CN107480780A - A kind of more threshold values polygamma functions select the design method of output artificial neuron more - Google Patents
A kind of more threshold values polygamma functions select the design method of output artificial neuron more Download PDFInfo
- Publication number
- CN107480780A CN107480780A CN201710782792.6A CN201710782792A CN107480780A CN 107480780 A CN107480780 A CN 107480780A CN 201710782792 A CN201710782792 A CN 201710782792A CN 107480780 A CN107480780 A CN 107480780A
- Authority
- CN
- China
- Prior art keywords
- threshold values
- artificial neuron
- function
- value
- input
- 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
- G06N3/04—Architecture, e.g. interconnection topology
-
- 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
- G06N3/06—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Evolutionary Computation (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Artificial Intelligence (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Neurology (AREA)
- Measurement And Recording Of Electrical Phenomena And Electrical Characteristics Of The Living Body (AREA)
Abstract
A kind of more threshold values polygamma functions select the technical field of the design method of output artificial neuron more, it is to belong to artificial intelligence, bionics, the technical field of circuit design, major technique is that artificial neuron is inputted by multichannel, when accumulated value is less than minimum threshold values, artificial neuron, it will not be activated, when cumulative value exceedes the threshold values of setting, artificial neuron is activated, artificial neuron is provided with multiple threshold values, cumulative value, activation primitive collection is passed to simultaneously and selects end-apparatus, activation primitive collection is according to the threshold values of reception, corresponding all functions below this threshold values are activated more, select threshold values of the end-apparatus according to reception simultaneously, open all passages corresponding to function that are activated, the value of various functions, from the port output of the respective function of setting, pass to next layer of neuron.
Description
Technical field
A kind of more threshold values polygamma functions select the technical field of the design method of output artificial neuron more, are to belong to artificial intelligence
Can, bionics, the technical field of circuit design, major technique is that artificial neuron is inputted by multichannel, when accumulated value is less than most
During small threshold values, artificial neuron, it will not be activated, when cumulative value exceedes the threshold values of setting, artificial neuron is activated, manually
Neuron is provided with multiple threshold values, cumulative value, while passes to activation primitive collection and selects end-apparatus, activation primitive collection is according to reception
Threshold values, corresponding all functions below this threshold values are activated more, while select threshold values of the end-apparatus according to reception, opens and all is swashed
Passage corresponding to function living, the value of various functions, from the port output of the respective function of setting, pass to next layer of neuron.
Background technology
Neuron is the elementary cell for forming brain, and the brain of the mankind is that have thousands of individual neurons according to certain rule
Form, for the mankind in order to simulate human brain, the design to artificial neuron is the most important thing, has artificial neuron to form people
Work network, artificial neural network are a kind of mathematical modulos for the structure progress information processing that application is similar to cerebral nerve cynapse connection
Type.In this model, composition network is coupled to each other between substantial amounts of artificial neuron, i.e. " neutral net ", to reach processing
The purpose of information.A kind of kinetic simulation for the distributed parallel information processing algorithm structure for imitating animal nerve network behavior feature
Type., with multichannel input stimulus are received, the part that " excitement " output is produced when exceeding certain threshold value by weighted sum is dynamic to imitate for it
The working method of thing neuron, and the weight coefficient of the structure being coupled to each other by these neural components and reflection strength of association makes
Its " collective behavior " has the various complicated information processing functions.Particularly it is this macroscopically have robust, it is fault-tolerant, anti-interference,
The formation of the flexible and strong function such as adaptability, self study can not only be updated by component performance, and pass through
Complicated interconnecting relation is achieved, thus artificial neural network is a kind of connection mechanism model, has many of complication system
Key character.Artificial neural network be applied to signal transacting, data compression, pattern-recognition, robot vision, knowledge processing and its
Using prediction, evaluation and the combinatorial optimization problem such as decision problem, scheduling, route planning.It can in Control System Design
For simulating controlled device characteristic, search and study control law, realizing fuzzy and intelligent control, therefore to the design of neuron
Very important, because fairly obvious, the shape of neuron is very more, although the mankind classify it, neuron has
Thousands of kinds, therefore different neurons also possesses different functions, the present invention is the design side of one of which neuron
Method, the design very simple of existing neuron is single, is exactly all inputs and multiplied by weight, is then added up, subtracted
Threshold values is removed, then activation primitive is set, passes to next layer of neuron.
The content of the invention
The brain of people is that many neurons are formed, therefore neuron is the elementary cell of neutral net, fairly obvious, nerve
First enormous amount, just there are the neuron of different shape, structure, physiologic character and function, neuron in the different parts of human body
Shape it is very strange very more, although the mankind classify to it, neuron has millions upon millions of kinds, therefore different nerves
Member also possesses different functions, and the present invention is that a kind of neuron therein is designed, due to the design of existing neuron
Very simple is single, and exactly all inputs and multiplied by weight are added up, and subtracts threshold values, then sets activation primitive, passes
Pass next layer of neuron, so form a network, and so simple design solve many forefathers of the mankind can not
Solve the problems, such as, tremendous influence, but a kind of this artificial neuron meta structure simply most simply, real generation are produced to All Around The World
The various shapes of neuron in boundary, various functions, therefore will invention various functions neuron design, this hair
Bright is exactly the design method of one of similar a variety of neuronal functions, and a kind of more threshold values polygamma functions select the artificial god of output more
Design method through member, it is characterized in that：More threshold values polygamma functions select more output artificial neuron be by input, artificial neuron,
Connecting line, end-apparatus, output end composition are selected, input receives the input of upper level artificial neuron such as the input of neuron
Or the input by other equipment, the effect of artificial neuron is added up after value and multiplied by weight input, if cumulative
Value be less than minimum threshold values, then artificial neuron would not be activated, without any reaction, if cumulative value is more than minimum
Threshold values, then artificial neuron is activated, and multiple threshold values are also set up on this minimum threshold values, when cumulative value is more than some valve
Value, this threshold values pass to activation primitive collection and select end-apparatus simultaneously, and activation primitive collection just starts all sharp below this threshold values
Function living, because the activation primitive of setting is different, therefore each function has a special connection to select end-apparatus more,
Connecting line is to be selected end-apparatus using such connection method, accumulator connection, threshold values is directly passed to and selects end-apparatus, in collection of functions
More corresponding some output ports for selecting end-apparatus of each function, how many function select the effect of end-apparatus with regard to how many bar line
It is such, according to the threshold values of input, the port that all respective functions below input threshold values are exported all connects, corresponding
Port output corresponding to the value from each function of function output below threshold values, the effect of output end is exactly that various activation primitives are defeated
The numerical value gone out is delivered to next layer of artificial neuron, and can be multiplied with weight, and wherein artificial neuron uses and such as divided into
Meter, artificial neuron is made up of 3 parts, and 1 is accumulator, and 2 be different threshold values, and 3 be different activation primitives, the work of accumulator
With being added up after input and multiplied by weight last layer, the design of different threshold values is such, sets minimum threshold values a,
a<b<c<D, when the value of input is less than a, then artificial neuron would not be activated, if the value of input is more than a, artificial neuron
Member is just activated, and the value at this moment inputted will be compared with different threshold values, for example the value inputted is less than d more than c, then just
The all activated function below c threshold values can be started, that is to say, that function corresponding to a, b, c threshold values is activated entirely, exports simultaneously
A, function f (x1), f (x2), f (x3) corresponding to b, c, select end-apparatus and use such design, it directly receives accumulator and is transmitted through
The threshold values come, according to the port of the selected output of threshold values, select end-apparatus and each function has a line more, and selecting end-apparatus can basis
Those interfaces of different threshold value settings and that functional link, the such design of the present invention just possess such function, work as input
Cumulative value exceedes threshold values, and artificial neuron is activated, and threshold values is transmitted to activation primitive collection simultaneously and selects end-apparatus, activation primitive collection
Function below input threshold values will be all activated, pass to and select end-apparatus, select end-apparatus and receive the threshold values passed over, gate
The passage of respective function, the value of different functions is passed to next layer of neuron.
Brief description of the drawings
Fig. 1 is that more threshold values polygamma functions select output artificial neuron's structure principle chart, i-1.i-2.i-3.i-4.i-5. i- more
6.i-7.i-8.i-9.i-10.i-11.i-12 represents input, and this input is a lot, and it is for masterpiece to draw 12 here
With o-1.o-2.o-3.o-4.o-5.o-6.o-7.o-8.o-9.o-10.i-11.i-12 represents output end, and this output end is very
More, it is for role of delegate to draw 12 here, and a-1 represents artificial neuron, and a-2 represents the insideAccumulator, a.b.c.d
It is to represent different threshold values, a-3 represents different activation primitive collection, and f (x1), f (x2), f (x3), f (x4) represent the inside
Different functions, this four functions are roles of delegate, and how many individual functions can be designed according to design requirement, and b-1 is represented and be selected end-apparatus,
B-2 representatives are selected end-apparatus and connected with accumulator, and b-3.b-4.b-5.b-6 represents activation primitive collection and connected with end-apparatus is selected, four here
It has been role of delegate, how many function, how many bar has been connected.
Implementation
Neuron species is various, Various Functions, invention emulates a kind of design method of neuron, using more threshold values polygamma functions
The methods for selecting output, it is by input that more threshold values polygamma functions select output artificial neuron more, artificial neuron, connecting line, is selected more
End-apparatus, output end composition, input receive the input of upper level artificial neuron or set by other such as the input of neuron
Standby input, the effect of artificial neuron is added up after value and multiplied by weight input, if cumulative value is less than most
Small threshold values, then artificial neuron would not be activated, without any reaction, if cumulative value is more than minimum threshold values, then
Artificial neuron is activated, and multiple threshold values are also set up on this minimum threshold values, when cumulative value is more than some threshold values, this valve
Value passes to activation primitive collection and selects end-apparatus simultaneously, and activation primitive collection just starts all activated function below this threshold values, by
In the activation primitive of setting be different, therefore each function more there is a special connection to select end-apparatus, connecting line is
Using such connection method, end-apparatus is selected in accumulator connection, and threshold values is directly passed to and selects end-apparatus, each letter in collection of functions
More corresponding some the port output ends for selecting end-apparatus of number, for how many function with regard to how many bar line, the effect for selecting end-apparatus is so
, according to the threshold values of input, the port that all respective functions below input threshold values are exported all connects, corresponding threshold value with
Under function output value from each function corresponding to port output, the effect of output end is exactly the number the output of various activation primitives
Value is delivered to next layer of artificial neuron, and the artificial neuron and the artificial neuron of other functions that the present invention designs network, structure
Into an artificial brain, it is possible to reach the function of imitating human brain, be to use more valves due to the artificial neuron of the present invention
Value polygamma function selects the form of output more, therefore can realize more functions, can use fewer artificial neuron of the invention
Member, reach sufficiently complex network function.
Claims (3)
1. a kind of more threshold values polygamma functions select the design method of output artificial neuron more, it is characterized in that：More threshold values polygamma functions are selected more
Output artificial neuron is by input, artificial neuron, connecting line, selects end-apparatus, output end composition, input is such as neuron
Input, receive upper level artificial neuron input or the input by other equipment, the effect of artificial neuron is defeated
Added up after the value and multiplied by weight that enter, if cumulative value is less than minimum threshold values, then artificial neuron would not be swashed
It is living, without any reaction, if cumulative value is more than minimum threshold values, then artificial neuron is activated, on this minimum threshold values
Multiple threshold values are also set up in face, and when cumulative value is more than some threshold values, this threshold values passes to activation primitive collection and selects end-apparatus simultaneously,
Activation primitive collection just starts all activated function below this threshold values, because the activation primitive of setting is different, therefore often
One function has a special connection to select end-apparatus more, and connecting line is selected using such connection method, accumulator connection
End-apparatus, threshold values is directly passed to and selects end-apparatus, more corresponding some output ports for selecting end-apparatus of each function in collection of functions, had
How many individual functions are with regard to how many bar line, and the effect for selecting end-apparatus is such, according to the threshold values of input, below input threshold values
The port of all respective function outputs all connects, port corresponding to the value from each function that the function below corresponding threshold value is exported
Output, the effect of output end are exactly that the numerical value of various activation primitives output is delivered to next layer of artificial neuron, and can be with
Weight is multiplied, and the such design of the present invention just possesses such function, when the cumulative value of input exceedes threshold values, artificial neuron
Member is activated, and threshold values is transmitted to activation primitive collection simultaneously and selects end-apparatus, activation primitive collection will be the function below input threshold values
All activation, pass to and select end-apparatus, select end-apparatus and receive the threshold values passed over, gate the passage of respective function, different letters
Several values passes to next layer of neuron.
2. according to the method for claim 1, artificial neuron is made up of using following design, artificial neuron 3 parts, and 1
It is accumulator, 2 be different threshold values, and 3 be different activation primitives, and the effect of accumulator is input and weight phase last layer
Added up after multiplying, the design of different threshold values is such, sets minimum threshold values a, a<b<c<D, when the value of input is less than a, that
Artificial neuron would not be activated, if the value of input is more than a, artificial neuron is just activated, and the value at this moment inputted is just
To be compared with different threshold values, for example the value inputted is less than d more than c, then will start all sharp below c threshold values
Function living, that is to say, that function corresponding to a, b, c threshold values is activated entirely, while exports function f (x1), f corresponding to a, b, c
(x2) 、f(x3)。
3. according to the method for claim 1, selecting end-apparatus using following design, it directly receives the valve that accumulator is transmitted through coming
Value, according to the port of the selected output of threshold values, select end-apparatus and each function has a line more, and selecting end-apparatus can be according to different
Those interfaces of threshold value setting and that functional link.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710782792.6A CN107480780A (en) | 2017-09-03 | 2017-09-03 | A kind of more threshold values polygamma functions select the design method of output artificial neuron more |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710782792.6A CN107480780A (en) | 2017-09-03 | 2017-09-03 | A kind of more threshold values polygamma functions select the design method of output artificial neuron more |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN107480780A true CN107480780A (en) | 2017-12-15 |
Family
ID=60604638
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710782792.6A Pending CN107480780A (en) | 2017-09-03 | 2017-09-03 | A kind of more threshold values polygamma functions select the design method of output artificial neuron more |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107480780A (en) |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN1107598A (en) * | 1993-06-14 | 1995-08-30 | 莫托罗拉公司 | Artificial neuron and method of using same |
| US6424961B1 (en) * | 1999-12-06 | 2002-07-23 | AYALA FRANCISCO JOSé | Adaptive neural learning system |
| CN1839397A (en) * | 2003-08-22 | 2006-09-27 | 西麦恩公司 | Neural network for processing arrays of data with existent topology, such as images, and application of the network |
| CN102035609A (en) * | 2010-12-15 | 2011-04-27 | 南京邮电大学 | Signal blind detection method based on a plurality of continuous unity feedback neural networks |
| CN103455843A (en) * | 2013-08-16 | 2013-12-18 | 华中科技大学 | Feedback artificial neural network training method and feedback artificial neural network calculating system |
-
2017
- 2017-09-03 CN CN201710782792.6A patent/CN107480780A/en active Pending
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN1107598A (en) * | 1993-06-14 | 1995-08-30 | 莫托罗拉公司 | Artificial neuron and method of using same |
| US6424961B1 (en) * | 1999-12-06 | 2002-07-23 | AYALA FRANCISCO JOSé | Adaptive neural learning system |
| CN1839397A (en) * | 2003-08-22 | 2006-09-27 | 西麦恩公司 | Neural network for processing arrays of data with existent topology, such as images, and application of the network |
| CN102035609A (en) * | 2010-12-15 | 2011-04-27 | 南京邮电大学 | Signal blind detection method based on a plurality of continuous unity feedback neural networks |
| CN103455843A (en) * | 2013-08-16 | 2013-12-18 | 华中科技大学 | Feedback artificial neural network training method and feedback artificial neural network calculating system |
Non-Patent Citations (1)
| Title |
|---|
| 韦一，沈继忠: "基于多阈值神经元的D型触发器设计", 《浙江大学学报(理学版)》 * |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN106982359B (en) | A kind of binocular video monitoring method, system and computer readable storage medium | |
| Niklas | Morphological evolution through complex domains of fitness. | |
| CN106295799B (en) | A kind of implementation method of deep learning multilayer neural network | |
| CN109063778A (en) | A kind of image aesthetic quality determines method and system | |
| CN107704925A (en) | The visual analysis system and method for deep neural network training process | |
| CN108875812B (en) | Driver behavior classification method based on branch convolutional neural network | |
| CN106227043A (en) | adaptive optimal control method | |
| CN109344888A (en) | A kind of image-recognizing method based on convolutional neural networks, device and equipment | |
| CN109522925A (en) | A kind of image-recognizing method, device and storage medium | |
| CN109711534A (en) | Dimensionality reduction model training method, device and electronic equipment | |
| CN107169566A (en) | Dynamic neural network model training method and device | |
| CN107220707A (en) | Dynamic neural network model training method and device based on 2-D data | |
| CN107516130A (en) | A kind of more threshold values polygamma functions select the design method of end output artificial neuron | |
| CN107480780A (en) | A kind of more threshold values polygamma functions select the design method of output artificial neuron more | |
| CN107886163A (en) | Single-object problem optimization method and device based on AGN and CNN | |
| CN109522185A (en) | A kind of method that model segmentation improves arithmetic speed | |
| CN107480779A (en) | Design method that is a kind of while exporting polymorphic function artificial neuron | |
| CN107563505A (en) | A kind of design method of external control implantation feedback artificial neuron | |
| CN107563503A (en) | A kind of codified selects the design method that threshold values selects function artificial neuron | |
| CN107480783A (en) | A kind of more threshold values list functions select the design method of end output artificial neuron | |
| CN107563506A (en) | A kind of voltage-frequency formula selects the design method that frequency exports artificial neuron | |
| CN107609636A (en) | A kind of polygamma function correspondingly exports the design method of feedback function artificial neuron | |
| CN107609640A (en) | A kind of threshold values selects the design method of end graded potential formula artificial neuron | |
| CN107657315A (en) | A kind of graded potential formula selects the design method of end output artificial neuron | |
| CN105373829B (en) | A kind of full Connection Neural Network structure |
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 |