sub-step 4.1, by M group test sample collection xtest_1, xtest_2..., xtest_MInput to decision-making value parallel respectively
Determining parallel M level width neural network, the corresponding N for obtaining every level width neural network that decision-making value determinestestDA output
ytest_m_d, (d=1,2 ... NtestD)。
Sub-step 4.2, to the N for every level width neural network that decision-making value determinestestDA output ytest_m_d, (d=1,2 ...
NtestD) average value is calculated, obtain the test output for every level width neural network that decision-making value determines
Sub-step 4.3 sets total classification number of test set as C, constructs R-matrix Rj(1≤j≤C)；It is defeated to obtain verifying
Y outv_mWith the R-matrix R of respective stagesjBetween error vector:
Jtest_mj=| | softmax (ytest_m)-Rj||2, 1≤j≤C；
Wherein, R-matrix RjThe element of jth row be all 1, remaining element is all 0, each R-matrix RjDimension
For C × Ntest_samples；Jtest_mjDimension be 1 × Ntest_samples, yv_mDimension be C × Ntest_samples。
(7) output of the every level width neural network determined to decision-making value judges are as follows:
The minimal error of current level width neural network is less than or equal to when prime decision-making value, then is judged as and works as prime
For the correct classification output stage of the output:
min(Jtest_mj)≤Tm。
The minimal error of current level width neural network is greater than when prime decision-making value, then is judged as when prime can not
Correctly classified to the output, which is transferred to next level width neural network and is tested, is so recycled, until this
Correct classification output stage is found in output:
min(Jtest_mj) > Tm。
(8) in step 4, the test for obtaining every level width neural network that decision-making value determines exports corresponding mark
Sign ytest_ind_mAre as follows:
Wherein, ytest_ind_mDimension be 1 × Ntest_samples。
Compared with prior art, the invention has the benefit that
(1) neural network of the invention has multistage base classifier, and every level-one is used to the different piece sample of learning data set
This, can adaptively determine the structure of neural network, realize computing resource according to problem and the complexity of data set
Optimization.
(2) neural network of the invention has the advantages that incremental learning, when new training data is available, realizes to working as
The judgement of preceding neural network, according to judging result, it is determined whether can correctly be classified to newly-increased training data, if cannot be into
Row properly separates, then learns new sample by increasing the new width radial basis function level-one new as neural network, and
Without re -training whole network.
(3) neural network of the invention can carry out concurrent testing when test, that is, simultaneously test data
To all grades of network, the decision-making value of every level-one obtained in training process determines each test sample finally by which
The neural network of level-one exports, and concurrent testing process greatly reduces waiting time when actual use network.
(4) neural network of the invention can be used as a kind of general learning framework, have very strong flexibility, each
Grade can use BP neural network, convolutional neural networks or other kinds of classifier according to actual needs.
Detailed description of the invention
The present invention is described in further details in the following with reference to the drawings and specific embodiments.
Fig. 1 is the schematic diagram and its training test process schematic diagram of parallel multi-level neural network of the invention；Wherein, Fig. 1
It (a) is parallel multi-level width neural networks principles figure of the invention；Fig. 1 (b) is parallel multi-level width neural network of the invention
Training and verification process schematic diagram；Fig. 1 (c) is the test process schematic diagram of parallel multi-level width neural network of the invention.
Fig. 2 is the structure chart of parallel multi-level width neural network of the invention.
Fig. 3 (a) is error distribution of the verifying collection of parallel multi-level width neural network of the invention in wherein level-one
Figure；Fig. 3 (b) is the Gaussian probability-density function of the statistical parameter in Fig. 3 (a).
Fig. 4 is test result of the parallel 26 level width neural network on MMNIST data set in the embodiment of the present invention
With the classification results comparison diagram of existing learning model.
Specific embodiment
Embodiment of the present invention is described in detail below in conjunction with embodiment, but those skilled in the art
It will be understood that following embodiment is merely to illustrate the present invention, and it is not construed as limiting the scope of the invention.
Using MNIST hand-written data collection, which is the image of 8 gray scale handwritten numerals 0~9, image
Size is 28 × 28, and 10 class, there is 60000 original training sample collection in total, and 10000 images are newly to learn as test set
One of the important general purpose image data collection of model training test.For the data set, with reference to Fig. 1 and Fig. 2, the present embodiment is used
Width radial primary function network is all made of width radial direction base as base classifier, i.e. every grade of parallel multi-level width neural network
Function Network, the series for choosing parallel width neural network is 26.
(1) verifying collection is obtained, base classifier is constructed.
Firstly, to NtrThe image pattern that=60000 original training samples are concentrated carries out 26 hypoelasticity transformation respectively, obtains
Collect x to M=26 verifyingv_1, xv_2..., xv_26, guarantee that the mistake for the verifying collection for having enough divides sample in the present embodiment
This, it includes N that each verifying, which is concentrated,val=10 data sets obtained by original trained set transformation.Wherein, the sample of each verifying collection
This number is the N of original training sample collectionval=10 times.
Secondly, designing parallel multi-level width neural network as base classifier using width radial primary function network；M
=26 width radial primary function networks link together, and form parallel multi-level width neural network Net1, Net2... NetM；
Each base classifier is as level-one, the different piece be absorbed in data set.
Finally, building width radial primary function network.Detailed process is as follows:
Building includes N0k=1000 Gaussian bases areRadial primary function network, should
The center of radial primary function network is a subset for being derived from original training sample collection at random, and standard deviation value is constant.Using
Sliding window obtains original training sample and concentrates the multiple groups local feature image of each image pattern, to obtain multiple groups part
Eigenmatrix obtains multiple radial primary function networks i.e. using multiple groups local feature matrix as the input data of Gaussian bases
For width radial primary function network.
2) every grade of parallel M level width neural network is trained and is verified, the parallel M level width after being trained
Neural network and the corresponding verifying of every level width neural network export yv_m(m=1,2 ..., M).
1st level width radial primary function network is trained using original training sample collection, after training, the instruction of mistake point
Practice the width radial primary function network that sample is sent to the 2nd grade, as a part of second training set, the net that the 2nd grade of Lai Xunlian
Network.The verifying collection obtained using step (1) to the training network verification when prime, while providing more error samples, makees
For a part of next stage training set.Include following sub-step specifically as shown in Fig. 1 (a), (b):
Sub-step 2.1, using original training sample collection as the 1st level width neural network Net1Input sample, to the 1st
Level width neural network Net1It is trained, the 1st level width neural network after being trained.
Sub-step 2.2 collects x using the first verifyingv_1The 1st level width neural network after training is verified, is obtained
The wrong classified sample set y of 1st level width neural networkvw_1。
Sub-step 2.3, by the wrong classified sample set y of the 1st level width neural networkvw_1As the 2nd level width nerve net
The input sample A of networkv_1；Training sample set A is randomly selected from original training sample concentration againv_2, make total input sample collection { Av_1
+Av_2In sample number be equal to the sample number that original training sample is concentrated, and by total input sample collection { Av_1+Av_2It is used as the 2nd
The input sample of level width neural network.
Sub-step 2.4, using total input sample collection { Av_1+Av_22nd level width neural network is trained, it obtains
The 2nd level width neural network after training；X is collected using the second verifyingv_2The 2nd level width neural network after training is tested
Card, obtains the wrong classified sample set y of the 2nd level width neural networkvw_2。
Sub-step 2.3 and 2.4 is repeated, 3rd level is trained respectively to M level width neural network, after being trained
Parallel M level width neural network and every level width neural network corresponding verifying output yv_m(m=1,2 ..., M),.
The specific training of above-mentioned width radial primary function network and verification process are as follows:
For the image pattern that original training sample is concentrated as input data, image size is M1×M2=28 × 28.It is sliding
Dynamic window size is r=13 × 13, and the initial position of sliding window is located at the upper left corner of each image pattern, selection sliding step
A length of 1 pixel, sliding window from left to right, successively slide from top to bottom, 60000 image patterns in sliding window
3 d image block stretch become matrix xk∈Rr×N, i.e., it is corresponding original each local feature image to be pressed into pixel composition respectively
Matrix forms a column vector after arranging the 2nd of each original matrix the to last Leie time sequence arrangement to the 1st；It will be N number of
Column vector sequence rearranges the local feature matrix x of one group of training image samplek(1≤k≤K), local feature matrix xk's
Each column represent a sample.Again local feature matrix xkIt inputs to including N0k=1000 Gaussian bases areRadial primary function network, output is denoted as:
Wherein,For the column vector comprising N=60000 element.
Sliding window slides a corresponding radial primary function network every time can be obtained K=(M after final sliding1-m
+1)(M2- m+1)=(28-13+1) × (28-13+1)=256 radial primary function network.
For each radial primary function network, sequence and down-sampling are introduced by the output of Gaussian bases to it.Needle
To each radial primary function network, to its Gaussian bases output data Φ Jing Guo nonlinear transformationkIntroduce sequence and under
Sampling.To the output data Φ of width radial primary function networkkEach column sum, obtain a row vector, row vector
Each element be each image to be processed local specific position the sum of pixel, it is special to the part of each image to be processed
It positions the sum of pixel set and carries out descending arrangement, obtain descending vectorUsing index skBy descending to
Measure akIn the corresponding home position of local specific position of each image to be processed be marked, the output data to be sorted
Φ′k=sort (Φk, sk)。
Down-sampling is carried out to the output data of sequence, sets down-sampling interval NkS=20, the output number through over-sampling
Are as follows:
Then the output number of total width radial primary function network isSampling output is Φks
=subsample (Φ 'k, NkS), then the output of Gaussian bases is Φ=[Φ1S, Φ2S..., ΦKS]。
Setting desired output is D=[D1, D2..., DC]；To the defeated of the Gaussian bases of width radial primary function network
Linear layer connection is carried out out, then the weight of linear layer are as follows: W=[W1, W2..., WC]；
Wherein, C=10 is the classification sum of original sample.
Obtain the classification output Y=[Y of width radial primary function network1, Y2..., YC]=Φ W；Specifically, pass through minimum
Change the least mean-square estimate that square error calculates the weight of linear layerSpecific formula are as follows:
The weight of linear layer is calculated by the pseudo inverse matrix that the Gaussian bases of width radial primary function network export Φ
Least mean-square estimate
Wherein, Φ+The pseudo inverse matrix of Φ is exported for the Gaussian bases of width radial primary function network.
Finally, the classification output of width radial primary function network is calculated are as follows:
And then the width radial primary function network after being trained, to the width radial primary function network after every grade of training
It is verified using corresponding verifying collection, the corresponding verifying of every level width radial primary function network after being trained exports yv_m(m
=1,2 ..., M).
Y is exported by the verifying of acquisitionv_m(m=1,2 ..., M) further obtains each verifying output yv_mCorresponding class
Distinguishing label yv_ind_m, the specific steps are as follows:
Firstly, setting total classification number that original training sample integrates as C, R-matrix R is constructedj(1≤j≤C)。
Wherein, R-matrix RjThe element of jth row be all 1, remaining element is all 0, each R-matrix RjDimension
For C × Ntr。
Secondly, exporting y according to the verifying of every level width neural network after trainingv_m, obtain verifying output yv_mWith it is corresponding
The R-matrix R of gradejBetween error vector:
Jv_mj=| | softmax (yv_m)-Rj||2, 1≤j≤C；
Wherein, Jv_mjDimension be 1 × Ntr；yv_mDimension be C × Ntr。
Finally, exporting y to verifyingv_mWith the R-matrix R of respective stagesjBetween error vector Jv_mjIt minimizes, obtains
The corresponding class label y of every level width neural network after trainingv_ind_m:
Wherein, yv_ind_mDimension be 1 × Ntr。
By the corresponding class label y of every level width neural network after trainingv_ind_mVerifying with every grade exports yv_mInto
Row compares, and can be obtained the correct classified sample set y of every level width neural networkvc_mWith wrong classified sample set yvw_m。
(3) pass through the decision-making value T that every level width neural network is calculatedm
The relatively difficult part of present networks is the determination of every level-one decision-making value, it is used to determine when test, often
One sample should be exported by which rank of network.After training and test, classify to correct classified sample set and mistake
Sample set carries out statistics calculating respectively.Assuming that in m level width neural network, correct classified sample set and wrong classified sample set
It is respectively as follows: yvc_mAnd yvw_m, the total sample number of correct classified sample set and wrong classified sample set is respectively as follows: Nvc_mAnd Nvw_m,
And Nvc_m+Nvw_m=Ntr。
In the above verification process, in order to guarantee that final sample has enough error sample collection, each verifying collection be can be
It include NvalA that original training sample collection is passed through to the verifying sample set that data convert, i.e., each verifying collection may include
NvalGroup verifying sample set, that is to say, that each sample number integrated of verifying is the N of original training samplevalTimes.
The error of two class sample sets is calculate by the following formula:
evc_m=| | softmax (yvc_m)-tvc_m||2；
evw_m=| | softmax (yvw_m)-tvw_m||2；
Wherein, tvc_mAnd tvw_mIt is correct classification samples y in m gradesvc_mWith wrong classification samples yvw_mCorresponding true mark
Label.Assuming that correct classification and mistake are classified, the mean value of these two types of sample statistics and variance are respectively: μc, uw, σc, σw, right therewith
Two Gaussian Profiles answered are respectively:
Its Gaussian probability-density function is respectively:
In the level-one of parallel multi-level width neural network, verifying collection error distribution and its probability density function such as Fig. 3
(a) and (b) is shown, then the decision-making value of m level width neural network are as follows:
Tm=min (evw_m)-ασw；
Wherein, α is a constant, for providing allowance, so that the wrong classification samples y of institutevw_mIt is refused when prime
Absolutely.
(4) it is tested by the parallel multi-level width neural network that test set determines decision-making value
As shown in Fig. 1 (c), specific test process are as follows:
Firstly, obtaining test set, detailed process are as follows: obtain original test sample collection Xtest；It is right by M data extending
M group test sample collection x should be obtainedtest_1..., xtest_m..., xtest_M, as test set；Wherein, original test sample collection
xtestMiddle test sample sum is Ntest_samples。
Above-mentioned data extending are as follows: to the original test sample collection XtestIn each sample carry out N respectivelytestDIt is secondary
Data transformation, correspondence obtain NtestDA test sample collection, the m as the parallel M level width neural network that decision-making value determines
The test set x of level width neural networktest_m。
The stability that above-mentioned test set acquisition methods can be tested in subsequent test process.
Secondly, by M group test sample collection xtest_1..., xtest_m..., xtest_MDecision-making value is inputed to parallel to determine
Parallel M level width neural network, test set is tested, i.e., is determined the corresponding decision-making value that inputs to of every group of test set
Every level width neural network tested, the corresponding N for obtaining every level width neural network that decision-making value determinestestDA survey
Try sample set output；To NtestDThe output of a test sample collection is averaged, and obtains every level width nerve that decision-making value determines
The test of network exports
Again, total classification number of test set is set as C, constructs R-matrix Rj(1≤j≤C)；Obtain verifying output yv_m
With the R-matrix R of respective stagesjBetween error vector:
Jtest_mj=| | softmax (ytest_m)-Rj||2, 1≤j≤C；
Wherein, R-matrix RjThe element of jth row be all 1, remaining element is all 0, each R-matrix RjDimension
For C × Ntest_samples；Jtest_mjDimension be 1 × Ntest_samples, yv_mDimension be C × Ntest_samples。
Finally, judging the output for every level width neural network that decision-making value determines, specifically: when prime is wide
The minimal error for spending neural network is less than or equal to the i.e. min (J when prime decision-making valuetest_mj)≤Tm, then it is judged as current
Grade is the correct classification output stage of the output.
The minimal error of current level width neural network is greater than the i.e. min (J when prime decision-making valuetest_mj) > Tm, then
It is judged as that working as prime can not correctly classify to the output, be transferred to next level width neural network for the output and test,
So circulation, until correct classification output stage is found in the output.And then obtain every level width neural network that decision-making value determines
Test export corresponding labelWherein, ytest_ind_mDimension be 1 ×
Ntest_samples。
If test sample cannot export for 25 grades in front, directly exported at last the 26th grade.
Test set finally can be obtained in the output L of whole networktest；Wherein, correct classification samples and wrong classification samples
Calculating, and then the precision of the sample classification of available parallel multi-level width neural network of the present invention can be counted.
Comparative example
Using original training sample collection same as the previously described embodiments, verifying collection and test set, random forest is respectively adopted
(RF), multilayer perceptron (MP), conventional radial primary function network (RBF), support vector machines (SVM), range learning system
(BLS), condition deep learning model (CDL), deepness belief network (DBL), convolutional neural networks LeNet-5, depth Bohr is hereby
Deep woods (gc) is used as base classifier at random for graceful machine (DBM) and depth, carries out learning classification, finally obtained various study sides
Method is as shown in Figure 4 to the precision of data classification.
From fig. 4, it can be seen that compared to the learning model of current mainstream: random forest (RF), multilayer perceptron (MP),
Conventional radial primary function network (RBF), support vector machines (SVM), range learning system (BLS), condition deep learning model
(CDL), deepness belief network (DBL), convolutional neural networks LeNet-5, depth Boltzmann machine (DBM) and depth are random
Deep woods (gc forest), the accuracy rate of the classification results of parallel multi-level width neural network (PMWNN) of the invention have non-
Often high competitiveness, the final nicety of grading of the method for the present invention is that 99.10%, WRBF is width radial primary function network.And
Compared to depth, deep woods learning model, the method for the present invention neural network have multistage base neural net at random, and every level-one is used to learn
The different piece sample for practising data set, can adaptively determine neural network according to problem and the complexity of data set
Structure realizes the optimization of computing resource；Meanwhile neural network of the invention can carry out concurrent testing when test,
Exactly test data is given simultaneously all grades of network, the decision-making value of every level-one obtained in training process is every to determine
A test sample is finally exported by which rank of neural network, when concurrent testing process greatly reduces actual use network
Waiting time.
In addition, incremental learning may be implemented in parallel multi-level width neural network of the invention, i.e., when there is new data
It waits, new width radial primary function network can be increased to learn new characteristic, it is wide without the entire parallel multi-level of re -training
Neural network is spent, this means that the network proposed can learn new knowledge under the premise of not forgeing old knowledge.New instruction
Practice data and inputs to current M grades of network, if wrong point of sample, they and the original training set Jing Guo data extending
New training dataset, the new width radial primary function network of training are established together, while being tested using new verifying collection
Card, and decision-making value is calculated, to establish M+1 grades of networks.Finally, new parallel multi-level width neural network will be by M+1
Level width radial primary function network composition.Meanwhile the parallel multi-level width neural network that the present invention designs can when test
With concurrent testing, all test samples can all give all grades of width radial primary function network, which decision-making value determines
One width radial primary function network distributes to corresponding test sample.The process withouts waiting for other grades of network output,
To parallelization when test, test process is accelerated.
Each level width neural network in parallel multi-level width neural network of the invention, can be width radial direction base
Function Network, BP neural network, convolutional neural networks or other classifiers, and every grade of base point of multistage width neural network
The type of class device can be different.
Obviously, various changes and modifications can be made to the invention without departing from of the invention by those skilled in the art
Spirit and scope.In this way, if these modification and variations of the invention belong to the model of the claims in the present invention and its equivalent technologies
Within enclosing, then the present invention is also intended to include these modifications and variations.
Claims (10)
1. a kind of learning method based on parallel multi-level width neural network, parallel multi-level width neural network includes multistage width
Neural network, wherein every level width neural network includes sequentially connected input layer, hidden layer and output layer, which is characterized in that
The learning method the following steps are included:
Step 1, original training sample collection is obtained, parallel M level width neural network Net is constructed1... Netm..., NetM(m=1,
2 ..., M), base classifier of every level width neural network as respective stages；Become by carrying out M data to original training sample collection
It changes, it is corresponding to obtain M verifying collection xv_1... xv_m... xv_M:
Wherein, the total sample number that original training sample integrates is Ntr；
Step 2, x is collected using original training sample collection and M verifyingv_1... xv_m, xv_MRespectively to parallel M level width neural network
Every grade be trained and verify, the parallel M level width neural network and every level width neural network after being trained are corresponding
Verifying output yv_m(m=1,2 ..., M)；Each verifying output y is obtained using minimum error methodv_mCorresponding label yv_ind_m, into
And the correct classified sample set of the verifying collection of every level width neural network of the parallel M level width neural network after being trained
yvc_mWith wrong classified sample set yvw_m；
Step 3, to the correct classification sample of the verifying collection of every level width neural network of the parallel M level width neural network after training
This collection yvc_mWith wrong classified sample set yvw_mStatistics calculating is carried out respectively, corresponds to every level width neural network after being trained
Decision-making value Tm；By the decision-making value T of every level width neural networkmAs the decision-making foundation of corresponding level width neural network, obtain
The parallel M level width neural network determined to decision-making value；
Step 4, test set is obtained, the input data for the parallel M level width neural network that test set is determined as decision-making value,
Every level width neural network that decision-making value determines is inputed to parallel to be tested, and every level width mind that decision-making value determines is obtained
Output through network；The error vector for obtaining every level width neural network, every level width neural network that decision-making value is determined
Output judged, so that the test for obtaining every level width neural network that decision-making value determines exports corresponding label
ytest_ind_m。
2. the learning method according to claim 1 based on parallel multi-level width neural network, which is characterized in that step 1
In, the data are transformed to be compressed or deformed by the sample that elastic registration concentrates original sample；Or the data become
It is changed to and is rotated, overturn by the sample that affine transformation concentrates original sample, zoomed in or out.
3. the learning method according to claim 1 based on parallel multi-level width neural network, which is characterized in that step 2
In, it is described that x is collected using original training sample collection and M verifyingv_1... xv_m... xv_MRespectively to parallel M level width neural network
Every grade be trained and verify, it includes following sub-steps:
Sub-step 2.1, using original training sample collection as the 1st level width neural network Net1Input sample, to the 1st level width
Neural network Net1It is trained, the first level width neural network after being trained；
Sub-step 2.2 collects x using the first verifyingv_1The 1st level width neural network after training is verified, obtains the 1st grade
The wrong classified sample set y of the verifying collection of width neural networkvw_1；
Sub-step 2.3, by the wrong classified sample set y of the first level width neural networkvw_1As the 2nd level width neural network
Input sample Av_1；Training sample set A is randomly selected from original training sample concentration againv_2, make total input sample collection { Av_1+Av_2}
In sample number be equal to the sample number that original training sample is concentrated, and by total input sample collection { Av_1+Av_2It is used as the 2nd level width
The input sample of neural network；
Sub-step 2.4, using total input sample collection { Av_1+Av_22nd level width neural network is trained, after being trained
The 2nd level width neural network；X is collected using the second verifyingv_2The 2nd level width neural network after training is verified, is obtained
The wrong classified sample set y of the verifying collection of 2nd level width neural networkvw_2；
And so on, 3rd level is trained respectively to M level width neural network, the parallel M level width mind after being trained
Corresponding verifying output y through network and every level width neural networkv_m(m=1,2 ..., M).
4. the learning method according to claim 1 based on parallel multi-level width neural network, which is characterized in that step 2
In, the minimum error method are as follows:
Firstly, setting total classification number that original training sample integrates as C, R-matrix R is constructedj(1≤j≤C)。
Wherein, R-matrix RjThe element of jth row be all 1, remaining element is all 0, each R-matrix RjDimension be C ×
Ntr；
Secondly, exporting y according to the verifying of every level width neural network after trainingv_m, obtain verifying output yv_mWith respective stages
R-matrix RjBetween error vector:
Jv_mj=| | softmax (yv_m)-Rj||2, 1≤j≤C；
Wherein, | | | |22 norms of representing matrix, softmax () are normalization exponential function；Jv_mjDimension be 1 × Ntr；
yv_mDimension be C × Ntr；
Finally, exporting y to verifyingv_mWith the R-matrix R of respective stagesjBetween error vector Jv_mjIt minimizes, is trained
The corresponding class label y of every level width neural network afterwardsv_ind_m:
Wherein, yv_ind_mDimension be 1 × Ntr。
5. the learning method according to claim 1 based on parallel multi-level width neural network, which is characterized in that step 3
In, it includes following sub-step that the statistics, which calculates:
Sub-step 3.1, the correct classification sample of the m level width neural network of the parallel M level width neural network after setting training
This collection and wrong classified sample set are respectively as follows: yvc_mAnd yvw_m, correct classified sample set and the sample that wrong classification samples are concentrated are total
Number is respectively as follows: Nvc_mAnd Nvw_m, and Nvc_m+Nvw_m=Ntr, then the error of correct classified sample set and wrong classified sample set is distinguished
Are as follows:
evc_m=| | softmax (yvc_m)-tvc_m||2；
evw_m=| | softmax (yvw_m)-tvw_m||2；
Wherein, tvc_mIt is correct classification samples y in m level width neural networkvc_mCorresponding true tag, tvw_mIt is m level width mind
Through classification samples y wrong in networkvw_mCorresponding true tag；
Sub-step 3.2, according to correct classified sample set yvc_mWith wrong classified sample set yvw_m, calculate separately out correct classification sample
This collection yvc_mMean value and variance be respectively μcAnd σc；Mistake classified sample set yvw_mMean value and variance be respectively: uwAnd σw；Then
Correct classified sample set yvc_mWith wrong classified sample set yvw_mCorresponding Gaussian Profile is respectively:
Correct classified sample set yvc_mWith wrong classified sample set yvw_mCorresponding Gaussian probability-density function is respectively:
Sub-step 3.3, according to wrong classified sample set yvw_mError evw_mAnd variances sigmaw, obtain determining for m level width neural network
Plan threshold value Tm=min (evw_m)-ασw；
Wherein, α is a constant, for providing allowance, so that the wrong classification samples y of institutevw_mIt is rejected when prime.
6. the learning method according to claim 2 based on parallel multi-level width neural network, which is characterized in that step 4
In, the acquisition test set are as follows: obtain original test sample collection xtest；It is corresponding to obtain M group test specimens by M data extending
This collection xrest_1..., xtest_m..., xtest_M, as test set.
7. the learning method according to claim 6 based on parallel multi-level width neural network, which is characterized in that the number
According to expansion are as follows: to the original test sample collection xtestIn each sample carry out N respectivelytestDThe secondary data transformation, it is corresponding
Obtain NtestDA test sample collection, the m level width neural network as the parallel M level width neural network that decision-making value determines
Test set xtest_m；
Wherein, original test sample collection xtestMiddle test sample sum is Ntest_samples。
8. the learning method according to claim 1 based on parallel multi-level width neural network, which is characterized in that step 4
In, the error vector for obtaining every level width neural network includes following sub-step:
Sub-step 4.1, by M group test sample collection xtest_1, xtest_2..., xtest_MDecision-making value is inputed to parallel respectively to determine
Parallel M level width neural network, the corresponding N for obtaining every level width neural network that decision-making value determinestestDA output
ytest_md, (d=1,2 ... NtestD)；
Sub-step 4.2, to the N for every level width neural network that decision-making value determinestestDA output ytest_m_d, (d=1,2 ... NtestD)
Average value is calculated, the test output for every level width neural network that decision-making value determines is obtained
Sub-step 4.3 sets total classification number of test set as C, constructs R-matrix Rj(1≤j≤C)；Obtain verifying output yv_m
With the R-matrix R of respective stagesjBetween error vector:
Jtest_mj=| | softmax (ytest_m)-Rj||2, 1≤j≤C；
Wherein, R-matrix RjThe element of jth row be all 1, remaining element is all 0, each R-matrix RjDimension be C ×
Ntest_samples；Jtest_mjDimension be 1 × Ntest_samples, yv_mDimension be C × Ntest_samples。
9. the learning method according to claim 8 based on parallel multi-level width neural network, which is characterized in that described right
The output for every level width neural network that decision-making value determines is judged are as follows:
The minimal error of current level width neural network is less than or equal to when prime decision-making value, then is judged as when prime is that this is defeated
Correct classification output stage out:
min(Jtest_mj)≤Tm；
The minimal error of current level width neural network is greater than when prime decision-making value, then is judged as when prime can not be defeated to this
Correctly classified out, which is transferred to next level width neural network and is tested, is so recycled, until the output is found
Correct classification output stage:
min(Jtest_mj) > Tm。
10. the learning method according to claim 9 based on parallel multi-level width neural network, which is characterized in that step 4
In, the test for obtaining every level width neural network that decision-making value determines exports corresponding label ytest_ind_mAre as follows:
Wherein, ytest_ind_mDimension be 1 × Ntest_samples。
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910331708.8A CN110110845B (en) | 2019-04-24 | 2019-04-24 | Learning method based on parallel multi-level width neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910331708.8A CN110110845B (en) | 2019-04-24 | 2019-04-24 | Learning method based on parallel multi-level width neural network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110110845A true CN110110845A (en) | 2019-08-09 |
| CN110110845B CN110110845B (en) | 2020-09-22 |
Family
ID=67486407
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910331708.8A Active CN110110845B (en) | 2019-04-24 | 2019-04-24 | Learning method based on parallel multi-level width neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110110845B (en) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111008647A (en) * | 2019-11-06 | 2020-04-14 | 长安大学 | Sample extraction and image classification method based on void convolution and residual linkage |
| CN111340184A (en) * | 2020-02-12 | 2020-06-26 | 北京理工大学 | Deformable reflector surface shape control method and device based on radial basis function |
| CN112966761A (en) * | 2021-03-16 | 2021-06-15 | 长安大学 | Extensible adaptive width neural network learning method |
| CN113449569A (en) * | 2020-03-27 | 2021-09-28 | 威海北洋电气集团股份有限公司 | Mechanical signal health state classification method and system based on distributed deep learning |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20160019454A1 (en) * | 2014-07-18 | 2016-01-21 | James LaRue | J Patrick's Ladder A Machine Learning Enhancement Tool |
| US20170132514A1 (en) * | 2012-12-24 | 2017-05-11 | Google Inc. | System and method for parallelizing convolutional neural networks |
| CN107784312A (en) * | 2016-08-24 | 2018-03-09 | 腾讯征信有限公司 | Machine learning model training method and device |
| CN108351985A (en) * | 2015-06-30 | 2018-07-31 | 亚利桑那州立大学董事会 | Method and apparatus for large-scale machines study |
-
2019
- 2019-04-24 CN CN201910331708.8A patent/CN110110845B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20170132514A1 (en) * | 2012-12-24 | 2017-05-11 | Google Inc. | System and method for parallelizing convolutional neural networks |
| US20160019454A1 (en) * | 2014-07-18 | 2016-01-21 | James LaRue | J Patrick's Ladder A Machine Learning Enhancement Tool |
| CN108351985A (en) * | 2015-06-30 | 2018-07-31 | 亚利桑那州立大学董事会 | Method and apparatus for large-scale machines study |
| CN107784312A (en) * | 2016-08-24 | 2018-03-09 | 腾讯征信有限公司 | Machine learning model training method and device |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111008647A (en) * | 2019-11-06 | 2020-04-14 | 长安大学 | Sample extraction and image classification method based on void convolution and residual linkage |
| CN111008647B (en) * | 2019-11-06 | 2022-02-08 | 长安大学 | Sample extraction and image classification method based on void convolution and residual linkage |
| CN111340184A (en) * | 2020-02-12 | 2020-06-26 | 北京理工大学 | Deformable reflector surface shape control method and device based on radial basis function |
| CN111340184B (en) * | 2020-02-12 | 2023-06-02 | 北京理工大学 | Deformable reflector surface shape control method and device based on radial basis function |
| CN113449569A (en) * | 2020-03-27 | 2021-09-28 | 威海北洋电气集团股份有限公司 | Mechanical signal health state classification method and system based on distributed deep learning |
| CN113449569B (en) * | 2020-03-27 | 2023-04-25 | 威海北洋电气集团股份有限公司 | Mechanical signal health state classification method and system based on distributed deep learning |
| CN112966761A (en) * | 2021-03-16 | 2021-06-15 | 长安大学 | Extensible adaptive width neural network learning method |
| CN112966761B (en) * | 2021-03-16 | 2024-03-19 | 长安大学 | Extensible self-adaptive width neural network learning method |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110110845B (en) | 2020-09-22 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110110845A (en) | Learning method based on parallel multi-level width neural network | |
| CN108564129B (en) | Trajectory data classification method based on generation countermeasure network | |
| CN107871100A (en) | The training method and device of faceform, face authentication method and device | |
| CN106529499A (en) | Fourier descriptor and gait energy image fusion feature-based gait identification method | |
| CN106295694A (en) | Face recognition method for iterative re-constrained group sparse representation classification | |
| CN112039687A (en) | Small sample feature-oriented fault diagnosis method based on improved generation countermeasure network | |
| CN111311702B (en) | Image generation and identification module and method based on BlockGAN | |
| CN110533116A (en) | Based on the adaptive set of Euclidean distance at unbalanced data classification method | |
| CN112949738B (en) | Multi-class unbalanced hyperspectral image classification method based on EECNN algorithm | |
| CN108108760A (en) | A kind of fast human face recognition | |
| CN106991355A (en) | The face identification method of the analytical type dictionary learning model kept based on topology | |
| CN106650667A (en) | Pedestrian detection method and system based on support vector machine | |
| CN104820825A (en) | Adaboost algorithm-based face recognition optimization method | |
| CN109344845A (en) | A kind of feature matching method based on Triplet deep neural network structure | |
| CN112819063B (en) | Image identification method based on improved Focal loss function | |
| CN110458178A (en) | The multi-modal RGB-D conspicuousness object detection method spliced more | |
| CN113139536A (en) | Text verification code identification method and equipment based on cross-domain meta learning and storage medium | |
| CN107886066A (en) | A kind of pedestrian detection method based on improvement HOG SSLBP | |
| CN109816030A (en) | A kind of image classification method and device based on limited Boltzmann machine | |
| CN109993042A (en) | A kind of face identification method and its device | |
| CN106778714A (en) | LDA face identification methods based on nonlinear characteristic and model combination | |
| CN109509188A (en) | A kind of transmission line of electricity typical defect recognition methods based on HOG feature | |
| CN104598898B (en) | A kind of Aerial Images system for rapidly identifying and its method for quickly identifying based on multitask topology learning | |
| CN110059705A (en) | A kind of OCR recognition result decision method and equipment based on modeling | |
| CN109948589A (en) | Facial expression recognizing method based on quantum deepness belief network |
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 |