CN110210616A - A kind of zero degree learning classification algorithm for indicating to instruct feature selecting based on classification - Google Patents
A kind of zero degree learning classification algorithm for indicating to instruct feature selecting based on classification Download PDFInfo
- Publication number
- CN110210616A CN110210616A CN201910330421.3A CN201910330421A CN110210616A CN 110210616 A CN110210616 A CN 110210616A CN 201910330421 A CN201910330421 A CN 201910330421A CN 110210616 A CN110210616 A CN 110210616A
- Authority
- CN
- China
- Prior art keywords
- classification
- sample
- feature
- training
- feature selecting
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
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
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Image Analysis (AREA)
Abstract
The present invention relates to a kind of to indicate to instruct the zero degree learning classification algorithm of feature selecting based on classification, the invention firstly uses the structures of neural network self-encoding encoder, while training sample visual signature is penetrated to the mapping of classification semanteme, and its reflection, in classification semantic space, sample and classification are lost.Define an additional feature selecting layer, it will be used for the expression of classification corresponding to the batch data of network training every time, reflection is mapped in sample visual signature space, it is poor to indicate to make with sample, and seek the mean value of all differences, again by result by sigmoid function, the significance distribution for obtaining a feature is indicated, the value of itself and feature selecting layer is lost.Utilize two loss functions, training pattern.Feature selecting layer proposed by the present invention and its training method, can choose out the feature for being conducive to zero degree learning classification, to obtain better zero degree learning classification effect.
Description
Technical field
The present invention relates to the technical fields of artificial intelligence, instruct feature selecting based on classification expression more particularly to a kind of
Zero degree learning classification algorithm.
Background technique
It is well known that the successful application with deep learning in a variety of machine learning tasks, has constantly attracted more next
More researchers utilizes deep learning, goes to solve the problems in more human lives, however, the algorithm mould based on deep learning
Type, outstanding performance tend to rely on the labeled data of magnanimity, many data it is limited application (such as hazard event inspection
Survey) in, it is difficult to reach adaptable level, therefore, it is necessary to zero degree learning methods to be used only instruction by transportable information
The data for practicing collection classification, identify test set class;The existing zero degree for indicating to instruct feature selecting based on classification
Learning classification algorithm uses deep neural network to carry out feature extraction to sample mostly, directly establishes between sample and classification expression
Mapping, and do not select purpose is carried out for the feature of sample, so that final classifying quality is bad.
Summary of the invention
In order to solve the above technical problems, the present invention provides a kind of algorithm for being able to carry out feature selecting, so as to be zero degree
The sample classification of study provides better sample characteristics, realizes that one kind of better zero degree learning classification effect is based on category table
Show the zero degree learning classification algorithm for instructing feature selecting.
A kind of zero degree learning classification algorithm for indicating to instruct feature selecting based on classification of the invention, mainly includes following step
It is rapid:
Step a:
Feature extraction is carried out to sample data using deep neural network, obtains training dataAnd test data
Step b:
The neural network model of one basic self-encoding encoder form of training, establishes the connection between sample and corresponding classification
System realizes classification;
Step c:
In the model of step b, by training sampleBefore being mapped into semantic embedding space, point is multiplied by a feature choosing
A layer mask is selected, feature selecting is carried out, re-maps in semantic embedding space, with matrix S composed by training classification semantic expressivenesss
It loses;
Step d:
By training dataCorresponding classification semantic expressivenessUsing trained model in step b, reflection is mapped to sample
In eigen space, and and training dataIt is operated, obtains target masko, production mask loses, in guiding step c
The generation of feature selecting layer mask.In conjunction with all losses, model is trained;
Step e:
When test, by test sampleIt by mask layers and is mapped in classification semantic space, obtains corresponding classification.
A kind of zero degree learning classification algorithm for indicating to instruct feature selecting based on classification of the invention, in step c, if special
The dimension of sign selection layer mask is consistent with the dimension of a sample x, is R1×lIf the training sample inputted every timeDimension be
Rb×l, mask is replicated b parts, forms new tensor maskx∈Rb×l, by its withMake dot product, obtains the sample after feature selecting
ThisIt by it using the model learnt in step b, is mapped into semantic embedding space, with training classification semantic expressiveness institute group
At matrix SsIt loses.
A kind of zero degree learning classification algorithm for indicating to instruct feature selecting based on classification of the invention, in step d, for
It is input to the data being trained in model every timeBy its corresponding classification semantic expressivenessUtilize what is learnt in step b
Model, reflection are mapped in sample characteristics space, withIt is poor to make, and takes mean value on 0 dimension direction, and result is passed through sigmoid letter
Number, the significance distribution for obtaining a feature indicate masko, later, it is lost with feature selecting layer mask, in conjunction with step
Loss in c, is trained model.
Compared with prior art the invention has the benefit that the invention firstly uses the knots of neural network self-encoding encoder
Structure, while training sample visual signature is to the mapping of classification semanteme, and its reflection is penetrated, in classification semantic space, by sample with
Classification is lost.An additional feature selecting layer is defined, category table corresponding to the batch data of network training will be used for every time
Show, reflection is mapped in sample visual signature space, and it is poor to indicate to make with sample, and seeks the mean value of all differences, then result is passed through
Sigmoid function, the significance distribution for obtaining a feature indicates, the value of itself and feature selecting layer is lost.In conjunction with all
Loss, is trained, and final model can be obtained, and a kind of algorithm being able to carry out feature selecting can be the sample of zero degree study
This classification provides better sample characteristics, realizes better zero degree learning classification effect.
Detailed description of the invention
Fig. 1 is flow chart of the invention;
Specific embodiment
With reference to embodiment, the embodiment of the present invention is furthur described in detail.Following embodiment is used for
Illustrate the present invention, but is not intended to limit the scope of the invention.
A kind of zero degree learning classification algorithm for indicating to instruct feature selecting based on classification of the invention, mainly includes following step
It is rapid:
Step a:
Feature extraction is carried out to sample data using depth convolutional neural networks, obtains training dataWith
Test dataTraining set has n1A sample, test set have n2A sample, and the dimension of each sample is 1 × l;
This part carries out feature extraction to picture sample data using depth convolutional neural networks.Depth convolutional neural networks
Be it is generally acknowledged at present, to image data carry out feature extraction the best way, its available very effective sample characteristics.
Step b:
The neural network model of one basic self-encoding encoder form of training, if sample space reflecting to classification semantic space
It penetrates as FCe, counter to be mapped as FCd;The then loss function (1) of available self-encoding encoder training:
Wherein, Ss∈Rb×mIndicate that the set of training classification semantic expressiveness, total b trained classification, and each classification are 1 × m
Matrix;It indicates the corresponding class label of training set sample, is the form of one-hot, T representing matrix turns
It sets, wherein entropy_loss () indicates to intersect entropy loss, and MSE () indicates mean square deviation loss;
The detailed process of formula (1) is: after sample is mapped into classification semantic space, being classified using entropy loss is intersected
Training;Sample reflection is emitted back towards sample space again and original sample carries out mean square deviation loss;Above-mentioned steps can train one it is basic
Self-encoding encoder form neural network model;
Step c:
In the model of step b, by training sampleBefore being mapped into semantic embedding space, point is multiplied by a feature choosing
A layer mask is selected, feature selecting is first carried out, is re-mapped in semantic embedding space, it is semantic with the training classification after corresponding standardization
Matrix S composed by indicatingsMatrix multiple is carried out, by result and corresponding label YsMake intersection entropy loss;Specifically, if it is special
The dimension of sign selection layer mask is consistent with the dimension of a sample x, is a R1×lTensor, if the training sample inputted every timeDimension be Rb×l, mask is replicated b parts, forms new tensor maskx∈Rb×l, by its withMake dot product, obtains feature
Sample after selectionBy it using the model learnt in step b, it is mapped into semantic embedding space, with corresponding mark
Classification semantic expressiveness after standardizationMatrix multiple is carried out, by result and corresponding label YsMake intersection entropy loss.
Step d:
By training dataCorresponding classification semantic expressivenessUsing trained model in step b, reflection is mapped to sample
In eigen space, and and training dataIt is operated, obtains target masko, production mask loses, in guiding step c
The generation of feature selecting layer mask.In conjunction in step c intersection entropy loss and mask loss, model is trained.Come in detail
It says, for being input to the data being trained in model every timeBy its corresponding classification semantic expressivenessUsing in step b
The model of study, reflection are mapped in sample characteristics space, withIt is poor to make, and takes mean value on 0 dimension direction, and result is passed through
Sigmoid function, the significance distribution for obtaining a feature indicate masko, later, it is square with feature selecting layer mask work
Differential loss lose, in conjunction in step c intersection entropy loss and mask loss, model is trained.The formula of objective function such as (2) institute
Show.
Step e:
When test, by test sampleIt by mask layers and is mapped in classification semantic space, obtains corresponding category table
ShowFor each test set sample, the classification semantic expressiveness with its arest neighbors is found in spaceAs its
Corresponding classification.
Experimentation and result explanation
The present invention tests on data set CUB-200-2011Birds (CUB) [1], entirely the picture of birds,
200 class in total, 150 classes are training set, and 50 classes are test set, and the semanteme of classification is 312 dimensions, there is 11788 pictures.
Feature extraction is carried out using picture sample of the GoogLeNet [2] to data set, it is finally available on test set
60.7% accuracy rate.Compared to the result 58.1% for the basic self-encoding encoder model for not increasing mask layers, 2.8 hundred are increased
Branch.
The pertinent literature of retrieval is given below:
[1]Catherine Wah,Steve Branson,Peter Welinder,Pietro Perona,and Serge
Belongie,“The caltech-ucsd birds-200-2011dataset,”California Institute of
Technology,2011.
[2]Christian Szegedy,Wei Liu,Yangqing Jia,Pierre Sermanet,Scott Reed,
Dragomir Anguelov,Dumitru Erhan,Vincent Vanhoucke,and Andrew Rabinovich,
“Going deeper with convolutions,”in CVPR,2015,pp.1–9.
The above is only a preferred embodiment of the present invention, it is noted that for the ordinary skill people of the art
For member, without departing from the technical principles of the invention, several improvements and modifications, these improvements and modifications can also be made
Also it should be regarded as protection scope of the present invention.
Claims (3)
1. a kind of zero degree learning classification algorithm for indicating to instruct feature selecting based on classification, which is characterized in that mainly include following
Step:
Step a:
Feature extraction is carried out to sample data using deep neural network, obtains training dataAnd test data
Step b:
The neural network model of one basic self-encoding encoder form of training, establishes contacting between sample and corresponding classification, real
Now classify;
Step c:
In the model of step b, by training sampleBefore being mapped into semantic embedding space, point is multiplied by a feature selecting layer
Mask carries out feature selecting, re-maps in semantic embedding space, with matrix S composed by training classification semantic expressivenesssIt damages
It loses;
Step d:
By training dataCorresponding classification semantic expressivenessUsing trained model in step b, reflection is mapped to sample spy
It levies in space, and and training dataIt is operated, obtains target masko, make mask loss, the feature in guiding step c
Select the generation of layer mask.In conjunction with all losses, model is trained;
Step e:
When test, by test sampleIt by mask layers and is mapped in classification semantic space, obtains corresponding classification.
2. a kind of zero degree learning classification algorithm for indicating to instruct feature selecting based on classification as described in claim 1, feature
It is, is R if the dimension of feature selecting layer mask is consistent with the dimension of a sample x in step c1×lIf inputting every time
Training sampleDimension be Rb×l, mask is replicated b parts, forms new tensorBy its withMake a little
Multiply, obtains the sample after feature selectingBy it using the model learnt in step b, it is mapped into semantic embedding space,
With matrix S composed by training classification semantic expressivenesssIt loses.
3. a kind of zero degree learning classification algorithm for indicating to instruct feature selecting based on classification as described in claim 1, feature
It is, in step d, for being input to the data being trained in model every timeBy its corresponding classification semantic expressivenessUsing the model learnt in step b, reflection is mapped in sample characteristics space, withIt is poor to make, and takes on 0 dimension direction
Value, by result by sigmoid function, the significance distribution for obtaining a feature indicates masko, later, it is selected with feature
It selects a layer mask to lose, in conjunction with the loss in step c, model is trained.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910330421.3A CN110210616A (en) | 2019-04-23 | 2019-04-23 | A kind of zero degree learning classification algorithm for indicating to instruct feature selecting based on classification |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910330421.3A CN110210616A (en) | 2019-04-23 | 2019-04-23 | A kind of zero degree learning classification algorithm for indicating to instruct feature selecting based on classification |
Publications (1)
Publication Number | Publication Date |
---|---|
CN110210616A true CN110210616A (en) | 2019-09-06 |
Family
ID=67786321
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910330421.3A Pending CN110210616A (en) | 2019-04-23 | 2019-04-23 | A kind of zero degree learning classification algorithm for indicating to instruct feature selecting based on classification |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110210616A (en) |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112465035A (en) * | 2020-11-30 | 2021-03-09 | 上海寻梦信息技术有限公司 | Logistics distribution task allocation method, system, equipment and storage medium |
CN113762005A (en) * | 2020-11-09 | 2021-12-07 | 北京沃东天骏信息技术有限公司 | Method, device, equipment and medium for training feature selection model and classifying objects |
-
2019
- 2019-04-23 CN CN201910330421.3A patent/CN110210616A/en active Pending
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113762005A (en) * | 2020-11-09 | 2021-12-07 | 北京沃东天骏信息技术有限公司 | Method, device, equipment and medium for training feature selection model and classifying objects |
CN112465035A (en) * | 2020-11-30 | 2021-03-09 | 上海寻梦信息技术有限公司 | Logistics distribution task allocation method, system, equipment and storage medium |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110533045B (en) | Luggage X-ray contraband image semantic segmentation method combined with attention mechanism | |
CN106920243A (en) | The ceramic material part method for sequence image segmentation of improved full convolutional neural networks | |
CN111639679A (en) | Small sample learning method based on multi-scale metric learning | |
CN105184298B (en) | A kind of image classification method of quick local restriction low-rank coding | |
CN110413924A (en) | A kind of Web page classification method of semi-supervised multiple view study | |
CN104036293B (en) | Rapid binary encoding based high resolution remote sensing image scene classification method | |
US20150147728A1 (en) | Self Organizing Maps (SOMS) for Organizing, Categorizing, Browsing and/or Grading Large Collections of Assignments for Massive Online Education Systems | |
CN109299663A (en) | Hand-written script recognition methods, system and terminal device | |
CN104751175B (en) | SAR image multiclass mark scene classification method based on Incremental support vector machine | |
CN117529755A (en) | Transfer learning in image recognition systems | |
Akhtar et al. | Attack to fool and explain deep networks | |
CN113761259A (en) | Image processing method and device and computer equipment | |
CN109902662A (en) | A kind of pedestrian recognition methods, system, device and storage medium again | |
CN108537257A (en) | The zero sample classification method based on identification dictionary matrix pair | |
CN109344898A (en) | Convolutional neural networks image classification method based on sparse coding pre-training | |
CN108960270A (en) | A kind of data scaling method and system based on manifold transfer learning | |
CN110210616A (en) | A kind of zero degree learning classification algorithm for indicating to instruct feature selecting based on classification | |
CN108764361A (en) | The operating mode's switch method of beam type oil pumping machine indicating diagram based on integrated study | |
CN110263855A (en) | A method of it is projected using cobasis capsule and carries out image classification | |
CN108229505A (en) | Image classification method based on FISHER multistage dictionary learnings | |
Pondenkandath et al. | Identifying cross-depicted historical motifs | |
Bi et al. | Critical direction projection networks for few-shot learning | |
CN110163149A (en) | Acquisition methods, device and the storage medium of LBP feature | |
CN109271833A (en) | Target identification method, device and electronic equipment based on the sparse self-encoding encoder of stack | |
CN117690178A (en) | Face image recognition method and system based on computer vision |
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 | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190906 |
|
RJ01 | Rejection of invention patent application after publication |