CN109711536A - One kind being based on homologous successional super sausage neural network model and its construction method - Google Patents
One kind being based on homologous successional super sausage neural network model and its construction method Download PDFInfo
- Publication number
- CN109711536A CN109711536A CN201811641493.1A CN201811641493A CN109711536A CN 109711536 A CN109711536 A CN 109711536A CN 201811641493 A CN201811641493 A CN 201811641493A CN 109711536 A CN109711536 A CN 109711536A
- Authority
- CN
- China
- Prior art keywords
- neuron
- neural network
- super
- network model
- super sausage
- 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
Landscapes
- Meat, Egg Or Seafood Products (AREA)
- Image Analysis (AREA)
Abstract
The present invention provides one kind to be based on homologous successional super sausage neural network model and its construction method, and using feed forward type structure, the super sausage neural network model includes: input layer, the first hidden layer and output layer;Wherein, the input layer includes M neuron, and first hidden layer includes K group neuron, and every group of neuron includes L super sausage neurons, and the output layer is a neuron, and wherein M, K and L are natural number;The input layer is for receiving M dimensional vector x=(x1,x2...xm), the corresponding one-component received in M dimensional vector of each neuron of the input layer, and by the M dimensional vector x=(x1,x2...xm) map to each super sausage neuron in first hidden layer.The present invention overcomes traditional modes to identify the drawbacks of optimum division of the similar sample in feature space is as target, and solves the problems, such as that homogeneous object communication paths are cut off.
Description
Technical field
The present invention relates to field of artificial intelligence, more particularly to one kind to be based on homologous successional super sausage neural network
Model and its construction method.
Background technique
The problem of machine learning is studied is that machine how to be allowed to carry out study to have intelligence.Since artificial neural network is ground
The basic starting point studied carefully is mimic biology neural network, so being the strong means for realizing machine learning.
Artificial neural network is interconnected by a large amount of neurons, the connection between them it is extremely complex thus have it is extremely strong
Nonlinear Mapping, parallel processing, distributed storage and fault-tolerant ability, this is the basis for realizing intelligence.Artificial neural network
Advantage is to realize that Nonlinear Mapping is approached, exactly because this feature, artificial intelligence many aspects it can be seen that refreshing
Presence through network.
In recent years, the history of the existing many decades of the research of pattern-recognition, achieves many achievements.With pattern-recognition
Development, the correct recognition rata of pattern classification is higher and higher, however these method discriminations seem there is a bottleneck, are difficult to break through.
The country has a collection of scholar to be made that contribution in terms of neural network model, as Wang Shoujue academician proposes based on higher dimensional space geometry point
Analyse theoretical neural network model.The thought that Wang Shoujue academician proposes is inspiring, and the starting point of thought is will be from image
The angle of thinking investigates artificial intelligence problem, and particularly, he queries for traditional mode identification, it is believed that should be from understanding
It is not the angle of division to think deeply pattern recognition problem, is based on this thought, Wang Yuanshi proposes bionic pattern identification and higher-dimension
The reason of spatial information theory, which thinks, cause this result, does not make full use of homologous because of these methods
Things of the like description has part directly this priori knowledge of connectivity.Therefore, all can only from feature space inhomogeneity very
This division is set out, however really not so in the practical rule of nature.In nature all things be all by things of the like description gradually
Change develops, therefore the feature set of homologous things of the like description has the direct connectivity in part.If it exists two it is similar and incomplete
Equal things, then the difference of the two things can centainly find a gradual change with gradual change or unquantized
Sequence, this sequence changes to another from one in the two homologous samples, and all modes in this sequence
Belong to same class.
Summary of the invention
In view of the above technical problems, the purpose of the present invention is to provide one kind to be based on homologous successional super sausage nerve net
Network model and its construction method, to overcome traditional mode to identify optimum division of the similar sample in feature space as target
Drawback, and solve the problems, such as that homogeneous object communication paths are cut off.
In order to achieve the above object, the technical solution adopted in the present invention is as follows:
According to an aspect of the invention, there is provided a kind of be based on homologous successional super sausage neural network model, adopt
With feed forward type structure, comprising: input layer, the first hidden layer and output layer;Wherein, the input layer includes M neuron, described
First hidden layer includes K group neuron, and every group of neuron includes L super sausage neurons, and the output layer is a neuron,
Wherein M, K and L are natural number;
The input layer is for receiving M dimensional vector x=(x1,x2...xm), each neuron correspondence of the input layer connects
The one-component in M dimensional vector is received, and by the M dimensional vector x=(x1,x2...xm) map to it is each super in first hidden layer
Sausage neuron.
In certain embodiments of the present invention, the super sausage neural network model is approached using the mapping of neural network,
Be divided into two stages: the first stage is the fuzzy clustering for realizing the sample of the input space, and second stage is covered based on super sausage
Mapping solve.
In certain embodiments of the present invention, the input layer is by the M dimensional vector x=(x1,x2...xm) map to institute
Each super sausage neuron stated in the first hidden layer is realized by following operation: L that input vector is clustered with jth represent a littleThe r power of distance, wherein r >=1.
In certain embodiments of the present invention, the super sausage neural network model is using in first principal direction of each cluster
L represent a littleAnd it represents to select by this L and constitutes a super sausage geometrical body Covering samples point.
In certain embodiments of the present invention, the first stage is obtained by non-supervisory method, and the second stage is logical
Supervised learning is crossed to obtain.
In certain embodiments of the present invention, the fuzzy clustering of the first stage is using the K based on Euclidean distance
Mean cluster, or using minimum distance as the hierarchical clustering algorithm of measuring similarity.
According to another aspect of the present invention, a kind of building side of above-mentioned super sausage neural network model is additionally provided
Method, comprising the following steps:
The sample of a certain type is subjected to fuzzy clustering, such sample is divided into multiple groups;
Super sausage neuron is established in each group;
Super sausage neuron is connected with linear neuron, to constitute super sausage neural network model.
It can be seen from the above technical proposal that the present invention is based on homologous successional super sausage neural network model and its structures
Construction method at least has the advantages that the present invention is based on homologous continuity rule, with better Approximation and more preferably
Noise robustness, with super sausage structure for basic neuron, the network model is not with inhomogeneity sample in feature space
Optimum division is as target, but using Optimal coverage of a kind of sample in feature space as objective, ensure that homologous similar
The integrality of things communication paths.
Detailed description of the invention
Fig. 1 is the schematic diagram of the super sausage neuron of the embodiment of the present invention.
Fig. 2 is that the embodiment of the present invention selects the distance between super sausage neuron schematic diagram.
Fig. 3 is the structural schematic diagram of the super sausage neural network of the embodiment of the present invention.
Specific embodiment
To make the objectives, technical solutions, and advantages of the present invention clearer, below in conjunction with specific embodiment, and reference
Attached drawing, the present invention is described in more detail.
In an embodiment of the present invention, it provides a kind of based on homologous successional super sausage neural network model.Such as Fig. 3
Shown, the present invention is based on homologous successional super sausage neural network models to use three layers of feed forward type structure comprising: input
Layer, the first hidden layer and output layer.Input layer includes M neuron, and the first hidden layer includes K group neuron, every group of neuron
Comprising L super sausage neurons, output layer is a neuron, and wherein M, K and L are natural number.
The input layer is for receiving M dimensional vector x=(x1,x2...xm), each neuron of input layer is corresponding to receive M dimension
One-component in vector, and by the M dimensional vector x=(x1,x2...xm) map to each super sausage nerve in the first hidden layer
Member.
In a certain embodiment, the input layer is by the M dimensional vector x=(x1,x2...xm) map in the first hidden layer
Each super sausage neuron is realized by following operation: L that input vector is clustered with jth represent a littleDistance
R power, wherein r >=1.
Further, which is represented a little using L in first principal direction of each clusterAnd it represents to select by this L and constitutes a super sausage geometrical body Covering samples point.
In a certain embodiment, which is approached using the mapping of neural network, be specifically divided into
Lower two stages: the first stage is the fuzzy clustering for realizing the sample of the input space, and second stage is covered based on super sausage
Mapping solves.Wherein, the above-mentioned first stage is obtained by non-supervisory method, and above-mentioned second stage is obtained by supervised learning.
Preferably, the fuzzy clustering of above-mentioned first stage is using the K mean cluster based on Euclidean distance, with better
The directionality fluid flow structure of sample data is excavated out, and using minimum distance as the hierarchical clustering algorithm of measuring similarity
Deng.Super sausage neuron is established in the group of cluster.Place mat is carried out for the super sausage geometrical body covering based on principal direction.
By above-mentioned model, mapping is realized to the covering of input space sample, is overcome traditional mode and is identified similar sample
Originally the drawbacks of optimum division in feature space is as target, and solve the problems, such as that homogeneous object communication paths are cut off.
This super sausage neural network model provided by the invention constructs on the basis of homologous continuity rule.By mould
The mapping approximation problem of type is divided into two stages, and the first stage is the fuzzy clustering for realizing input space sample, and second stage is
Mapping based on super sausage overlay model solves.First stage is obtained by unsupervised learning, and second stage passes through supervised learning
It obtains.The neural network model is not only clear in structure, but also algorithm should be readily appreciated that and have operability well;The neural network
Model not only has good non-linear comparison ability, can be good at classifying to the sample for the classification trained,
But also have also can good rejection to the sample of untrained type.
According to the homologous principle of continuity, from a sample to another sample there are a continuous path, on the path
The point of process belongs to such sample point.An optimal covering is used, the sample space of this class is covered, thus is being instructed
The information between classification is had ignored in experienced process;And the classifier of traditional division mode, emphasis is differentiation between classification, and
Think little of the sample space of classification itself.Both modes are each defective, and have their own advantages, and the invention proposes a kind of by the two
The method combined constructs a kind of based on homologous successional super sausage neural network model, the sample to the classification trained
Originally can be good at classifying, it also can good rejection to the sample of untrained type.
The construction method of the super sausage neural network model is described in detail below.
Firstly, the sample of a certain type is carried out fuzzy clustering, such sample is divided into multiple groups (cluster), it is equal using k-
Value-based algorithm or minimum distance are the hierarchical clustering algorithm of similarity measurement, secondly, establishing super sausage nerve in each group
Member finally connects super sausage neuron with linear neuron, to constitute super sausage neural network structure (Hyper
Sausage Neural Network/HSNN)。
Super sausage neuron is the expansion product of a straightway and hypersphere, its expression formula are as follows:
Above-mentioned neuron is the simplest form of super sausage neuron, referred to as super sausage nerve primitive, wherein S (x1,
x2;It r) is the range of super sausage neuron covering,For the minimum range of the core of point x to super sausage neuron.It is super fragrant
The coverage area of enteric nervous member is as shown in Figure 1, the distance for selecting super sausage nerve primitive is as shown in Figure 2.
The network structure of SNN network is as shown in figure 3, be the feedforward shape neural network with double-layer structure, structure shape
Formula is similar with RBF network structure, and RBF neuron has been replaced as super sausage neuron.This method is covered compared with RBF network
Lid range is much smaller, can more accurately describe sample space.This method is eventually adding compared with traditional bionic pattern identification
One linear neuron to consider the influence of sample between classification can preferably classify for the classification learnt.
The detailed process being trained to super sausage neural network model of the invention is described in detail below.
Step1: input data X, input pre- cluster number of clusters K, allowable error upper bound ε.
Step2: the determination of super sausage neuron.Fuzzy clustering is carried out to every a kind of sample first, the method for cluster uses
K- mean algorithm, and using minimum distance as hierarchical clustering algorithm of measuring similarity etc..It is established in the group of cluster super fragrant
Enteric nervous member.The common output of super sausage neuron is as follows:
Wherein x indicates that input sample vector, y indicate that the core of the super sausage neuron of principal direction, d (x, y) are x point to super sausage
The minimum range of the core of neuron, σ2For sample variance, the radius of super sausage neuron represent.
Step3: the determination of linear neuron weight.Using under the least square method direct solution or gradient of guidance learning
Drop method recursive resolve.Wherein least square solution is as follows:
Order matrix:
Wherein, oijIndicate that i-th of sample j-th is surpassing the output valve on sausage neuron.Using least square solution w, obtain
To multiple linear regression:
O ω=y (6)
Wherein y is desired output, and least square solution can be obtained by multiple linear regression are as follows:
ω=(O'O)-1O'y (7)
Step4: suitable threshold value is chosen to obtained estimated value, is classified.
So far, whole training process of super sausage neural network model are completed, when use only needs to substitute into test vector
Mapping function value is sought in mapping expression formula.
So far, attached drawing is had been combined the present embodiment is described in detail.According to above description, those skilled in the art
It should be to the present invention is based on homologous successional super sausage neural network models and its construction method clear understanding.
It should be noted that in attached drawing or specification text, the implementation for not being painted or describing is affiliated technology
Form known to a person of ordinary skill in the art, is not described in detail in field.In addition, the above-mentioned definition to each element and method is simultaneously
It is not limited only to various specific structures, shape or the mode mentioned in embodiment, those of ordinary skill in the art can carry out letter to it
It singly changes or replaces.
Particular embodiments described above has carried out further in detail the purpose of the present invention, technical scheme and beneficial effects
It describes in detail bright, it should be understood that the above is only a specific embodiment of the present invention, is not intended to restrict the invention, it is all
Within the spirit and principles in the present invention, any modification, equivalent substitution, improvement and etc. done should be included in guarantor of the invention
Within the scope of shield.
Claims (7)
1. one kind is based on homologous successional super sausage neural network model, using feed forward type structure characterized by comprising
Input layer, the first hidden layer and output layer;Wherein,
The input layer includes M neuron, and first hidden layer includes K group neuron, and every group of neuron includes L super fragrant
Enteric nervous member, the output layer are a neuron, and wherein M, K and L are natural number;
The input layer is for receiving M dimensional vector x=(x1,x2...xm), each neuron of the input layer is corresponding to receive M dimension
One-component in vector, and by the M dimensional vector x=(x1,x2...xm) map to each super sausage in first hidden layer
Neuron.
2. super sausage neural network model according to claim 1, which is characterized in that the super sausage neural network model
It is approached using the mapping of neural network, be divided into two stages: the first stage is the fuzzy clustering for realizing the sample of the input space, the
Two-stage is that the mapping based on the covering of super sausage solves.
3. super sausage neural network model according to claim 2, which is characterized in that the input layer by the M tie up to
Measure x=(x1,x2...xm) each super sausage neuron for mapping in first hidden layer realized by following operation: input vector
It is represented a little with L of jth clusterThe r power of distance, wherein r >=1.
4. super sausage neural network model according to claim 3, which is characterized in that the super sausage neural network model
It is represented a little using L in first principal direction of each clusterAnd be made of this L representative point one it is super fragrant
Intestines geometrical body Covering samples point.
5. super sausage neural network model according to claim 2, which is characterized in that the first stage passes through non-supervisory
Method obtains, and the second stage is obtained by supervised learning.
6. super sausage neural network model according to claim 2, which is characterized in that the fuzzy clustering of the first stage
Using the K mean cluster based on Euclidean distance, or using minimum distance as the hierarchical clustering algorithm of measuring similarity.
7. a kind of construction method of super sausage neural network model described in any one of claims 1-6, which is characterized in that including
Following steps:
The sample of a certain type is subjected to fuzzy clustering, such sample is divided into multiple groups;
Super sausage neuron is established in each group;
Super sausage neuron is connected with linear neuron, to constitute super sausage neural network model.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811641493.1A CN109711536A (en) | 2018-12-29 | 2018-12-29 | One kind being based on homologous successional super sausage neural network model and its construction method |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811641493.1A CN109711536A (en) | 2018-12-29 | 2018-12-29 | One kind being based on homologous successional super sausage neural network model and its construction method |
Publications (1)
Publication Number | Publication Date |
---|---|
CN109711536A true CN109711536A (en) | 2019-05-03 |
Family
ID=66260284
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201811641493.1A Pending CN109711536A (en) | 2018-12-29 | 2018-12-29 | One kind being based on homologous successional super sausage neural network model and its construction method |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109711536A (en) |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111428799A (en) * | 2020-03-30 | 2020-07-17 | 北京市威富安防科技有限公司 | Image recognition model construction method and device, computer equipment and storage medium |
CN112446486A (en) * | 2019-08-30 | 2021-03-05 | 中国科学院半导体研究所 | Construction method of high-order neuron model |
-
2018
- 2018-12-29 CN CN201811641493.1A patent/CN109711536A/en active Pending
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112446486A (en) * | 2019-08-30 | 2021-03-05 | 中国科学院半导体研究所 | Construction method of high-order neuron model |
CN111428799A (en) * | 2020-03-30 | 2020-07-17 | 北京市威富安防科技有限公司 | Image recognition model construction method and device, computer equipment and storage medium |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN106250812B (en) | A kind of model recognizing method based on quick R-CNN deep neural network | |
CN106682696B (en) | The more example detection networks and its training method refined based on online example classification device | |
CN110263697A (en) | Pedestrian based on unsupervised learning recognition methods, device and medium again | |
CN103150580B (en) | A kind of high spectrum image semisupervised classification method and device | |
CN108021947B (en) | A kind of layering extreme learning machine target identification method of view-based access control model | |
Waheed et al. | Deep learning algorithms-based object detection and localization revisited | |
CN111476261A (en) | Community-enhanced graph convolution neural network method | |
Cai et al. | A multiobjective simultaneous learning framework for clustering and classification | |
CN108549866B (en) | Remote sensing airplane identification method based on dense convolutional neural network | |
Suganthan | Hierarchical overlapped SOM's for pattern classification | |
CN109711536A (en) | One kind being based on homologous successional super sausage neural network model and its construction method | |
CN109840518A (en) | A kind of visual pursuit method of combining classification and domain adaptation | |
CN104463251A (en) | Cancer gene expression profile data identification method based on integration of extreme learning machines | |
CN106647272A (en) | Robot route planning method by employing improved convolutional neural network based on K mean value | |
Zhou et al. | A novel K-means image clustering algorithm based on glowworm swarm optimization | |
CN116108383A (en) | Ship track prediction method based on improved goblet sea squirt multi-output support vector | |
Mengcan et al. | Constrained voting extreme learning machine and its application | |
CN105512675A (en) | Memory multi-point crossover gravitational search-based feature selection method | |
CN110490234A (en) | The construction method and classification method of classifier based on Cluster Classification associative mechanism | |
Chen et al. | Multi-granularity regularized re-balancing for class incremental learning | |
CN101546389A (en) | Primary direction neural network system | |
CN102799902A (en) | Enhanced relationship classifier based on representative samples | |
CN104537660B (en) | Image partition method based on Multiobjective Intelligent body evolution clustering algorithm | |
CN114373093A (en) | Fine-grained image classification method based on direct-push type semi-supervised deep learning | |
Shen et al. | Hybridization of particle swarm optimization with the K-Means algorithm for clustering analysis |
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: 20190503 |
|
RJ01 | Rejection of invention patent application after publication |