CN103310229A - Multitask machine learning method and multitask machine learning device both used for image classification - Google Patents
Multitask machine learning method and multitask machine learning device both used for image classification Download PDFInfo
- Publication number
- CN103310229A CN103310229A CN2013102406335A CN201310240633A CN103310229A CN 103310229 A CN103310229 A CN 103310229A CN 2013102406335 A CN2013102406335 A CN 2013102406335A CN 201310240633 A CN201310240633 A CN 201310240633A CN 103310229 A CN103310229 A CN 103310229A
- Authority
- CN
- China
- Prior art keywords
- matrix
- regression
- label
- view data
- machine learning
- 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
Images
Abstract
The invention discloses a multitask machine learning method and a multitask machine learning device both used for image classification. The method and the device are characterized in that low rank approximation of a residual structure and a covariance matrix of a regression matrix are utilized simultaneously, probability modeling is performed on the residual structure, the regression matrix, a low rank decomposition of the regression matrix and the covariance matrix of the regression matrix, learning of parameters of a probability model is performed through a variational deduction method or a sampling method, and a regression matrix high in accuracy is acquired finally and used for image classification. By the scheme, on one side, correlativity information among multitasks in the residual structure is utilized, so that parameter learning accuracy can be improved to improve classification accuracy; on the other side, by performing low rank approximation on the covariance matrix of the regression matrix, calculating complexity of an algorithm can be effectively lowered.
Description
Technical field
The invention belongs to technical field of image processing, relate to a kind of multitask machine learning method of Images Classification, particularly a kind of residual error structure and regression matrix covariance matrix low-rank of utilizing approaches the multitask machine learning method that carries out Images Classification.
Background technology
Along with the arrival of large data age, mass data is excavated and is seemed particularly important.In mass data is excavated, how to utilize the information of from data with existing, excavating out to instruct the excavation of new data to become a new study hotspot.Particularly when the sample size of some task is less, utilizes multi-task learning can effectively reduce the time cost that mass data excavates and improve the acquisition of information accuracy.For example, in the view data classification task, utilize correlativity between the different classification task of view data can improve the nicety of grading of task; In bioinformatics, utilize correlativity between the different Detection tasks of DNA can promote the accuracy of detection of task.
The multitask machine learning method can be regarded as the multi-output regression problem:
Label matrix for N sample d output;
Observing matrix for N sample D dimensional feature;
Be regression matrix;
Bias vector for task; 1
NFor N ties up complete 1 column vector, subscript T representing matrix turn order; ∈ is the residual matrix of d * N.Dimension N wherein, d is natural number.Traditional multitask machine learning method adds that by finding the solution the error loss function optimization problem of regression coefficient matrix canonical constraint learns W:
Wherein ||
FBe the matrix F norm; R (W) is the canonical constraint of W, and common constraint is the matrix F norm, matrix l
2,1Norm, matrix l
P, qNorm etc.But the canonical bound term of this class norm has only been introduced the positive correlation between the task, and does not set up the negative correlation between the task.
For the negative correlation between the introducing task, be necessary W is carried out the modeling of correlativity between the task.A kind of constraint commonly used is tr (Σ
-1W Ω
-1W
T)+R (Σ, Ω).Here W is considered as one and obeys the matrix Gaussian distribution
Stochastic variable, wherein
It is matrix K ronecker product signs.Find the solution the optimization problem that the error loss function adds regression matrix covariance matrix constraint this moment and learn W.
But above-mentioned machine learning method has all been ignored the structure of residual matrix ∈.Also comprised the relevant information between the task in the residual matrix, correlation information can further promote the accuracy of learning regression matrix W between comprising in the residual matrix of the task if can further utilize.Because covariance matrix Σ and Ω dimension may be larger, if can not effectively approach it, may make multitask machine learning algorithm complexity excessive simultaneously.
Summary of the invention
The object of the invention is to overcome above-mentioned prior art shortcoming, propose a kind of multitask machine learning method and device thereof of the Images Classification that approaches based on residual error structure and regression matrix covariance matrix low-rank, thereby can effectively promote the prediction accuracy of regression matrix and then promote the precision of view data classification.
For achieving the above object, the multitask machine learning method for Images Classification of the present invention is at first introduced the view data of tens each classification task of the width of cloth and corresponding label data thereof; And then above-mentioned data are set up probability model, by the multitask machine learning method that utilizes residual error structure and regression matrix covariance matrix low-rank to approach, regression matrix is learnt; Utilize at last multi-task learning to regression matrix test pattern is classified.Concrete steps comprise:
S10, the training image data of input classification task and corresponding label data thereof extract feature to the method that view data uses image to process, and obtain observing matrix and the label matrix of view data;
S20 multiply by poor between the observing matrix by calculating label matrix and regression matrix, and the residual error structure is carried out probabilistic Modeling, namely sets residual matrix and obeys a certain probability distribution ∈~F (), and label matrix Y obeys average and is thus
Distribution F ();
S30 approaches the covariance matrix low-rank of regression matrix W and to carry out probabilistic Modeling, wherein:
S301, the low-rank decomposition of setting regression matrix is W=VZ, wherein V and Z are the low-rank matrix, with VV
TForm approach capable covariance matrix Σ, with Z
TThe form of Z is approached row covariance matrix Ω.And to W, V, Z carries out probabilistic Modeling:
W~F
W(VZ)
V~F
V(Σ)
Z~F
Z(Ω)
F wherein
W, F
V, F
ZBe respectively W, V, the probability distribution that Z sets;
S302 sets respectively probability distribution F to capable covariance matrix Σ and the row covariance matrix Ω of regression matrix W
Σ, F
Ω;
S40 learns the parameter in the probability model described in S20 and the S30 by the method for variation deduction or sampling.
S50, repeating step S40 is until the convergence of all parameters.
S60, the regression matrix W that utilizes multitask machine learning method study to obtain, the observing matrix that multiply by the test pattern of input obtains the estimated value of test pattern label matrix, estimates that with this label matrix classifies to the test pattern data.
Another object of the present invention also is to provide a kind of multitask machine learning classification device for Images Classification, it is characterized in that, comprises following three modules:
1) pre-processing image data module, it is input as view data and corresponding label data thereof, output is respectively observing matrix and the label matrix of view data, is used for the method that view data uses image to process is extracted observing matrix and the label matrix that feature obtains view data;
2) probabilistic Modeling parameter learning module, it is input as observing matrix and the corresponding label matrix thereof of view data, and it is output as regression matrix; Be used for:
The difference that multiply by between the observing matrix by calculating label matrix and regression matrix obtains residual matrix, and residual matrix is carried out probabilistic Modeling, and sets residual matrix and obey a certain probability distribution;
Regression matrix is carried out two decomposition of low-rank matrix, and wherein each split-matrix is respectively as the matrix that approaches of the capable covariance matrix of regression matrix and row covariance matrix; Regression matrix and two split-matrix are carried out probabilistic Modeling;
Method by variation deduction or sampling is learnt the parameter in the above-mentioned probability model;
3) view data sort module, it is input as regression matrix and test pattern data observation matrix, it is output as the estimated value of test pattern label matrix, be used for utilizing the multitask machine learning method to learn the regression matrix W that obtains, the observing matrix that multiply by the test pattern of input obtains the label matrix estimated value of test pattern, estimates that with this label matrix classifies to the test pattern data.
Advantage of the present invention is: utilize simultaneously the low-rank of residual error structure and regression matrix covariance matrix to approach and carry out the algorithm complex that the multitask machine learning can effectively be excavated the correlativity between the view data classification task and be reduced covariance matrix study; Parameter by probability model is upgraded, and can effectively improve the study precision of regression matrix, thereby can improve the precision of view data classification.
Description of drawings
Fig. 1 is the probability graph model of Gauss's Generalized Inverse Matrix Gauss model;
Fig. 2 is the process flow diagram of this multitask machine learning method embodiment that is used for Images Classification;
The schematic diagram of a kind of multitask machine learning classification device embodiment for Images Classification of Fig. 3 the present invention.
Embodiment
In order to make purpose of the present invention, technical scheme and advantage clearer, below in conjunction with drawings and Examples, the present invention is further elaborated.Should be appreciated that specific embodiment described herein only in order to explain the present invention, is not intended to limit the present invention.
On the contrary, the present invention contain any by claim definition in substituting of making of marrow of the present invention and scope, modification, equivalent method and scheme.Further, in order to make the public the present invention is had a better understanding, in hereinafter details of the present invention being described, detailed some specific detail sections of having described.There is not for a person skilled in the art the description of these detail sections can understand the present invention fully yet.
The embodiment of the invention one provides a kind of multitask machine learning method for Images Classification, and it may further comprise the steps:
S10, input picture characteristic and the corresponding label data of each data carry out filtering, extract color and vein, yardstick invariant features etc. view data;
S20, the difference that multiply by between the observing matrix X by calculating label matrix Y and regression matrix W obtains residual matrix, and the residual error structure is carried out probabilistic Modeling, wherein sets residual matrix and obeys the matrix Gaussian distribution
Label matrix obedience average is thus
The matrix Gaussian distribution
σ wherein
2For greater than zero real number, I
NThe unit matrix of expression N * N.
S30 approaches the covariance matrix low-rank of regression matrix and to carry out probabilistic Modeling, obtains Gauss's Generalized Inverse Matrix Gauss model (Gaussian Matrix Generalized Inverse Gaussian Model).It is defined as follows:
Ω~MGIG
D(Ψ
1,Φ
1,ν
1)
Σ~MGIG
d(Ψ
2,Φ
2,ν
2)
Then claim this Series correlation variable to consist of Gauss's Generalized Inverse Matrix Gauss model.Wherein
The expression Gaussian distribution, MGIG representing matrix generalized inverse Gaussian distribution.κ
l, κ
2Be positive number, dimension d, D, K are natural numbers.V and Z consist of the low-rank decomposition of regression matrix W, thereby its row covariance matrix Σ and row covariance matrix Ω exist respectively low-rank to approach VV
TAnd Z
TZ.Then this distribution relation is combined with Gauss's Generalized Inverse Matrix Gaussian distribution and just can obtain Gauss's Generalized Inverse Matrix Gauss regression model (Gaussian Matrix Generalized Inverse Gaussian Regression Model), its probability graph model as shown in Figure 1.
S40 learns the parameter in the probability model by the method that variation is inferred.S401 obtains under expectation step (E-phase) renewal
Wherein<〉is illustrated in the expectation of parameter in the given situation of other parameters, Ω
W=(<Ω
-1〉+σ
-2XX
T)
-1The renewal of Ω and Σ is estimated by the mode of sampling; S402 obtains under maximization step, (M ?phase) upgraded
S403 repeats update algorithm S401, and S402 is until the convergence of all parameters.
S50, the regression matrix W that utilizes study to arrive, the observing matrix that multiply by the test pattern data obtains the estimated value of test pattern label matrix
, with this estimated matrix view data is classified.
The parameter learning of S40 also can be realized by the method for sampling.Adopting the method for sampling, namely is to use the method for a large amount of training image data statistics sample frequencies that parameter is upgraded, and belongs to conventional method, does not do exemplifying at this.
Figure 3 shows that a kind of multitask machine learning classification device for Images Classification.This device comprises three modules: the pre-processing image data module; Probabilistic Modeling parameter learning module; The view data sort module.
The label data that is input as view data and correspondence thereof of pre-processing image data module, output is respectively observing matrix and the corresponding label matrix thereof of view data, and its function of finishing is identical with embodiment with the above-mentioned described function of method of view data being carried out filtering, extraction color, texture, yardstick invariant features for the multitask machine learning method of Images Classification with embodiment.
The input of probabilistic Modeling parameter learning module comprises observing matrix and the corresponding label matrix thereof of view data, and it is output as regression matrix.The probabilistic Modeling that its function of finishing and embodiment and above-mentioned multitask machine learning method for Images Classification approach residual error structure, regression matrix covariance matrix low-rank and the described function of method of using variation deduction to learn to parameter are identical with embodiment.
The view data sort module, it is input as regression matrix and view data observing matrix, it is output as the estimated value of label matrix, and its function of finishing is identical with embodiment to the described function of method that the test pattern observing matrix carries out the estimation of test pattern label matrix with above-mentioned multitask machine learning method for Images Classification with embodiment.
The above only is preferred embodiment of the present invention, not in order to limiting the present invention, all any modifications of doing within the spirit and principles in the present invention, is equal to and replaces and improvement etc., all should be included within protection scope of the present invention.
Claims (2)
1. a multitask machine learning method that is used for Images Classification is characterized in that comprising the steps:
S10, the view data of input classification task and corresponding label data thereof extract feature to the method that view data uses image to process, and obtain observing matrix and the label matrix of view data;
S20, the difference that multiply by between the observing matrix by calculating label matrix and regression matrix obtains residual matrix, and residual matrix is carried out probabilistic Modeling, and sets residual matrix and obey a certain probability distribution;
S30 carries out two decomposition of low-rank matrix to regression matrix, and wherein each split-matrix is respectively as the matrix that approaches of the capable covariance matrix of regression matrix and row covariance matrix; Regression matrix and two split-matrix are carried out probabilistic Modeling;
S40 learns the parameter in the probability model described in S20 and the S30 by the method for variation deduction or sampling;
S50, repeating step S40 is until the convergence of all parameters;
S60, the regression matrix W that utilizes multitask machine learning method study to obtain, the observing matrix that multiply by the test pattern of input obtains the label matrix estimated value of test pattern, estimates that with this label matrix classifies to the test pattern data.
2. a multitask machine learning classification device that is used for Images Classification is characterized in that, comprising:
The pre-processing image data module, it is input as view data and corresponding label data thereof, output is respectively observing matrix and the label matrix of view data, is used for the method that view data uses image to process is extracted observing matrix and the label matrix that feature obtains view data;
Probabilistic Modeling parameter learning module, it is input as observing matrix and the corresponding label matrix thereof of view data, and it is output as regression matrix; Be used for:
The difference that multiply by between the observing matrix by calculating label matrix and regression matrix obtains residual matrix, and residual matrix is carried out probabilistic Modeling, and sets residual matrix and obey a certain probability distribution;
Regression matrix is carried out two decomposition of low-rank matrix, and wherein each split-matrix is respectively as the matrix that approaches of the capable covariance matrix of regression matrix and row covariance matrix; Regression matrix and two split-matrix are carried out probabilistic Modeling;
Method by variation deduction or sampling is learnt the parameter in the above-mentioned probability model;
The view data sort module, it is input as regression matrix and test pattern data observation matrix, it is output as the estimated value of test pattern label matrix, be used for utilizing the multitask machine learning method to learn the regression matrix W that obtains, the observing matrix that multiply by the test pattern of input obtains the label matrix estimated value of test pattern, estimates that with this label matrix classifies to the test pattern data.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201310240633.5A CN103310229B (en) | 2013-06-15 | 2013-06-15 | A kind of multitask machine learning method for image classification and device thereof |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201310240633.5A CN103310229B (en) | 2013-06-15 | 2013-06-15 | A kind of multitask machine learning method for image classification and device thereof |
Publications (2)
Publication Number | Publication Date |
---|---|
CN103310229A true CN103310229A (en) | 2013-09-18 |
CN103310229B CN103310229B (en) | 2016-09-07 |
Family
ID=49135423
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201310240633.5A Expired - Fee Related CN103310229B (en) | 2013-06-15 | 2013-06-15 | A kind of multitask machine learning method for image classification and device thereof |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN103310229B (en) |
Cited By (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103678580A (en) * | 2013-12-07 | 2014-03-26 | 浙江大学 | Multitask machine learning method for text classification and device thereof |
CN106537420A (en) * | 2014-07-30 | 2017-03-22 | 三菱电机株式会社 | Method for transforming input signals |
CN108133224A (en) * | 2016-12-01 | 2018-06-08 | 富士通株式会社 | For assessing the device and method of classification task complexity |
WO2019058300A1 (en) * | 2017-09-21 | 2019-03-28 | International Business Machines Corporation | Data augmentation for image classification tasks |
CN111506832A (en) * | 2020-04-20 | 2020-08-07 | 盐城禅图智能科技有限公司 | Heterogeneous object completion method based on block matrix completion |
CN112329604A (en) * | 2020-11-03 | 2021-02-05 | 浙江大学 | Multi-modal emotion analysis method based on multi-dimensional low-rank decomposition |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20100329517A1 (en) * | 2009-06-26 | 2010-12-30 | Microsoft Corporation | Boosted face verification |
CN102819748A (en) * | 2012-07-19 | 2012-12-12 | 河南工业大学 | Classification and identification method and classification and identification device of sparse representations of destructive insects |
CN103150713A (en) * | 2013-01-29 | 2013-06-12 | 南京理工大学 | Image super-resolution method of utilizing image block classification sparse representation and self-adaptive aggregation |
US20130151441A1 (en) * | 2011-12-13 | 2013-06-13 | Xerox Corporation | Multi-task learning using bayesian model with enforced sparsity and leveraging of task correlations |
-
2013
- 2013-06-15 CN CN201310240633.5A patent/CN103310229B/en not_active Expired - Fee Related
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20100329517A1 (en) * | 2009-06-26 | 2010-12-30 | Microsoft Corporation | Boosted face verification |
US20130151441A1 (en) * | 2011-12-13 | 2013-06-13 | Xerox Corporation | Multi-task learning using bayesian model with enforced sparsity and leveraging of task correlations |
CN102819748A (en) * | 2012-07-19 | 2012-12-12 | 河南工业大学 | Classification and identification method and classification and identification device of sparse representations of destructive insects |
CN103150713A (en) * | 2013-01-29 | 2013-06-12 | 南京理工大学 | Image super-resolution method of utilizing image block classification sparse representation and self-adaptive aggregation |
Non-Patent Citations (3)
Title |
---|
CEDRIC ARCHAMBEAU等: "Sparse Bayesian Multi-Task Learning", 《ADVANCES IN NEURAL INFORMATION PROCESSING SYSTEMS》 * |
WONYUL LEE等: "Simultaneous Multiple Response Regression and Inverse Covariance Matrix Estimation via Penalized Gaussian Maximum Likelihood", 《JOURNAL OF MULTIVARIATE ANALYSIS》 * |
刘成等: "基于多任务学习的自然图像分类研究", 《计算机应用研究》 * |
Cited By (16)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103678580A (en) * | 2013-12-07 | 2014-03-26 | 浙江大学 | Multitask machine learning method for text classification and device thereof |
CN103678580B (en) * | 2013-12-07 | 2017-08-08 | 浙江大学 | A kind of multitask machine learning method and its device for text classification |
CN106537420A (en) * | 2014-07-30 | 2017-03-22 | 三菱电机株式会社 | Method for transforming input signals |
CN106537420B (en) * | 2014-07-30 | 2019-06-11 | 三菱电机株式会社 | Method for converted input signal |
CN108133224A (en) * | 2016-12-01 | 2018-06-08 | 富士通株式会社 | For assessing the device and method of classification task complexity |
CN108133224B (en) * | 2016-12-01 | 2021-11-16 | 富士通株式会社 | Method for evaluating complexity of classification task |
GB2580002A (en) * | 2017-09-21 | 2020-07-08 | Ibm | Data augmentation for image classification tasks |
US10614346B2 (en) | 2017-09-21 | 2020-04-07 | International Business Machines Corporation | Data augmentation for image classification tasks |
US10496902B2 (en) | 2017-09-21 | 2019-12-03 | International Business Machines Corporation | Data augmentation for image classification tasks |
GB2580002B (en) * | 2017-09-21 | 2021-01-13 | Ibm | Data augmentation for image classification tasks |
US11120309B2 (en) | 2017-09-21 | 2021-09-14 | International Business Machines Corporation | Data augmentation for image classification tasks |
WO2019058300A1 (en) * | 2017-09-21 | 2019-03-28 | International Business Machines Corporation | Data augmentation for image classification tasks |
US11238317B2 (en) | 2017-09-21 | 2022-02-01 | International Business Machines Corporation | Data augmentation for image classification tasks |
CN111506832A (en) * | 2020-04-20 | 2020-08-07 | 盐城禅图智能科技有限公司 | Heterogeneous object completion method based on block matrix completion |
CN111506832B (en) * | 2020-04-20 | 2023-07-28 | 郑州信息科技职业学院 | Heterogeneous object completion method based on block matrix completion |
CN112329604A (en) * | 2020-11-03 | 2021-02-05 | 浙江大学 | Multi-modal emotion analysis method based on multi-dimensional low-rank decomposition |
Also Published As
Publication number | Publication date |
---|---|
CN103310229B (en) | 2016-09-07 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Karimpouli et al. | Image-based velocity estimation of rock using convolutional neural networks | |
CN103310229A (en) | Multitask machine learning method and multitask machine learning device both used for image classification | |
Kroese et al. | Spatial process simulation | |
Tahmasebi et al. | MS-CCSIM: accelerating pattern-based geostatistical simulation of categorical variables using a multi-scale search in Fourier space | |
CN102810161B (en) | Method for detecting pedestrians in crowding scene | |
CN105740894A (en) | Semantic annotation method for hyperspectral remote sensing image | |
CN103440512A (en) | Identifying method of brain cognitive states based on tensor locality preserving projection | |
CN103226595B (en) | The clustering method of the high dimensional data of common factor analyzer is mixed based on Bayes | |
CN103235947B (en) | A kind of Handwritten Numeral Recognition Method and device | |
CN106599230A (en) | Method and system for evaluating distributed data mining model | |
CN103337060B (en) | A kind of texture image high precision denoising filter based on partial fractional differential equation | |
CN105678715A (en) | Method for removing image noise based on kernel regression total variation | |
CN105005760A (en) | Pedestrian re-identification method based on finite mixture model | |
CN103544488B (en) | A kind of face identification method and device | |
CN102148987A (en) | Compressed sensing image reconstructing method based on prior model and 10 norms | |
CN103942749A (en) | Hyperspectral ground feature classification method based on modified cluster hypothesis and semi-supervised extreme learning machine | |
CN104484886A (en) | Segmentation method and device for MR image | |
Lu et al. | Existence and controllability results for stochastic fractional evolution hemivariational inequalities | |
CN104077765B (en) | Image segmentation device, image partition method | |
CN111428678A (en) | Method for generating confrontation network remote sensing image sample expansion under space constraint condition for ground object change detection | |
Martínez et al. | Probing the geometry of data with diffusion fréchet functions | |
CN113449878B (en) | Data distributed incremental learning method, system, equipment and storage medium | |
Chen et al. | Rsmamba: Remote sensing image classification with state space model | |
CN105938524A (en) | Microorganism association network prediction method and apparatus | |
Shen et al. | Fast adaptive optimization of weighted vector median filters |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
C06 | Publication | ||
PB01 | Publication | ||
C10 | Entry into substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
C14 | Grant of patent or utility model | ||
GR01 | Patent grant | ||
CF01 | Termination of patent right due to non-payment of annual fee | ||
CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20160907 Termination date: 20210615 |