CN112560904A - Small sample target identification method based on self-adaptive model unknown element learning - Google Patents

Small sample target identification method based on self-adaptive model unknown element learning Download PDF

Info

Publication number
CN112560904A
CN112560904A CN202011388919.4A CN202011388919A CN112560904A CN 112560904 A CN112560904 A CN 112560904A CN 202011388919 A CN202011388919 A CN 202011388919A CN 112560904 A CN112560904 A CN 112560904A
Authority
CN
China
Prior art keywords
learning
meta
task
small sample
representing
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202011388919.4A
Other languages
Chinese (zh)
Inventor
庄连生
李厚强
杨健
樊硕
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
University of Science and Technology of China USTC
Original Assignee
University of Science and Technology of China USTC
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by University of Science and Technology of China USTC filed Critical University of Science and Technology of China USTC
Priority to CN202011388919.4A priority Critical patent/CN112560904A/en
Publication of CN112560904A publication Critical patent/CN112560904A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • General Engineering & Computer Science (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Software Systems (AREA)
  • Medical Informatics (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Image Analysis (AREA)

Abstract

The invention provides a small sample target recognition algorithm based on self-adaptive model unknown element learning, which comprises the following steps: step S101: initializing small sample target identification model parameters based on adaptive model unknown element learning; step S201: using a meta-learning method to iteratively update the specific parameters of the task; step S301: and optimizing the small sample identification model by utilizing the synthetic gradient direction. The method effectively improves the model generalization capability of the first-order MAML algorithm and the model convergence speed while maintaining the basic characteristics of the first-order MAML algorithm, and simultaneously maintains the space-time overhead almost consistent with the first-order MAML algorithm.

Description

Small sample target identification method based on self-adaptive model unknown element learning
Technical Field
The invention belongs to the field of small sample learning in machine learning, and particularly relates to a construction method of a small sample target recognition algorithm (hereinafter, referred to as a small sample target recognition algorithm based on adaptive MAML) based on adaptive model unknown element learning.
Background
The existing machine learning method can well solve the classification task under a large-capacity balanced data set, but under-fitting can be caused by insufficient data volume when small sample learning is carried out, so that the performance is poor. Whether a good model can be obtained by training with only a small amount of data becomes a key problem for small sample learning.
The meta-learning can well process small sample classification tasks, aims to enable a machine to learn by learning, and performs pre-learning on a plurality of similar small sample classification tasks by a systematic and data-driven method, so that the previously learned knowledge is adopted when a new task is faced, the decision process is guided, the new task is adapted to, and the purpose of fast learning of a model is achieved.
The small sample object recognition algorithm closely related to the present invention is a model-agnostic meta-learning (MAML) algorithm. However, the MAML algorithm needs to calculate the second-order gradient in the outer layer update phase, which results in a training process that occupies a large space-time overhead. Currently, there are a series of first order MAML algorithms that simplify the computation by ignoring second order gradient terms (fomal, reple), but lose some of the gradient information, resulting in varying degrees of accuracy loss.
Disclosure of Invention
In order to solve the problem of small sample classification in the field of machine learning, the invention provides a small sample target identification method based on adaptive model unknown meta-learning, which comprises the following steps:
step S101: initializing small sample identification model parameters based on self-adaptive model unknown element learning;
step S201: using a meta-learning method to iteratively update the specific parameters of the task;
step S301: and optimizing the small sample identification model by utilizing the synthetic gradient direction.
Further, the step S101 includes:
initializing parameters of a small sample identification model based on self-adaptive unknown element learning, wherein the parameters comprise: meta-model parameter theta, inner layer cyclic learning rate alpha, outer layer meta cyclic learning rate beta, inner layer cyclic super gradient step length alpha0And outer layer element circulation super gradient step length beta0。
Further, the step S201 includes:
sub-step S201 a: dividing a meta-learning task according to the auxiliary data set;
sub-step S201 b: inputting the divided meta-learning tasks into a small sample identification model based on self-adaptive model unknown meta-learning;
sub-step S201 c: and updating the specific parameters of the task according to a random gradient descent algorithm.
Further, the sub-step S201 a: the dividing of the meta-learning task according to the auxiliary data set specifically includes:
dividing small sample learning tasks according to N-type and K-type sample formats, and selecting 5-type and 3-type sample formats by default for explanation, wherein 5-type and 3-type samples refer to 5 types sampled from an auxiliary data set, 4 samples of each type are divided into a support set and a query set, the support set comprises 5 types, 3 samples of each type comprise 1 residual sample of each type in the 5 types, the model trains each task on the support set, and the training precision is verified on the query set.
Further, the sub-step S201 c: according to the divided meta-learning task, updating the specific parameters of the task by using a random gradient descent algorithm, wherein the expression of the specific parameters of the updated task is as follows:
Figure BDA0002811559710000021
wherein,
Figure BDA0002811559710000022
representing the concrete parameters of the j step of the meta-learning task i; alpha is alphatRepresenting the inner loop learning rate of the t iteration;
Figure BDA0002811559710000023
gradient operations representing continuous functions; l isiAnd k is the number of the small sample learning tasks.
Further, the step S301 includes:
sub-step S301 a: updating the average gradient direction according to the specific parameters of the task and the meta-learning module;
sub-step S301 b: updating the inner-layer cyclic learning rate according to the specific parameters of the task and the meta-learning module;
sub-step S301 c: and updating the outer-layer meta-learning rate according to the task specific parameters and the meta-learning module.
Further, the sub-step S301 a: updating the average gradient direction according to the specific parameters of the task and the meta-learning module, and updating the average gradient direction by adopting a first-order approximate gradient technology, wherein the specific expression is as follows:
Figure BDA0002811559710000024
wherein, thetatA meta-learning parameter representing a t-th iteration; beta is at-1Representing the outer-layer meta-learning rate of the t-th iteration;
Figure BDA0002811559710000025
and (5) representing the specific parameters of the small sample learning task at the kth step of the meta-learning task i.
Further, the sub-step S301 b: updating the inner-layer cyclic learning rate according to the specific parameters of the task and the meta-learning module; the expression for updating the inner loop learning rate is as follows:
Figure BDA0002811559710000031
wherein alpha istRepresenting the inner loop learning rate of the t iteration; alpha is alpha0Representing the inner layer circulation super-gradient step size;
Figure BDA0002811559710000032
representing concrete parameters of a small sample learning task in the jth step of the meta-learning task i;
Figure BDA0002811559710000035
gradient operations representing continuous functions; l isiRepresenting the loss function of the evaluation meta-learning task i.
Further, the sub-step S301 c: updating the outer-layer meta-learning rate according to the specific parameters of the task and the meta-learning module; the expression for updating the outer-layer meta-learning rate is:
Figure BDA0002811559710000033
wherein, thetatA meta-learning parameter representing a t-th iteration; beta is atRepresenting the outer-layer meta-learning rate of the t-th iteration; beta is a0Representing the step size of the outer layer element cyclic super-gradient;
Figure BDA0002811559710000034
and (5) representing the specific parameters of the small sample learning task at the kth step of the meta-learning task i.
Has the advantages that:
the invention provides a small sample target identification method based on self-adaptive model unknown element learning, which effectively improves the model generalization capability of a first-order MAML algorithm and the model convergence speed while maintaining the basic characteristics of the first-order MAML, and simultaneously maintains the space-time overhead almost consistent with the first-order MAML algorithm.
Drawings
FIG. 1 is a flow chart of a small sample target identification method based on adaptive model agnostic meta-learning according to the present invention;
FIG. 2(a) is a small sample classification task of the method of the present invention on Omniglot datasets;
FIG. 2(b) is a small sample classification task of the method of the present invention on the Mini-ImageNet dataset;
FIG. 3(a) 5-class-5-sample experiments on Omniglot datasets;
FIG. 3(b) 5-class-1-sample experiments on Omniglot data set;
FIG. 3(c) 5-class-5-sample experiments on Mini-ImageNet dataset;
FIG. 3(d) 5-class-1-sample experiments on Mini-ImageNet dataset;
fig. 4 is a convergence test of the adaptive MAML-based small sample target recognition algorithm according to the embodiment of the present invention.
Detailed Description
The technical solutions in the embodiments of the present invention will be described clearly and completely with reference to the accompanying drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, rather than all embodiments, and all other embodiments obtained by a person skilled in the art based on the embodiments of the present invention belong to the protection scope of the present invention without creative efforts.
The invention provides a small sample target identification method based on adaptive model unknown element learning, which comprises the following steps:
step S101: initializing adaptive MAML based small sample identification model parameters.
The method comprises the following steps of initializing parameters of a small sample identification model based on the self-adaptive MAML, wherein the parameters comprise: meta-model parameter theta, inner layer cyclic learning rate alpha, outer layer meta cyclic learning rate beta, inner layer cyclic super gradient step length alpha0And outer layer element circulation super gradient step length beta0. In general, the present invention recommends for alpha0And beta0By adopting the super-gradient step length setting of 1e-4, specific learning rate and meta-model parameters need to be manually selected to be proper values aiming at different tasks so as to achieve better effect.
Step S201: and (3) performing iterative updating on the specific parameters of the task by using a meta-learning method, wherein the iterative updating comprises the following steps:
sub-step S201 a: the meta-learning task is partitioned according to the auxiliary data set.
The invention divides the small sample learning task according to the N-type and K-sample formats, and selects the 5-type and 3-sample formats by default for explanation. The 5-class and 3-sample means that 5 classes of samples are sampled from the auxiliary data set, 4 samples of each class are divided into a support set and a query set, the support set comprises 5 classes of samples, 3 samples of each class, and the query set comprises the remaining 1 sample of each class of the 5 classes. The model trains each task on the support set, and verifies the training precision on the query set. In general, the present invention recommends partitioning the small sample learning task using a 5-class, 5-sample format or a 5-class, 1-sample format.
Sub-step S201 b: and inputting the divided meta-learning tasks into a small sample recognition model based on the self-adaptive MAML.
Sub-step S201 c: and updating specific parameters of the task by using a random gradient descent algorithm according to the divided meta-learning task.
The expression for updating the specific parameters of the task is as follows:
Figure BDA0002811559710000041
wherein,
Figure BDA0002811559710000042
representing the concrete parameters of the j step of the meta-learning task i; alpha is alphatRepresenting the inner loop learning rate of the t iteration;
Figure BDA0002811559710000043
gradient operations representing continuous functions; l isiRepresents a dedicated loss function for the evaluation meta-learning task i.
Step S301: optimizing the small sample identification model by using the direction of the synthetic gradient, wherein the method comprises the following steps:
sub-step S301 a: and updating the average gradient direction according to the task specific parameters and the meta-learning module.
Updating the average gradient direction by adopting a first-order approximate gradient technology, wherein the specific expression is as follows:
Figure BDA0002811559710000044
wherein, thetatA meta-learning parameter representing a t-th iteration; beta is at-1Representing the outer-layer meta-learning rate of the t-th iteration;
Figure BDA0002811559710000045
and (4) representing the k-th step task specific parameters of the meta-learning task i.
Sub-step S301 b: updating the inner-layer cyclic learning rate according to the specific parameters of the task and the meta-learning module;
the expression for updating the inner loop learning rate is as follows:
Figure BDA0002811559710000051
wherein alpha istRepresenting the inner loop learning rate of the t iteration; alpha is alpha0Representing the inner layer circulation super-gradient step size;
Figure BDA0002811559710000052
representing the concrete parameters of the j step of the meta-learning task i;
Figure BDA0002811559710000053
gradient operations representing continuous functions; l isiRepresents a dedicated loss function for the evaluation meta-learning task i.
Sub-step S301 c: updating the outer-layer meta-learning rate according to the specific parameters of the task and the meta-learning module; the expression for updating the outer-layer meta-learning rate is:
Figure BDA0002811559710000054
wherein, thetatA meta-learning parameter representing a t-th iteration; beta is atRepresenting the outer-layer meta-learning rate of the t-th iteration; beta is a0Representing the step size of the outer layer element cyclic super-gradient;
Figure BDA0002811559710000055
and (4) representing the k-th step task specific parameters of the meta-learning task i.
The first embodiment is as follows: and (3) performing model precision test on Omniglot and Mini-ImageNet data sets by a small sample target recognition algorithm based on the self-adaptive MAML.
For the present image classification problem (the second embodiment, the third embodiment and the same way), the loss function L of the small sample target recognition algorithm based on the adaptive MAMLiAre cross entropy loss functions.
For the present image classification problem (the second embodiment, the same way as the third embodiment), it is based onThe adaptive MAML small sample target recognition algorithm uses a standard four-layer convolutional neural network, each layer of which has a size of 3 × 32, followed by a ReLU activation function, a BN layer, and a pooling layer. For the present embodiment (the second embodiment, the same way as the third embodiment), the specific task parameters
Figure BDA0002811559710000056
I.e. the parameters of the standard four-layer convolutional neural network.
As can be seen from fig. 2(a) and (b), compared with other first-order MAML algorithms, the small sample target recognition algorithm based on the adaptive MAML has the advantage of model generalization capability in the small sample classification task on the Omniglot and Mini-ImageNet data set, and can effectively alleviate the model precision loss caused by neglecting the second-order gradient term.
Example two: and (3) testing convergence rate of a small sample target recognition algorithm based on the self-adaptive MAML and other first-order MAML algorithms.
As can be seen from fig. 3(a) and 3(c), when the division specification of 5-class-5-samples is used on the omniroot and Mini-ImageNet data sets, the adaptive MAML-based small sample target identification algorithm has an advantage in convergence speed compared with the replay algorithm belonging to the first-order MAML algorithm, and can reach smaller loss values with a small number of iterations.
As can be seen from fig. 3(b) and fig. 3(d), when the division specification of the 5-class-1-sample is used on the omniroot and Mini-ImageNet data set, the adaptive MAML-based small sample target identification algorithm still has advantages in convergence speed compared with the replay algorithm belonging to the first-order MAML algorithm, and can reach smaller loss values through a small number of iterations.
Example three: and (3) carrying out convergence test on the small sample target identification algorithm based on the self-adaptive MAML.
Through tests, the outer layer circulation super gradient step length beta is modified within a certain range (1 e-3-1 e-6)0The convergence influence on the small sample target identification algorithm based on the self-adaptive MAML is very small. Based on the invention, the super-gradient step setting of 1e-4 is recommended for the outer loop.
Adaptive MAML based small sampleThe convergence of the target identification algorithm depends on alpha of the step size of the inner loop super gradient0Is appropriately selected. The gradient of the loss function along with the inner loop is changed, as shown in FIG. 4, and based on the experiment of FIG. 4, the present invention recommends the super-gradient step setting of 1e-4 for the inner loop.
So far, the present invention has been described in detail with reference to the embodiments and the drawings of the embodiments. From the above description, those skilled in the art should clearly recognize the present invention.
It is to be noted that, in the attached drawings or in the description, the implementation modes not shown or described are all the modes known by the ordinary skilled person in the field of technology, and are not described in detail. In addition, the above definitions of the various elements are not limited to the specific structures, shapes or modes mentioned in the embodiments, and those skilled in the art may easily modify or replace them, for example:
(1) directional phrases used in the embodiments, such as "upper", "lower", "front", "rear", "left", "right", etc., refer only to the orientation of the attached drawings and are not intended to limit the scope of the present invention;
(2) the embodiments described above may be mixed and matched with each other or with other embodiments based on design and reliability considerations, i.e. technical features in different embodiments may be freely combined to form further embodiments.
The above-mentioned embodiments further explain the objects, technical solutions and advantages of the present invention in detail. It should be understood that the above-mentioned embodiments are only some specific examples of the present invention, and are not intended to limit the present invention, and any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims (9)

1. A small sample target identification method based on adaptive model unknown element learning comprises the following steps:
step S101: initializing small sample target identification model parameters based on adaptive model unknown element learning;
step S201: using a meta-learning method to iteratively update the specific parameters of the task;
step S301: and optimizing the small sample identification model by utilizing the synthetic gradient direction.
2. The method for small sample target recognition based on adaptive model-agnostic learning as claimed in claim 1, wherein the step S101 comprises:
initializing parameters of a small sample identification model based on self-adaptive unknown element learning, wherein the parameters comprise: meta-model parameter theta, inner layer cyclic learning rate alpha, outer layer meta cyclic learning rate beta, inner layer cyclic super gradient step length alpha0And outer layer element circulation super gradient step length beta0。
3. The method for identifying small sample objects based on adaptive model agnostic learning as claimed in claim 1, wherein the step S201 comprises:
sub-step S201 a: dividing a meta-learning task according to the auxiliary data set;
sub-step S201 b: inputting the divided meta-learning tasks into a small sample identification model based on self-adaptive model unknown meta-learning;
sub-step S201 c: and updating the specific parameters of the task according to a random gradient descent algorithm.
4. The adaptive model agnostic learning-based small sample object recognition method as claimed in claim 1, wherein the sub-step S201 a: dividing a meta-learning task according to an auxiliary data set, specifically comprising:
and dividing a plurality of small sample learning tasks according to the N-type and K-sample formats. The default selection of 5-class and 3-sample formats is used for explanation, 5-class and 3-sample means that 5 classes and 4 samples of each class are sampled from an auxiliary data set and are divided into a support set and a query set, the support set comprises 5 classes and 3 samples of each class, the query set comprises the remaining 1 sample of each class in the 5 classes, a model trains each task on the support set, and the training precision is verified on the query set.
5. The adaptive model agnostic learning-based small sample object recognition method as claimed in claim 1, wherein the sub-step S201 c: according to the divided meta-learning task, updating the specific parameters of the task by using a random gradient descent algorithm, wherein the expression of the specific parameters of the updated task is as follows:
Figure FDA0002811559700000011
wherein,
Figure FDA0002811559700000012
representing concrete parameters of a small sample learning task in the jth step of the meta-learning task i; alpha is alphatRepresenting the inner loop learning rate of the t iteration;
Figure FDA0002811559700000013
gradient operations representing continuous functions; l isiAnd k is the number of the small sample learning tasks.
6. The method for small sample target recognition based on adaptive model-agnostic learning as claimed in claim 1, wherein the step S301 comprises:
sub-step S301 a: updating the average gradient direction according to the specific parameters of the task and the meta-learning module;
sub-step S301 b: updating the inner-layer cyclic learning rate according to the specific parameters of the task and the meta-learning module;
sub-step S301 c: and updating the outer-layer meta-learning rate according to the task specific parameters and the meta-learning module.
7. The adaptive model agnostic learning-based small sample object recognition method of claim 1, wherein the sub-step S301 a: updating the average gradient direction according to the specific parameters of the task and the meta-learning module, and updating the average gradient direction by adopting a first-order approximate gradient technology, wherein the specific expression is as follows:
Figure FDA0002811559700000021
wherein, thetatA meta-learning parameter representing a t-th iteration; beta is at-1Representing the outer-layer meta-learning rate of the t-th iteration;
Figure FDA0002811559700000022
and (5) representing the specific parameters of the small sample learning task at the kth step of the meta-learning task i.
8. The adaptive model agnostic learning-based small sample object recognition method of claim 1, wherein the sub-step S301 b: updating the inner-layer cyclic learning rate according to the specific parameters of the task and the meta-learning module; the expression for updating the inner loop learning rate is as follows:
Figure FDA0002811559700000023
wherein alpha istRepresenting the inner loop learning rate of the t iteration; alpha is alpha0Representing the inner layer circulation super-gradient step size;
Figure FDA0002811559700000024
representing concrete parameters of a small sample learning task in the jth step of the meta-learning task i;
Figure FDA0002811559700000027
gradient operations representing continuous functions; l isiRepresenting the loss function of the evaluation meta-learning task i.
9. The adaptive model agnostic learning-based small sample object recognition method of claim 1, wherein the sub-step S301 c: updating the outer-layer meta-learning rate according to the specific parameters of the task and the meta-learning module; the expression for updating the outer-layer meta-learning rate is:
Figure FDA0002811559700000025
wherein, thetatA meta-learning parameter representing a t-th iteration; beta is atRepresenting the outer-layer meta-learning rate of the t-th iteration; beta is a0Representing the step size of the outer layer element cyclic super-gradient;
Figure FDA0002811559700000026
and (5) representing the specific parameters of the small sample learning task at the kth step of the meta-learning task i.
CN202011388919.4A 2020-12-01 2020-12-01 Small sample target identification method based on self-adaptive model unknown element learning Pending CN112560904A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202011388919.4A CN112560904A (en) 2020-12-01 2020-12-01 Small sample target identification method based on self-adaptive model unknown element learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202011388919.4A CN112560904A (en) 2020-12-01 2020-12-01 Small sample target identification method based on self-adaptive model unknown element learning

Publications (1)

Publication Number Publication Date
CN112560904A true CN112560904A (en) 2021-03-26

Family

ID=75047735

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202011388919.4A Pending CN112560904A (en) 2020-12-01 2020-12-01 Small sample target identification method based on self-adaptive model unknown element learning

Country Status (1)

Country Link
CN (1) CN112560904A (en)

Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111738301A (en) * 2020-05-28 2020-10-02 华南理工大学 Long-tail distribution image data identification method based on two-channel learning

Patent Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111738301A (en) * 2020-05-28 2020-10-02 华南理工大学 Long-tail distribution image data identification method based on two-channel learning

Similar Documents

Publication Publication Date Title
Arora et al. Understanding gradient descent on the edge of stability in deep learning
Chen et al. A tutorial on network embeddings
US20220076150A1 (en) Method, apparatus and system for estimating causality among observed variables
Ba et al. Blending diverse physical priors with neural networks
CN111814897A (en) Time series data classification method based on multi-level shape
CN111027636B (en) Unsupervised feature selection method and system based on multi-label learning
Constantinopoulos et al. An incremental training method for the probabilistic RBF network
CN108108455A (en) Method for pushing, device, storage medium and the electronic equipment of destination
Xiong et al. Recursive learning for sparse Markov models
CN108197225A (en) Sorting technique, device, storage medium and the electronic equipment of image
Kimura et al. A column-wise update algorithm for nonnegative matrix factorization in Bregman divergence with an orthogonal constraint
Neto et al. Opposite maps: Vector quantization algorithms for building reduced-set SVM and LSSVM classifiers
CN113255873A (en) Clustering longicorn herd optimization method, system, computer equipment and storage medium
Castellano et al. An empirical risk functional to improve learning in a neuro-fuzzy classifier
Ren et al. Structured optimal graph-based clustering with flexible embedding
Lember et al. Regenerativity of Viterbi process for pairwise Markov models
Nguyen et al. InfoCNF: An efficient conditional continuous normalizing flow with adaptive solvers
CN114399653A (en) Fast multi-view discrete clustering method and system based on anchor point diagram
CN118094216A (en) Multi-modal model optimization retrieval training method and storage medium
CN112560904A (en) Small sample target identification method based on self-adaptive model unknown element learning
KR20210119208A (en) Training method for model that imitates expert and apparatus thereof
Chen et al. On balancing neighborhood and global replacement strategies in MOEA/D
Kolaczyk et al. Statistical models for network graphs
CN114124729A (en) Dynamic heterogeneous network representation method based on meta-path
Nguyen et al. A novel online Bayes classifier

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: 20210326

RJ01 Rejection of invention patent application after publication