CN106156845A - A kind of method and apparatus for building neutral net - Google Patents
A kind of method and apparatus for building neutral net Download PDFInfo
- Publication number
- CN106156845A CN106156845A CN201510129042.XA CN201510129042A CN106156845A CN 106156845 A CN106156845 A CN 106156845A CN 201510129042 A CN201510129042 A CN 201510129042A CN 106156845 A CN106156845 A CN 106156845A
- Authority
- CN
- China
- Prior art keywords
- neuron
- factor
- output
- hidden layer
- neutral net
- 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
Abstract
It relates to a kind of method and apparatus for building neutral net.Described method includes: obtain the association relation model between target data and its influence factor, and described association relation model characterizes the relatedness between described target data and its influence factor;According to described association relation model, set the network topology of described neutral net;And utilize sample data that described neutral net is trained.According to the disclosure, it is provided that a kind of method being used for building neutral net of improvement, wherein set the network topology of neutral net by excavating the association relation model between inputoutput data.Utilize the method that the training time of network in the case of not affecting model prediction accuracy, can be greatly saved.
Description
Technical field
It relates to nerual network technique, relate more particularly to a kind of for building nerve net
The method and apparatus of network.
Background technology
Neutral net is a complicated network system, and it is by being referred to as " god in a large number
Through unit " simple process unit interconnect widely and formed.Neutral net can be anti-
Reflecting many basic features of human brain function, it is the non-linear dynamic study of a high complexity
System.Generally, neural network model is carried out table by network topology, node feature and learning rules
Show.The network topology of neutral net include the number of plies of network, the quantity of each layer neuron and
It is connected with each other mode between each neuron.Neutral net has proved to be a kind of highly effective
Method, and be widely used in prediction, image and speech pattern recognition and function optimization
In field.But simultaneously neutral net also has shortcoming associated with it, such as associate complexity,
Interpretability difference etc..And owing to neutral net existing the most unnecessary connection and limit power,
Training process may be caused the most loaded down with trivial details but also time-consuming.
At article " the Improving neural networks by delivered by Hinton et al.
preventing co-adaptation of feature detectors”CoRR,abs/1207.0580
(2012), in, the construction method of a kind of neutral net is disclosed, wherein when neural metwork training,
Randomly choose so that the weight of node of some hidden layer in network does not works, in order to reduce
There is annexation to imply the common effect of node, thus improve the performance of neutral net, prevent
Model over-fitting problem.
For purposes of illustration, figure 1 illustrates the neutral net proposed in this article
The flow chart of model training method.As it is shown in figure 1, first initialize the topology of neutral net
Structure.This can be included in step S101 especially, first set the input of neutral net because of
Son, the output factor;The number of the neuron in step S102, setting hidden layer;And
In step 103, to performing full connection between the neuron of each hidden layer.Then sample number is inputted
According to performing network training.Especially, sample data can be inputted in step S104, adjust god
Through number and the annexation of unit, train connection weight.This such as includes for each sample
The partial nerve unit making hidden layer and input layer at random does not works.Utilizing training sample
Being trained rear neural metwork training to complete, at this moment the connection weight of neuron is determined.?
The method provided in this article is particularly well-suited to the less situation of training sample, less at sample
In the case of can by eliminate over-fitting problem improve neutral net performance.But, work as instruction
When white silk sample is abundant, the method for this random drop hidden layer node but influences whether finally
The precision of prediction of neutral net.
Therefore, there is a kind of technology to building for neutral net in the art to change
The demand entered.
Summary of the invention
In view of this, present disclosure discloses a kind of method and apparatus for building neutral net,
Eliminate on it is at least part of or alleviate the problems referred to above.
According in the first aspect of the disclosure, it is provided that a kind of side for building neutral net
Method.The method can include obtaining the incidence relation mould between target data and its influence factor
Type, described association relation model characterizes associating between described target data and its influence factor
Property;According to described association relation model, set the network topology of described neutral net;And
Utilize sample data that described neutral net is trained.
In an embodiment of the first aspect according to the disclosure, according to described incidence relation
Model, the network topology setting described neutral net may include that based on described incidence relation
Model, determines the input factor and the output factor of described neural network and determines described nerve
The input factor of network and the connection exporting the factor and the neuron in hidden layer.
In another embodiment of the first aspect according to the disclosure, determine described neutral net
The input factor and the output factor and hidden layer in the connection of neuron may include that basis
Described association relation model, determines that each input factor of described neutral net is defeated for each
Go out the weight of the factor;And according to described weight, determine the described input factor and the output factor
Connection with the neuron of described hidden layer.
In the another embodiment of the first aspect according to the disclosure, according to described incidence relation
Model, the network topology setting described neutral net may include that for the first hidden layer,
Add the same number of neuron of number and the described output factor, described neuron respectively with
The corresponding output factor connects, and according to described association relation model, described in adding
Neuron is connected with the described input factor;And replicate for the neuron added,
To produce the neuron setting number for each output factor at described first hidden layer.
In another embodiment of the first aspect according to the disclosure, described method can enter one
Step includes for other hidden layers: add in the upper hidden layer with other hidden layers described
For the neuron that the neuron of each output factor is identical, and a hidden layer on described
In and carrying out between the neuron of the identical output factor in other hidden layers described complete
Connect.
In a further embodiment according to the first aspect of the disclosure, wherein according to described association
Relational model, the network topology setting described neutral net can also include: according to described pass
Connection relational model sets the number of the neuron of each hidden layer.
In the another embodiment of the first aspect according to the disclosure, wherein according to described association
The number of the neuron that relational model sets each hidden layer may include that determine defeated with each
Go out the weight that the factor is associated and account for the ratio of all weight sums;And based on a determination that described
Ratio, determines the number of neuron relevant to each output factor in each hidden layer.
In another embodiment of the first aspect according to the disclosure, based on a determination that described ratio
Example, determines that the number of neuron relevant to each output factor in each hidden layer can wrap
Include: determined the number of described neuron by following formula:
Wherein kjIndicate the number of the neuron being associated with the jth output factor;mjInstruction
Number with the input factor that the jth output factor is associated;And μjInstruction and jth
The weight that the output factor is associated accounts for the ratio of all weight sums;α is between 1 to 10
Predetermined constant.
In an embodiment of the first aspect according to the disclosure, described acquisition number of targets
According to and influence factor between association relation model may include that use Granger cause-and-effect diagram
Model, extracts including at least target data and the polynary time series data of possible influence factor thereof
Data element between incidence relation.
In another embodiment of the first aspect according to the disclosure, can remain set
In the case of the structure of the network topology of fixed described neutral net, described neutral net is entered
Row training.
Second aspect according to the disclosure, it is provided that a kind of equipment for building neutral net.
This equipment may include that model acquisition module, and being configured to obtain target data affects with it
Association relation model between factor, described association relation model characterize described target data with
Relatedness between its influence factor;Topology setting module, is configured to according to described association
Relational model, sets the network topology of described neutral net;And network training module, quilt
It is configured to utilize sample data that described neutral net is trained.
The third aspect according to the disclosure, it is provided that a kind of computer program product, it includes
Having computer program code, when being loaded in computer equipment, it makes this computer
Equipment performs the method for the first aspect according to the disclosure.
According to the disclosure, it is provided that a kind of method being used for building neutral net of improvement, its
In initially set neutral net by excavating the association relation model between inputoutput data
Network topology.Utilize the method can be in the case of not affecting model prediction accuracy, greatly
The big training time saving network.
Accompanying drawing explanation
By the embodiment combined shown by accompanying drawing is described in detail, the disclosure upper
Stating and other features will be apparent from, label identical in the accompanying drawing of the disclosure represents identical
Or similar parts.In the accompanying drawings:
Fig. 1 schematically shows a kind of side built for neutral net of the prior art
The example of method;
Fig. 2 schematically show an embodiment according to the disclosure for building god
Flow chart through the method for network;
Fig. 3 schematically shows the sampled data of an embodiment according to the disclosure
Example;
Fig. 4 schematically show an embodiment according to the disclosure target data and
The schematic diagram of the association relation model between its influence factor;
Fig. 5 A to Fig. 5 C schematically shows the hidden of an embodiment according to the disclosure
The schematic diagram of the structure containing layer;And
Fig. 6 schematically show an embodiment according to the disclosure for building god
Block diagram through the equipment of network.
Detailed description of the invention
Hereinafter, each illustrative embodiments of the disclosure will be described in detail with reference to the attached drawings.
It should be noted that, the preferred implementation being merely possible to example that these accompanying drawings and description relate to.
Can be it should be noted that according to description subsequently, it is easy to dream up knot disclosed herein
The replacement embodiment of structure and method, and can be in the public affairs claimed without departing from the disclosure
These alternate embodiments are used in the case of the principle opened.
Should be appreciated that providing these illustrative embodiments is only used to make art technology
Personnel better understood when and then realize the disclosure, and limits these public affairs the most by any way
The scope opened.The most in the accompanying drawings, for purposes of illustration, by optional step, module,
Unit etc. illustrate with dotted line frame.
Term as used herein " includes ", " comprising " and similar terms should be understood that
It is open term, i.e. " include/including but not limited to ".Term "based" is " at least partly
Ground based on ".Term " embodiment " expression " at least one embodiment ";" another implements term
Example " expression " at least one further embodiment ".The related definition of other terms will be retouched below
Be given in stating.
For purposes of illustration, before specifically describing the disclosure, will be first to neutral net
It is briefly described.Neutral net is by processing unit a large amount of, simple (referred to as neuron)
The complex networks system being interconnected to form widely, it reflects the many of human brain function
Basic feature, is the Kind of Nonlinear Dynamical System of a high complexity.The topology of neutral net
Structure is by including input layer, one or more layers hidden layer and output layer.Input layer is responsible for
Receive input data (the input factor), and be output to hidden layer by excitation function.
Neutral net can comprise one or more hidden layer, and each hidden layer includes some neurons.
Neuron is a multiple input single output information process unit, for carrying out the information of input
Nonlinear Processing.In brief, hidden layer is responsible for receiving the output of input layer, and through too drastic
Encourage function and be output to output layer.It should be noted that the neuron of hidden layer is not
Can observe from its exterior, say, that cannot feel the existence of neuron from outside.
Output layer is responsible for receiving the output of hidden layer, and produces the defeated of whole network by excitation function
Go out result.
Hereinafter, will be used for building god to according to embodiment of the present disclosure with reference to accompanying drawing
It is described through the method for network.In the following description, will be with reference to such as PM2.5, SO2
Air Quality Forecast embodiment of this disclosure be described.However, it is necessary to explanation
It is that the disclosure is not limited to that, but can be used for any other occasion needed.
Fig. 2 schematically show an embodiment according to the disclosure for building god
Flow chart through the method for network.As it is shown in figure 1, first in step S201, obtain target
Association relation model between data and its influence factor, described association relation model characterizes institute
State the relatedness between target data and its influence factor.
In embodiment of the present disclosure, will first obtain the number of targets such as such as PM2.5, SO2
According to and influence factor between association relation model, the reflection of described association relation model is described
Incidence relation between target data and its influence factor, i.e. target data with which affect because of
Element is associated and correlation degree (optional).This association relation model can be by the suitableeest
When mode from gather extracting data.In an embodiment according to the disclosure, can
To use Granger carsal graph model, extract including at least target data and may affect because of
Incidence relation between the data element of the polynary time series data of element.
Gather data and such as can include meteorological data and air quality data, and also permissible
Preferably include at least one in traffic data, density data of population and polluter data.
For collecting data, first data are carried out pretreatment.Such as can be normalized to tool
There is the data sequence of identical time scale, the most such as, by difference processing, it is performed steadily
Change processes, the most such as, retrain based on duration and time lag and be translated into sample sequence.
Hereafter, cause-and-effect diagram iterative learning procedure can be used to identify the data of polynary time series data
Association between element is to obtain cause-and-effect diagram sequence.For example, it is possible to first data sequence is drawn
It is divided into the section of predetermined number, obtains the cause-and-effect diagram of each section of sequence.May then pass through calculating
The error of fitting of each sample in whole sequence and the cause-and-effect diagram obtained, and by each sample
Again incorporate in the sample segments corresponding to the cause-and-effect diagram minimum with its error of fitting.For newly
Sample Refreshment cause-and-effect diagram, and merge similar cause-and-effect diagram and corresponding sample segments.Such as,
Can be based on the weight calculation Euclidean distance of each time series data element in cause-and-effect diagram, thus
Determine similar cause-and-effect diagram.Subsequently, it may be determined that the cause-and-effect diagram of renewal and former cause-and-effect diagram it
Between change.If cause-and-effect diagram does not change, then it is assumed that cause-and-effect diagram iteration convergence, and
Terminate this cause-and-effect diagram iterative learning procedure.Otherwise, the error of fitting of each sample is recalculated
And each sample is incorporated into again the sample corresponding to the cause-and-effect diagram minimum with its error of fitting
Duan Zhong, and run next round calculating.The cause and effect finally given after determining iteration convergence is closed
It is that figure is association relation model.
For purposes of illustration, according to the disclosure is schematically shown in figure 3
The example of the collection data of embodiment, and schematically show in the diagram according to this
Incidence relation figure between target data and the influence factor thereof of a disclosed embodiment
Schematic diagram.
As it is shown on figure 3, gather data can be included in the carbon monoxide number in different time points
According to, PM2.5 data, SO2 data, traffic data, temperature data, humidity data, wind
Speed data etc..By incidence relation extracting method the most given above, for gathering data
After performing incidence relation analysis, can obtain target data SO2 and PM2.5 affects with each
Association relation model between factor, as shown in Figure 4.In the diagram it is clear that
SO2 is temperature dependent with traffic, and PM2.5 is relevant to traffic and wind speed.Under in the diagram
The form in portion shows each influence factor weight for each target data.
Then in step S202, according to described association relation model, described neutral net is set
Network topology.
Different from traditional method presets the network topology of neutral net based on experience,
In embodiment of the present disclosure, the network setting neutral net based on association relation model is opened up
Flutter.Such as may determine that the input factor in the input layer of neuroid, output layer defeated
Go out the connection pass of the factor and the described input factor and the output factor and the neuron in hidden layer
System.The number of hidden layer can be and every according to by engineering test or by empirically determined
The data of the neuron in Ceng can rule of thumb determine or according to association relation model by formula
Determine.It will be described in detail below.
Associating between the target data characterized according to association relation model with its influence factor
Relation, it may be determined that go out in the input factor in the input layer of neutral net and output layer is defeated
Go out the factor.Especially, for the target data characterized in association relation model and have with it
They can be identified as exporting the factor and the input factor by the influence factor of relatedness.
Such as, according to the association relation model shown in Fig. 4, it may be determined that the output factor is PM2.5
And SO2, and inputting the factor is traffic data, temperature data and air speed data.It is determined by
The input factor and the output factor, it may be determined that go out data that neuron in input layer receives and
The data of the neuron output in output layer.Therefore, according in embodiment of the present disclosure,
Based on the incidence relation mould between target data and the influence factor thereof gathering extracting data
Type determines the input factor and the output factor, and is not based on empirical relation and sets.By this
The mode of kind can reduce unnecessary input data in the case of ensureing precision of prediction, this
Will assist in the time saving network training.
Additionally, further, it is possible to determine that each input factor of neutral net is defeated for each
Going out the weight of the factor, as shown in the bottom form in Fig. 4, this weight reflects that each inputs
The factor is for exporting influence degree or the importance of the factor.Described weight can be subsequently
For determining described input layer, the described output layer neuron with described hidden layer in step
Connect.
For the topological structure of neutral net, in addition to input layer and output layer, each
The number of the neuron in hidden layer is also to need to set.In embodiment of the present disclosure,
The number of the neuron in each hidden layer can determine in various ways.For example, it is possible to
Rule of thumb formula determines, or alternatively determines this number also according to association relation model
Mesh.
Such as, in an embodiment according to the disclosure, the nerve in each hidden layer
The number of unit uses following empirical equation to determine:
In addition it is also possible to divide based on the association relation model shown in Fig. 4.Example
As, can determine that the weight being associated with each output factor accounts for according to association relation model
The ratio of all weight sums;It is then based on the described ratio determined, exports described first
Described neuron is distributed between the factor and the second output factor.Such as with the first output factor
The weight that SO2 and second output factor PM2.5 is associated accounts for ratio μ of all weight sums1
And μ2It is respectively as follows:
In this manner it is possible to according to the ratio-dependent determined go out in each hidden layer for first output because of
The neuron node of son is the joint that 3 (0.4*7 ≈ 3) are used for the neuron of the second output factor
Point is 4 (0.6*7 ≈ 4).
According in another embodiment of the disclosure, can utilize according to association relation model
Another way determines the number of the neuron in each hidden layer.For example, it is possible to according to pass
Connection relational model, determines that the weight being associated with each output factor accounts for all weight sums
Ratio;Be then based on the described ratio determined, determine in each hidden layer with each output because of
The number of the neuron that son is relevant.Especially, in a particular instance, can be based on logical
Cross following formula to determine the number of described neuron:
Wherein kjIndicate the number of the neuron being associated with the jth output factor;mjInstruction and jth
The number of the input factor that the individual output factor is associated;And μjInstruction with jth output because of
The weight that son is associated accounts for the ratio of all weight sums;α is making a reservation between 1 to 10
Constant.Such as, for the association relation model shown in Fig. 4, can by with first output because of
Sub-SO2 and the weight being associated with second output factor PM2.5 account for all weight sums respectively
Ratio μ1And μ2It is identified as 0.4 and 0.6, as shown in Equation 2.Then, according to upper
The formula 3 that face is given, it may be determined that in hidden layer for first output the factor and second output because of
The interstitial content of son is respectively
By such mode, it is possible to determine that each hidden layer includes 9 nodes altogether,
Wherein 4 is that 5 for second output factor PM2.5 for the first output factor S O2.
After determining hidden layer number, the hidden layer structure of neutral net can be performed.Example
As for the first hidden layer, can first added the same number of god of number and the output factor
Through unit, and described neuron is connected with the corresponding output factor respectively.Then, according to institute
State the input factor represented by association relation model and the incidence relation between the output factor, will
With the input that the output neuron that is connected of the factor is connected to relation factor-related with this output because of
Son.Fig. 5 A shows after adding neuron and it being connected with the input factor and the output factor
The diagram of neutral net.As shown in Figure 5A, wherein with the addition of and the first output factor S O2
Two neurons corresponding with second output factor PM2.5, wherein utilize with a circle filled
It is shown for the neuron that the first output factor S O2 is added, and utilizes with crossed grid filling
Circle is shown for the neuron that second output factor PM2.5 is added.In fig. 5, institute
Add neuron one of them with first output factor S O2 be connected, and foundation Fig. 4 show
The association relation model gone out, is connected this neuron with traffic and temperature input factor, and another
One neuron and second output factor PM2.5 connect, and according to the association shown in Fig. 4
Relational model, this neuron also inputs the factor with traffic and wind speed and is connected.
It follows that can replicate for the neuron added, to imply described first
Layer produces the neuron setting number for each output factor.For above-mentioned 4 nerves
Unit is to export the factor for the first output factor S O2 and 5 neurons for second
The situation of PM2.5, can additionally replicate three and export factor S O2 institute with above for first
Add neuron node, and replicate four with above for second export factor PM2.5 add
The neuron node added.These nodes and the node being replicated replicated with input the factor and defeated
Go out the factor and there is identical annexation.As shown in Figure 5 B.
Further, in the case of neutral net also has other hidden layers, can be for other
Hidden layer: add and export the factor for each in the upper hidden layer of these other hidden layers
The identical neuron of neuron.And in a upper hidden layer and in these other hidden layers
Entirely connect between the neuron of the identical output factor.Such as the second hidden layer,
The neuron identical with the neuron number in the first hidden layer, the most altogether 9 god can be added
Through unit, and wherein 4 for first output factor S O2,5 for second output the factor
PM2.5.The neuron of the second hidden layer added replaces ground floor neuron and the input factor
Connect, and at the first hidden layer and the neuron for the identical output factor of the second hidden layer
Between perform full connection, as shown in Figure 5 C.For the 3rd or the hidden layer of more top, permissible
Employing similar fashion determines, the most for simplicity, is no longer described in detail.Herein,
Although it should be noted that the node in each hidden layer and the interstitial content of other hidden layers
It is similar, but the not necessarily phase of the excitation function in the neuron in each hidden layer
With.It is to say, these neurons can be set to tool according to reality application or needs
There is identical excitation function, or be set to that there is different excitation functions.
By operations described above, initial neural network topology structure can be formed, such as figure
Shown in 5C, wherein there are two hidden layers.Additionally, can also be apparent from from this Fig. 5 C,
Entirely do not connect between all neurons of each hidden layer, but only hidden at each
Containing layer performs between the neuron of the identical output factor full connection.According to the disclosure
Method in, decrease the unnecessary connection between hidden layer by association relation model,
Therefore this will accelerate the training of neutral net, does not interferes with precision of prediction simultaneously.
Then in step S203, utilize sample data that described neutral net is trained.Instruction
Practice method and can use any method of the prior art, the most more common supervised learning side
Method, its using the corresponding output factor in sample data as expected value output valve, and by this phase
Hope that output is compared to adjust weight as real output value, until real output value and phase
Hope the difference minimum of output valve or less than predetermined threshold value.In the training process, can be right
In each sample data, the topological structure of neutral net is made all to keep constant.That is, do not change
The number of the neuron in the initial neutral net set of change and annexation thereof.The most permissible
Further eliminate and make because changing annexation or make some neuron different operatings at random
The precision of prediction become reduces problem.
According to the disclosure, it is provided that a kind of method being used for building neutral net of improvement, wherein
Neutral net is initially set by excavating the association relation model between inputoutput data
Network topology.Utilize the method can be in the case of not affecting model prediction accuracy, significantly
Save the training time of network.
Additionally, in the disclosure, additionally provide a kind of equipment for constructing neural network,
It is described in greater detail below with reference to Fig. 6.
As shown in Figure 6, equipment 600 can include that model acquisition module 610, topology set mould
Block 620 and network training module 630.This model acquisition module 610 can be configured to obtain
Take the association relation model between target data and its influence factor, described association relation model
Characterize the relatedness between described target data and its influence factor.This topology setting module 620
Can be configured to according to described association relation model, the network setting described neutral net is opened up
Flutter.Network training module 630 can be configured to, with sample data to described neutral net
It is trained.
In an embodiment according to the disclosure, described topology setting module 620 can wrap
Factoring and connection determine module 621.This factor and connection determine that module 622 can be configured
By the association relation model provided based on model acquisition module 619, determine described nerve net
The input factor on road and export the factor and determine input layer and output layer with in hidden layer
The connection of neuron.Especially, the described factor and connection determine that module is further configured to
Determine each input factor weight for each output factor of neutral net;And according to
Described weight, determines the connection of the described input factor and the output factor and the neuron of hidden layer.
In an embodiment according to the disclosure, described topology setting module 620 includes hidden
Containing layer constructing module 624.Described hidden layer constructing module 624 can be configured to: for
One hidden layer, adds the same number of neuron of number and the output factor, described neuron
Connect with the corresponding output factor respectively, and according to described association relation model, will add
Described neuron be connected with the described input factor;And carry out for the neuron added
Replicate, to produce the god setting number for each output factor at described first hidden layer
Through unit.Further, described hidden layer constructing module 624 can be configured to for other
Hidden layer: add with in the upper hidden layer of other hidden layers described for each output because of
The identical neuron of neuron of son, and on described in a hidden layer and described other
Entirely connecting between the neuron of the identical output factor in hidden layer.
In an embodiment according to the disclosure, described topology setting module 620 is all right
Including neuron number setting module 626.This neuron number setting module can be as previously mentioned
The number of the neuron during rule of thumb formula sets each hidden layer, it is also possible to be configured
According to described association relation model set the number of neuron of each hidden layer.Especially
Ground, described neuron number setting module 626 is configured to be come by operations described below
Determine the number of the neuron of each hidden layer: determine the weight being associated with each output factor
Account for the ratio of all weight sums;And based on a determination that described ratio, determine each implicit
The number of neuron relevant to each output factor in Ceng.Specifically, described neuron number
Mesh setting module can be configured to be determined the number of described neuron by following formula
Mesh:
Wherein kjIndicate the number of the neuron being associated with the jth output factor;mjInstruction and jth
The number of the input factor that the individual output factor is associated;And μjInstruction with jth output because of
The weight that son is associated accounts for the ratio of all weight sums;α is making a reservation between 1 to 10
Constant.
According in another embodiment of the disclosure, described network training module 630 is further
It is configured to the situation of structure of the network topology of described neutral net set by maintain
Under, described neutral net is trained.I.e., in the training process, change initially
The number of the neuron in the neutral net set and annexation thereof.So can be further
Eliminate because of change annexation or make some neuron different operatings at random and cause pre-
Survey precision and reduce problem.
In superincumbent description, the disclosure is entered by the embodiment of Primary Reference Air Quality Forecast
Go description;But it should be noted that the disclosure is also possible to apply predicts field at other
Close.Additionally, in superincumbent description, with reference to Fig. 3 with Fig. 4 to using data and associating pass
It is that model is described, but the disclosure is not limited to that, associates in actual applications
Relational model is likely more complexity.It is described above for extracting association relation model
Use data and for training the sample data of neutral net;It should be noted that these two groups
Data can be identical data or the identical data of part, it is also possible to is diverse number
According to, extract and network training purpose as long as they can be used in incidence relation.Above
The description of the extraction of association relation model with reference to ad hoc fashion, but it should be noted that
It is also possible to use any other for extracting the mode of association relation model to incidence relation mould
Type extracts.It addition, determine the neuron number in hidden layer based on association relation model
It is example that purpose mode is also not limited to be given above, and actually can also use any other
Suitable mode.
Additionally it should also be understood that embodiment of the present disclosure can with software, hardware or
Being implemented in combination in of software and hardware.Hardware components can utilize special logic to realize;Soft
Part part can store in memory, by suitable instruction execution system, the most micro-process
Device or special designs hardware perform.It will be understood by those skilled in the art that above-mentioned
Method and apparatus can use computer executable instructions and/or be included in processor control generation
Realize in Ma, such as at the mounting medium, all of such as disk, CD or DVD-ROM
Programmable memory or such as optics or electronic signal such as read only memory (firmware)
Such code is provided in the data medium of carrier.Equipment and the assembly thereof of the present embodiment can
With by such as super large-scale integration or gate array, such as logic chip, transistor etc.
Quasiconductor or such as field programmable gate array, programmable logic device etc. able to programme
The hardware circuit of hardware device realizes, it is also possible to soft with performed by various types of processors
Part realizes, it is also possible to realized by the combination such as firmware of above-mentioned hardware circuit and software.
Although describe the disclosure by reference to the embodiment being presently contemplated that, but should
Understand that the disclosure is not limited to disclosed embodiment.On the contrary, the disclosure is intended to appended
Various amendments included in spirit and scope by the claims and equivalent arrangements.Following right
The scope required meets broad interpretation, in order to comprise all such amendments and equivalent structure
And function.
Claims (20)
1. for the method building neutral net, including:
Obtain the association relation model between target data and its influence factor, described incidence relation
Model characterizes the relatedness between described target data and its influence factor;
According to described association relation model, set the network topology of described neutral net;And
Utilize sample data that described neutral net is trained.
Method the most according to claim 1, wherein according to described association relation model,
The network topology setting described neutral net includes:
Based on described association relation model, determine the input factor of described neural network and output because of
Son and determine described neutral net the input factor and output the factor and hidden layer in nerve
The connection of unit.
Method the most according to claim 2, wherein determines the described input factor and described
The output factor includes with the connection of the neuron in hidden layer:
Determine each input factor weight for each output factor of described neutral net;With
And
According to described weight, determine the described input factor and the described output factor and described hidden layer
The connection of neuron.
Method the most according to claim 2, wherein according to described association relation model,
The network topology setting described neutral net includes: for the first hidden layer,
Adding the same number of neuron of number and the described output factor, described neuron is respectively
Connect with the corresponding output factor, and according to described association relation model, the institute that will add
State neuron to be connected with the described input factor;And
Replicate for the neuron added, to produce for each at described first hidden layer
The neuron setting number of the individual output factor.
Method the most according to claim 4, farther includes: for other hidden layers,
Add with in the upper hidden layer of other hidden layers described for each output factor
The neuron that neuron is identical, and
On described in a hidden layer with in other hidden layers described for identical output because of
Entirely connect between the neuron of son.
Method the most according to claim 1, wherein according to described association relation model,
The network topology setting described neutral net also includes:
The number of the neuron of each hidden layer is set according to described association relation model.
Method the most according to claim 6, wherein comes according to described association relation model
The number of the neuron setting each hidden layer includes:
Determine that the weight being associated with each output factor accounts for the ratio of all weight sums;And
Based on a determination that described ratio, determine in each hidden layer to each output factor relevant
The number of neuron.
Method the most according to claim 7, wherein based on a determination that described ratio, really
The number of neuron relevant to each output factor in fixed each hidden layer includes: under by
Formula determines the number of described neuron:
Wherein kjIndicate the number of the neuron being associated with the jth output factor;mjInstruction and jth
The number of the input factor that the individual output factor is associated;And μjInstruction with jth output because of
The weight that son is associated accounts for the ratio of all weight sums;α is making a reservation between 1 to 10
Constant.
Method the most according to claim 1, wherein obtain target data and affect because of
Association relation model between element includes: use Granger carsal graph model, extracts and at least wraps
Between data element containing target data and the polynary time series data of possible influence factor thereof
Incidence relation.
Method the most according to claim 1, wherein at the described nerve set by maintenance
In the case of the structure of network of network topology, described neutral net is trained.
11. 1 kinds for building the equipment of neutral net: including:
Model acquisition module, is configured to obtain associating between target data and its influence factor
Relational model, described association relation model characterizes between described target data and its influence factor
Relatedness;
Topology setting module, is configured to, according to described association relation model, set described nerve
Network of network topology;And
Network training module, is configured to, with sample data and instructs described neutral net
Practice.
12. equipment according to claim 11, wherein said topology setting module includes
The factor and connection determine that module, the described factor and connection determine that module is configured to:
Based on described association relation model, determine the input factor of described neural network and output because of
Son and determine the connection of neuron in the described input factor and the output factor and hidden layer.
13. equipment according to claim 12, the wherein said factor and connection determine mould
Block is further configured to:
Determine each input factor weight for each output factor of described neutral net;With
And
According to described weight, determine the described input factor and the described output factor and described hidden layer
The connection of neuron.
14. equipment according to claim 12, wherein said topology setting module also wraps
Including hidden layer constructing module, described hidden layer constructing module is configured to: imply for first
Layer,
Adding the same number of neuron of number and the described output factor, described neuron is respectively
Connect with the corresponding output factor, and according to described association relation model, the institute that will add
State neuron to be connected with the described input factor;And
Replicate for the neuron added, to produce for each at described first hidden layer
The neuron setting number of the individual output factor.
15. equipment according to claim 14, wherein said hidden layer constructing module enters
One step is configured to: for other hidden layers,
Add with in the upper hidden layer of other hidden layers described for each output factor
The neuron that neuron is identical, and
On described in a hidden layer with in other hidden layers described for identical output because of
Entirely connect between the neuron of son.
16. equipment according to claim 11, wherein said topology setting module also wraps
Including neuron number setting module, described neuron number setting module is configured to according to institute
State association relation model to set the number of the neuron of each hidden layer.
17. equipment according to claim 16, wherein said neuron number sets mould
Block is configured to:
Determine that the weight being associated with each output factor accounts for the ratio of all weight sums;And
Based on a determination that described ratio, determine in each hidden layer to each output factor relevant
The number of neuron.
18. equipment according to claim 17, wherein said neuron number sets mould
Block is configured to be determined the number of described neuron by following formula:
Wherein kjIndicate the number of the neuron being associated with the jth output factor;mjInstruction and jth
The number of the input factor that the individual output factor is associated;And μjInstruction with jth output because of
The weight that son is associated accounts for the ratio of all weight sums;α is making a reservation between 1 to 10
Constant.
19. equipment according to claim 11, wherein said Relation acquisition module is joined
It is set to: use Granger carsal graph model, extracts including at least target data and possible shadow thereof
Incidence relation between the data element of the polynary time series data of the factor of sound.
20. equipment according to claim 11, wherein said network training module is joined
It is set in the case of the structure of the network topology of the described neutral net set by maintaining, right
Described neutral net is trained.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201510129042.XA CN106156845A (en) | 2015-03-23 | 2015-03-23 | A kind of method and apparatus for building neutral net |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201510129042.XA CN106156845A (en) | 2015-03-23 | 2015-03-23 | A kind of method and apparatus for building neutral net |
Publications (1)
Publication Number | Publication Date |
---|---|
CN106156845A true CN106156845A (en) | 2016-11-23 |
Family
ID=58063491
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201510129042.XA Pending CN106156845A (en) | 2015-03-23 | 2015-03-23 | A kind of method and apparatus for building neutral net |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN106156845A (en) |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109635706A (en) * | 2018-12-04 | 2019-04-16 | 武汉灏存科技有限公司 | Gesture identification method, equipment, storage medium and device neural network based |
CN110462637A (en) * | 2017-03-24 | 2019-11-15 | 华为技术有限公司 | Neural Network Data processing unit and method |
WO2024046462A1 (en) * | 2022-09-02 | 2024-03-07 | 深圳忆海原识科技有限公司 | Port model object calling method and system, platform, intelligent device, and storage medium |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2002063555A2 (en) * | 2001-02-07 | 2002-08-15 | Worldcom, Inc. | An artificial intelligence trending system |
CN101226056A (en) * | 2007-06-01 | 2008-07-23 | 清华大学 | Car body obliqueness detection voice alarm device based on CAN bus |
US20090281974A1 (en) * | 2008-04-08 | 2009-11-12 | Infosys Technologies Limited | System and method for adaptive data masking |
CN101682195A (en) * | 2007-03-26 | 2010-03-24 | Vpec株式会社 | Power system |
CN102075352A (en) * | 2010-12-17 | 2011-05-25 | 北京邮电大学 | Method and device for predicting network user behavior |
CN102098530A (en) * | 2010-12-02 | 2011-06-15 | 惠州Tcl移动通信有限公司 | Method and device for automatically distinguishing quality of camera module |
-
2015
- 2015-03-23 CN CN201510129042.XA patent/CN106156845A/en active Pending
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2002063555A2 (en) * | 2001-02-07 | 2002-08-15 | Worldcom, Inc. | An artificial intelligence trending system |
CN101682195A (en) * | 2007-03-26 | 2010-03-24 | Vpec株式会社 | Power system |
CN101226056A (en) * | 2007-06-01 | 2008-07-23 | 清华大学 | Car body obliqueness detection voice alarm device based on CAN bus |
US20090281974A1 (en) * | 2008-04-08 | 2009-11-12 | Infosys Technologies Limited | System and method for adaptive data masking |
CN102098530A (en) * | 2010-12-02 | 2011-06-15 | 惠州Tcl移动通信有限公司 | Method and device for automatically distinguishing quality of camera module |
CN102075352A (en) * | 2010-12-17 | 2011-05-25 | 北京邮电大学 | Method and device for predicting network user behavior |
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110462637A (en) * | 2017-03-24 | 2019-11-15 | 华为技术有限公司 | Neural Network Data processing unit and method |
CN109635706A (en) * | 2018-12-04 | 2019-04-16 | 武汉灏存科技有限公司 | Gesture identification method, equipment, storage medium and device neural network based |
CN109635706B (en) * | 2018-12-04 | 2020-09-01 | 武汉灏存科技有限公司 | Gesture recognition method, device, storage medium and device based on neural network |
WO2024046462A1 (en) * | 2022-09-02 | 2024-03-07 | 深圳忆海原识科技有限公司 | Port model object calling method and system, platform, intelligent device, and storage medium |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN104166731B (en) | A kind of overlapping community discovery system and method for social networks | |
CN102606557A (en) | Health evaluation method for hydraulic system based on fault observer and SOM (self-organized mapping) | |
CN107977748B (en) | Multivariable distorted time sequence prediction method | |
CN110223785A (en) | A kind of infectious disease transmission network reconstruction method based on deep learning | |
CN105184368A (en) | Distributed extreme learning machine optimization integrated framework system and method | |
CN109886554A (en) | Unlawful practice method of discrimination, device, computer equipment and storage medium | |
CN109558893A (en) | Fast integration sewage treatment method for diagnosing faults based on resampling pond | |
CN107025341A (en) | A kind of photovoltaic DC-to-AC converter method for diagnosing faults | |
CN106709588A (en) | Prediction model construction method and equipment and real-time prediction method and equipment | |
CN108879732A (en) | Transient stability evaluation in power system method and device | |
CN106156845A (en) | A kind of method and apparatus for building neutral net | |
CN112017070A (en) | Method and system for evaluating transient stability of power system based on data enhancement | |
CN105335626B (en) | A kind of group lasso characteristic grouping methods of Excavation Cluster Based on Network Analysis | |
CN112580807A (en) | Neural network improvement demand automatic generation method and device based on efficiency evaluation | |
CN107016440A (en) | The multiresolution deep neural network intelligent diagnosing method of machine driving failure | |
CN115409262A (en) | Railway data center key performance index trend prediction method and abnormity identification method | |
KR20160132850A (en) | Contextual real-time feedback for neuromorphic model development | |
Gorbachev et al. | A synergistic effect in the measurement of neuro-fuzzy system | |
CN110796243B (en) | Continuous operation monitoring data simulation generation method and device | |
Chen | Hotel management evaluation index system based on data mining and deep neural network | |
CN111695989A (en) | Modeling method and platform of wind-control credit model | |
CN107679478A (en) | The extracting method and system of transmission line of electricity space load state | |
Meng et al. | Nonlinear system simulation based on the BP neural network | |
De Aguiar et al. | Using reservoir computing for forecasting of wind power generated by a wind farm | |
CN113326182B (en) | Software defect prediction method based on sampling and ensemble learning |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
C06 | Publication | ||
PB01 | Publication | ||
C10 | Entry into substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
WD01 | Invention patent application deemed withdrawn after publication | ||
WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20161123 |