CN108171329A - Deep learning neural network training method, number of plies adjusting apparatus and robot system - Google Patents
Deep learning neural network training method, number of plies adjusting apparatus and robot system Download PDFInfo
- Publication number
- CN108171329A CN108171329A CN201711330260.5A CN201711330260A CN108171329A CN 108171329 A CN108171329 A CN 108171329A CN 201711330260 A CN201711330260 A CN 201711330260A CN 108171329 A CN108171329 A CN 108171329A
- Authority
- CN
- China
- Prior art keywords
- neural network
- deep learning
- learning neural
- training
- input data
- 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
- 238000013528 artificial neural network Methods 0.000 title claims abstract description 189
- 238000012549 training Methods 0.000 title claims abstract description 184
- 238000013135 deep learning Methods 0.000 title claims abstract description 167
- 238000000034 method Methods 0.000 title claims abstract description 38
- 241001269238 Data Species 0.000 claims description 12
- 238000004590 computer program Methods 0.000 claims description 11
- 210000005036 nerve Anatomy 0.000 claims description 5
- 230000003466 anti-cipated effect Effects 0.000 claims description 4
- 238000004364 calculation method Methods 0.000 claims 1
- 230000008569 process Effects 0.000 abstract description 7
- 239000010410 layer Substances 0.000 description 150
- 210000004218 nerve net Anatomy 0.000 description 8
- 230000001815 facial effect Effects 0.000 description 7
- 238000012360 testing method Methods 0.000 description 6
- 238000004458 analytical method Methods 0.000 description 5
- 230000019771 cognition Effects 0.000 description 5
- 230000001537 neural effect Effects 0.000 description 5
- 230000004069 differentiation Effects 0.000 description 4
- 230000000694 effects Effects 0.000 description 4
- 230000008859 change Effects 0.000 description 3
- 238000003780 insertion Methods 0.000 description 2
- 230000037431 insertion Effects 0.000 description 2
- 239000011229 interlayer Substances 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000009471 action Effects 0.000 description 1
- 210000004556 brain Anatomy 0.000 description 1
- 239000003086 colorant Substances 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 235000013399 edible fruits Nutrition 0.000 description 1
- 230000006870 function Effects 0.000 description 1
- 238000011478 gradient descent method Methods 0.000 description 1
- 230000010365 information processing Effects 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 230000009467 reduction Effects 0.000 description 1
- 239000002356 single layer Substances 0.000 description 1
- 230000008667 sleep stage Effects 0.000 description 1
- 238000012706 support-vector machine Methods 0.000 description 1
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
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Image Analysis (AREA)
Abstract
The present invention relates to a kind of deep learning neural network training method, number of plies adjusting apparatus and robot systems.This method includes：Obtain training input data；Training input data is inputted current depth learning neural network to be trained to obtain training result information, current depth learning neural network includes input layer, hidden layer, grader and output layer；If training result information does not meet preset condition, the number of plies of hidden layer in current depth learning neural network is then adjusted to update current depth learning neural network, and it returns and training input data is inputted into the step of current depth learning neural network is trained, obtains training result information；Otherwise using current depth learning neural network as the deep learning neural network after training.In this way, the adjustment by carrying out hidden layers numbers according to training result information in the training process, until training result information meets preset condition, it can be ensured that deep learning neural network is trained successfully, and training success rate is high.
Description
Technical field
The present invention relates to machine learning techniques field, more particularly to a kind of deep learning neural network training method, layer
Number adjusting apparatus and robot system.
Background technology
Neural network is a kind of algorithm mathematics model for carrying out distributed parallel information processing, is usually used in intelligence machine knowledge
Not, deep learning neural network can be used sample data and carry out deep learning.Wherein, deep learning includes supervised learning and nothing
Supervised learning.
In traditional technology, the deep learning neural network progress that the fixed number of plies is typically inputted using sample data is unsupervised
Study obtains output label using the deep learning progress supervised learning of the fixed number of plies of sample data input with label,
To complete the training to deep learning neural network, and passed through after training using test data test.For example, output label is
ID card No. may be used head portrait and obtain the ID card No. of the people or using voice as sample data, input head portrait
The ID card No. of the people is obtained as sample data, input voice.
However, during supervised learning, if the number of plies of deep learning neural network is very few so that top layer concept
It is excessively specific, with unrelated or even contradiction the details of output label differentiation so that top layer concept cannot be abundant with output label
Fitting, that can cause sample data much to cannot get correct output label, the failure of deep learning neural metwork training after inputting；
And if the number of plies of deep learning neural network is excessive, it can so that top layer concept is excessively abstracted, lacks the details of differentiation, and having
" noise " details of output label can just be distinguished by being added in during supervised learning, so that top layer concept and output label mistake
In fitting, that can cause sample data much cannot after all obtaining correct output label but test data input after inputting
Correct output label fails so as to cause deep learning neural metwork training.
Invention content
Based on this, it is necessary to which, for traditional low success rate of problem of deep learning neural metwork training, providing one kind can
Improve deep learning neural network training method, number of plies adjusting apparatus and the machine of the success rate of training deep learning neural network
People's system.
A kind of deep learning neural network training method, including：
Obtain training input data；
The trained input data input current deep learning neural network is trained, obtains training result information,
The current deep learning neural network includes input layer, hidden layer, grader and output layer；
If the training result information does not meet preset condition, hidden layer in the current deep learning neural network is adjusted
The number of plies to update current deep learning neural network, and return described by the trained input data input current depth study
Neural network is trained, the step of obtaining training result information；
If the training result information meets the preset condition, using the current deep learning neural network as instruction
Deep learning neural network after white silk.
A kind of number of plies adjusting apparatus, including：
Data acquisition module, for obtaining trained input data；
Network training module, for the trained input data input current deep learning neural network to be trained,
Training result information is obtained, the current deep learning neural network includes input layer, hidden layer, grader and output layer；
The number of plies adjusts module, for when the training result information does not meet preset condition, adjusting the current depth
The number of plies of hidden layer is to update current deep learning neural network in learning neural network, and returns to the network training module again
The trained input data input current deep learning neural network is trained, obtains training result information；
Training terminate module, for when the training result information meets the preset condition, by the current depth
Learning neural network is as the deep learning neural network after training.
Training input data is inputted current depth by above-mentioned deep learning neural network training method and number of plies adjusting apparatus
Learning neural network is trained to obtain training result information, if training result information does not meet the successful preset condition of training,
The number of plies of hidden layer in current deep learning neural network is then adjusted to update current deep learning neural network, and returns and again will
Training input data inputs updated current deep learning neural network and is trained, and whether meets default item with replicate analysis
Part；If training result information meets preset condition, expression train successfully, using current deep learning neural network as after trained
Deep learning neural network.In this way, the adjustment by carrying out hidden layers numbers according to training result information in the training process, until
Training result information meets preset condition, it can be ensured that deep learning neural network is trained successfully.It is deep compared to traditional training
Spend the training success rate of learning neural network mode, the application deep learning neural network training method and number of plies adjusting apparatus
It is high.
A kind of medium, is stored with computer program, and above-mentioned depth is realized when the computer program of storage is executed by processor
The step of learning neural network training method.
A kind of robot system can be run on a memory and on a processor including memory, processor and storage
Computer program, the processor realize the step of above-mentioned deep learning neural network training method when performing the computer program
Suddenly.
Above-mentioned medium and robot system, similarly can be true due to realizing above-mentioned deep learning neural network training method
Training successfully for deep learning neural network is protected, training success rate is high.
Description of the drawings
Fig. 1 is the flow chart of deep learning neural network training method in an embodiment；
Fig. 2 is the flow chart of deep learning neural network training method in another embodiment；
Fig. 3 is the illustraton of model of the deep learning neural network containing multiple hidden layers；
Fig. 4 is the core thinking schematic diagram of deep learning；
Fig. 5 is that there are one the depth of input layer, a grader, a hidden layer and an output layer for tool in an embodiment
Practise the illustraton of model of neural network；
Fig. 6 is the structure chart of number of plies adjusting apparatus in an embodiment.
Specific embodiment
During carrying out supervised learning to deep learning neural network, if the number of plies is excessive, then can so that top layer is general
The details for being excessively abstracted, lacking differentiation is read, and " making an uproar for output label can just be distinguished by being added during supervised learning
Sound " details so that top layer concept is excessively fitted with output label, adds in training in order to be fitted in top layer concept
The noise of non-distinctive feature, necessarily leads to that error rate can be increased in the test below.Such as output label includes " white
Man ", " white woman ", " black man ", " black woman ", and top layer conceptual abstraction has arrived " man ", " woman " feature, suddenly
Omit " black ", " white " feature, sample datas all at this time can be fitted to " man ", on " woman " this 2 labels, because
It is deep learning using unsupervised learning from bottom to top and top-down supervised learning, from bottom to top without prison
During educational inspector practises, this few class sample data of " white man ", " black man ", it is clear that can be corresponding with " man " top layer concept, it is " white
Color woman ", " black woman " this few class sample data, it is clear that can be corresponding with " woman " top layer concept；Have top-down
In supervised learning, " white man ", " black man " this several class, it is clear that can be corresponding with " man " top layer concept, " white female
People ", " black woman " this several class, it is clear that can be corresponding with " woman " top layer concept；Pass through supervised learning, deep learning nerve
Network meeting adjust automatically network weight, eventually so that " man+noise 1 " top layer concept corresponds to " black man ", " man+make an uproar
2 " top layer concept of sound corresponds to " white man ", and " woman+noise 3 " top layer concept corresponds to " black woman ", " woman+noise 4 " top
Layer concept corresponds to " white woman ".Due to being fitted repeatedly during supervised learning, so as to for training data
It is to have achieved the effect that abundant fitting.But in use, during input " white man " test data, obtained top layer concept is " male
People+noise 2 ", but because noise 2 is not to discriminate between the distinctive feature of white man and black man, pass through grader at this time
The output label corresponded to may be " white man ", it is also possible to " black man " or even can be obtained due to the interference of noise 2
Output label may be " white woman " or " black woman ", so as to will cause test when error rate increase, failure to train.
If the number of plies is very few, then can cause top layer concept it is excessively specific, with the unrelated or even contradiction of output label differentiation
Details, so that top layer concept cannot be fitted with output label, that is to say, that can not cause top layer concept and output label one
One is mapped.Such as output label include " man ", " woman ", and the feature of top layer concept in addition to include distinguish men and women must
It wants outside feature, further includes hair feature, features of skin colors.Because deep learning is using unsupervised learning from bottom to top and certainly
Downward supervised learning is pushed up, in top-down supervised learning, if just beginning with very much " bob man ", " black
Man ", " white woman ", " long hair woman " training sample, then will cause in top layer concept formed " bob man ",
" black man ", " white woman ", the concept of " long hair woman ", wherein " bob man ", " black man " pass through grader pair
" man " output label is answered, " white woman ", " long hair woman " correspond to " woman " output label by grader, but if later
There are the training sample of a large amount of " long hair men ", " black woman " again, top layer concept will be caused to be adjusted to " long hair man ", " black
Color man ", " black woman ", " long hair woman " concept, but adjust after obviously again can not be fitted " bob man ", " white
Those sample datas of woman ", so as to cause deep learning neural network with the constant adjustment network weight of the variation of sample data
Value, but can not fully be fitted, failure to train always.
The application provides a kind of deep learning neural network training method, number of plies adjusting apparatus, medium and robot system.
With reference to figure 1, in one embodiment, deep learning neural network training method includes the following steps：
S110：Obtain training input data.
Training input data is the sample data being trained for deep learning neural network.For example, training input number
According to can be facial image or voice.Specifically, training input data can be obtained by acquiring, for example, passing through
Camera acquisition, which obtains facial image or acquired by voice receiver, obtains voice；Training input data can also be from number
It is obtained according to library lookup, for example, storing up facial image or voice in databases in advance, facial image or language is searched for from database
Sound can obtain training input data.
S130：Training input data is inputted current deep learning neural network to be trained, obtains training result information.
Current deep learning neural network is the deep learning neural network created, including input layer, hidden layer, is divided
Class device and output layer.Wherein, grader includes Rogers spy's recurrence grader, SVM (Support Vector Machine supports
Any one of vector machine) grader.Specifically, when performing step S130 for the first time, corresponding current depth learns nerve net
Network is the deep learning neural network for being pre-created storage.
S150：If training result information does not meet preset condition, hidden layer in current deep learning neural network is adjusted
The number of plies is to update current deep learning neural network, and return to step S130.
Preset condition is to represent to train successful condition, can be set according to actual demand.It can be by training result
Information carries out analyzing and determining whether training result information meets preset condition.Training result information does not meet preset condition, represents
It does not train also successfully, top layer concept and the output label of current deep learning neural network cannot be fitted, at this point, by adjusting hidden
The number of plies of layer is to update current deep learning neural network, to increase top layer concept and the degree of output label fitting.It adjusts hidden
The number of plies of layer can increase new hidden layer or reduction hidden layer.
Specifically, training input data is inputted updated current depth study nerve by return to step S130 again
Network is trained, and is obtained updated current deep learning neural network and is corresponded to the training result information that training obtains.For example,
After increasing hidden layer, training data is inputted into the current deep learning neural network after increasing hidden layer again, obtains corresponding training knot
Fruit information.
S170：If training result information meets preset condition, using current deep learning neural network as training after
Deep learning neural network.
Training result information meets preset condition, and expression is trained successfully, and current deep learning neural network can reach top layer
Concept is fitted with label；At this point, terminate training, using current deep learning neural network as the deep learning nerve net after training
Network.Specifically, if the training result information for meeting preset condition is the corresponding instruction of updated current deep learning neural network
Practice result information, then using the updated current deep learning neural network of correspondence as the deep learning neural network after training.
Above-mentioned deep learning neural network training method, will training input data input current deep learning neural network into
Row training obtains training result information, if training result information does not meet the successful preset condition of training, adjusts current depth
The number of plies of hidden layer is to update current deep learning neural network in learning neural network, and returning will train input data defeated again
Enter updated current deep learning neural network to be trained, whether preset condition is met with replicate analysis；If training result
Information meets preset condition, and expression is trained successfully, using current deep learning neural network as the deep learning nerve after training
Network.In this way, the adjustment by carrying out hidden layers numbers according to training result information in the training process, until training result information
Meet preset condition, it can be ensured that deep learning neural network is trained successfully.Compared to traditional training deep learning nerve net
Network mode, the training success rate of the application deep learning neural network training method are high.
In one embodiment, training input data includes no label input data, tape label input data and each band mark
Sign the corresponding expected label of input data.
No label input data is not carry the data of label, and tape label input data is to carry the data of label.
Each tape label input data corresponds to an expected label, and tape label input data and expected label can correspond to storage in advance.
Wherein, label and expected label are a kind of identification informations, for example, be single facial image without label input data, tape label
Input data is to carry the facial image of ID card No., it is contemplated that label is right by the facial image for carrying ID card No.
The real identity card number of the user answered.
It please refers to Fig.2, step S130 includes step S131 to step S135.
S131：Using no label input data as the input data of current deep learning neural network, to current depth
It practises neural network and carries out unsupervised learning, obtain the initial neural network of parameter.
Wherein, the quantity of no label input data can have multiple；It is successively that each no label input data input is current deep
Learning neural network is spent, unsupervised learning is taken turns in corresponding completion one, then the number of the corresponding unsupervised learning of a wheel unsupervised learning
Equal to the quantity of no label input data.It is appreciated that or choose present count from all no label input datas
Amount inputs current deep learning neural network to complete a wheel unsupervised learning without label input data, other are unselected
It does not use；At this point, quantity of the number of the corresponding unsupervised learning of a wheel unsupervised learning less than no label input data, specifically
To be equal to preset quantity.
S133：Tape label input data and the corresponding expected label of tape label input data is initially refreshing as parameter
Input data and anticipated output through network, to parameter, initial neural network carries out supervised learning, the correspondence band exported
The physical tags of label input data.
Wherein, the quantity of tape label input data can have multiple；It is successively that each tape label input data and tape label is defeated
Enter input data and anticipated output of the corresponding expected label of data as the initial neural network of parameter, a corresponding wheel of completing has prison
Educational inspector practises, then the number of the corresponding supervised learning of a wheel supervised learning is equal to the quantity of tape label input data.It can manage
The tape label input data input parameter for solving or preset quantity being chosen from all tape label input datas is initially refreshing
Through network to complete a wheel supervised learning, other are unselected not to use；At this point, supervised learning is corresponding prison for a wheel
The number that educational inspector practises is less than the quantity of tape label input data, specially equal to preset quantity.
S135：The number of statistics supervised learning obtains training total degree, and the quantity of statistics tape label input data obtains
Total sample number believes training total degree, total sample number and the corresponding physical tags of each tape label input data as training result
Breath.
Total sample number is total quantity of the tape label input data obtained；Training total degree is statistics supervised learning
Number obtains, wherein, it is corresponding to carry out once thering is supervision to learn often using a tape label input data and corresponding expected label
It practises.If carrying out a wheel supervised learning needs using all tape label input datas, a wheel supervised learning is completed, it is corresponding
The number of supervised learning is equal to total sample number, that is, total degree is trained to be equal to total sample number；Two-wheeled supervised learning is completed, it is corresponding
The number of supervised learning is equal to 2 times of total sample number, that is, trains total degree equal to 2 times of total sample number, and so on.It can be with
Understand, if carrying out the part tape label input data that a wheel supervised learning uses all tape label input datas, complete one
Supervised learning is taken turns, the number of corresponding supervised learning is less than total sample number, that is, total degree is trained to be less than total sample number, according to this class
It pushes away.
Unsupervised learning and supervised learning are the modes of learning of deep learning.Worked as by that will be inputted without label input data
On the basis of preceding deep learning neural network progress unsupervised learning obtains the initial neural network of parameter, inputted further according to tape label
Initial neural network carries out supervised learning to parameter for data and the corresponding expected label of tape label input data, with reference to unsupervised
Study and supervised learning are trained, and closer to global optimum, training effect is good.
With reference to figure 3 and Fig. 4, learning structure is regarded as a network, the core thinking of deep learning is as follows：
The first step：Using unsupervised learning from bottom to top
1st, successively structure monolayer neuronal is first.
2nd, tuning is carried out using wake-sleep algorithms every layer.One layer is only adjusted every time, is successively adjusted.
This process can be regarded as the process of a feature learning, be to distinguish maximum with traditional neural network
Part.
Wake-sleep algorithms include:
1st, the wake stages：Cognitive process is weighed by the input feature vector (Input) of lower floor and upward cognition (Encoder)
Each layer of abstract representation (Code) is generated again, then a reconstruction information is generated by current generation weight (Decoder)
(Reconstruction), input feature vector and reconstruction information residual error are calculated, the downlink for declining modification interlayer using gradient generates power
Weight.Namely " if reality imagines different with me, change my generation weight so that the thing that I imagines becomes and reality
Equally ".
2nd, the sleep stages：Generating process by Upper Concept and downward generation weight, generates the state of lower floor, then profit
An abstract scene is generated with cognition weight.Using initial upper layer concept and the residual error of newly-built abstract scene, declined using gradient
Change the upward cognition weight of interlayer.Namely " if the scene in dream is not the corresponding concepts in my brain, change my cognition
Weight so that this scene is exactly this concept in my view ".
Specifically, first train first layer with no nominal data, when training first learns the parameter of first layer, and (this layer can be seen
Work is to obtain one so that the hidden layer of the three-layer neural network of output data and input data difference minimum), due to model
The limitation of capacity and sparsity constraints so that obtained model can learn the structure to data in itself, so as to obtain
Than the feature that input data has more expression ability；After study obtains (n-1)th layer, using n-1 layers of output data as n-th layer
Input data, thus training n-th layer respectively obtains the parameter of each layer.
Second step：Top-down supervised learning
This step is to obtain each layer parameter on the basis of in first step study, and a classification is added in the coding layer most pushed up
Device then by the supervised learning of tape label input data, goes to finely tune entire network parameter using gradient descent method.
The first step of deep learning is substantially a network parameter initialization procedure.It is different from traditional neural network initial value
Random initializtion, deep learning neural network be by being obtained in the structure of unsupervised learning input data, thus this
Initial value is closer to global optimum, so as to obtain better effect.
Specifically, in the present embodiment, preset condition includes：Training total degree is less than or equal to default times of total sample number
Number, and the average value of the residual error of the expection label and physical tags of all tape label input datas is less than or equal to default residual error threshold
Value.
Preset condition is represents the successful condition of training.If training total degree is less than or equal to default times of total sample number
Number, and the average value of the residual error of the expection label and physical tags of all tape label input datas is less than or equal to default residual error threshold
Value, then training result information meets preset condition, and top layer concept is fully fitted with output label and is not fitted excessively at this time；It is no
Then, preset condition is not met, represents that top layer concept is not fitted successfully with output label.By according to training total degree, sample
Sum and the average value of the residual error of expected label and physical tags judge whether to meet preset condition, can accurately analyse whether into
Work(is fitted.It is appreciated that in other embodiments, preset condition can also be other.
Wherein, preset multiple can be pre-set according to actual needs；Specifically, preset multiple can be more than 1, can also
It is less than 1 more than zero.For example, often carrying out a wheel supervised learning using in the embodiment of all tape label input datas, setting is pre-
If multiple is more than 1；It often carries out a wheel supervised learning and uses the part tape label input data in all tape label input datas
Embodiment in, can also set preset multiple be more than 0 be less than 1.Wherein, average value inputs number by calculating each tape label respectively
According to expection label and the residual errors of corresponding physical tags obtain multiple residual errors, calculate being averaged for multiple residual errors again and be worth to.
Specifically, expection label and the corresponding tape label input data last time input parameter for choosing each tape label input data are initial
The physical tags that neural network obtains calculate residual error.For example, the expection label of tape label input data A is z0, input for the first time
The physical tags of tape label input data A outputs are z1, and the physical tags that last time inputs tape label input data A outputs are
Z2, then the corresponding residual error of tape label input data is the residual error of z0 and z2.
In one embodiment, with reference to figure 2, step S130 further includes step S120 before.
S120：Creating tool, there are one input layer, a grader, a hidden layer and an output layers and top-down successively
The deep learning neural network of arrangement obtains current deep learning neural network.
To be used as initial neural network there are one the deep learning neural network of hidden layer by creating tool, with minimum number of layers
Based on hidden layer, subsequently to increase hidden layer.For example, with reference to figure 5, for the deep learning nerve net created before step S130
Network.Step S120 can be before step S110 or after step silo.In the present embodiment, step S120 exists
It is performed after step S110.
Specifically, in the present embodiment, in step S150, the number of plies of hidden layer in current deep learning neural network is adjusted with more
New current deep learning neural network is one hidden layer of insertion in current deep learning neural network, obtains new current depth
Spend learning neural network.In this way, step S150 is often performed once, the number of plies of hidden layer will increase in current deep learning neural network
Add one layer.Such as initial hidden layer only has 1 layer；After step S150 is performed 1 time, hidden layer becomes 2 layers；It is hidden after step S150 is performed 2 times
Layer becomes 3 layers；And so on.By being stepped up the number of plies of hidden layer, and the analytical judgment for passing through preset condition checks fitting
Situation, enabling the state being just fully fitted is progressivelyed reach, without because blindly increasing the number of plies leads to over-fitting.
Specifically, the number of plies of hidden layer in current deep learning neural network is adjusted to update current depth study nerve net
Network, including：The hidden layer being newly inserted into is created, by the output of the last one hidden layer in deep learning neural network and the hidden layer being newly inserted into
Input be connected by encoding and decoding network, using the hidden layer being newly inserted into output as grader input, to update current depth
Learning neural network.
The present embodiment is by initially only setting one layer of hidden layer, by carrying out unsupervised learning to hidden layer, then plus classification
The top-down carry out supervised learning of device, if training result information meets preset condition, is successfully fitted, and completion has supervision to instruct
Practice, if not meeting preset condition, be fitted failure, the connection between the previous hidden layer of cut-out grader and grader, and classifying
A new hidden layer is inserted between the previous hidden layer of device and grader layer, and the previous hidden layer of grader is connected to new hidden layer, is incited somebody to action
New hidden layer is connected to grader layer, then to repeating unsupervised learning and supervised learning, until being successfully fitted.In this way, energy
The previous hidden layer concept of grader is exactly the concept for being enough fully to be fitted with output label when reaching abundant fitting, can be just
Reach fully fitting without being excessively fitted.
Specifically, in the present embodiment, the number of nodes for the hidden layer being newly inserted into is less than or equal to what the hidden layer being newly inserted into was connected
The number of nodes of hidden layer.The hidden layer that the hidden layer being newly inserted into is connected is the last one hidden layer of link sort device before updating.It is logical
It crosses and the number of nodes for the hidden layer being newly inserted into is allowed to be less than or equal to the number of nodes of the last one hidden layer rather than more than the last one hidden layer
Number of nodes, can cause in this way input grader top layer concept it is more abstract, neglecting can not be reflected with output label
The feature penetrated, taking out can be with output label energy fully corresponding feature.
It is appreciated that in other embodiments, step S120 can also be the current depth for first establishing a multiple hidden layers
Learning neural network.If training result information does not meet preset condition, return to step after a hidden layer is reduced in step S150
S130。
In one embodiment, with continued reference to Fig. 2, after step S170, S180 is further included：By the deep learning after training
Neural network exports.Deep learning neural network after training is the deep learning neural network of successfully fitting, after output training
Deep learning neural network can be used for Machine self-learning.
With reference to figure 6, in one embodiment, a kind of number of plies adjusting apparatus is provided, including data acquisition module 110, net
Network training module 130, number of plies adjustment module 150 and training terminate module 170.
Data acquisition module 110 is used to obtain trained input data.
Network training module 130 obtains for input data input current deep learning neural network will to be trained to be trained
To training result information, current deep learning neural network includes input layer, hidden layer, grader and output layer.
Number of plies adjustment module 150 is used for when training result information does not meet preset condition, adjustment current depth study god
The number of plies through hidden layer in network returns to network training module 130 again by training to update current deep learning neural network
Input data input current deep learning neural network is trained, and obtains training result information.
Preset condition is to represent to train successful condition.Specifically, network training module 130 again will training input data
It inputs updated current deep learning neural network to be trained, obtains updated current deep learning neural network and correspond to
The training result information that training obtains.
Training terminate module 170 is used for when training result information meets preset condition, and current depth is learnt nerve net
Network is as the deep learning neural network after training.
Training input data input current deep learning neural network is trained and is instructed by above-mentioned number of plies adjusting apparatus
Practice result information, if training result information does not meet the successful preset condition of training, adjust current deep learning neural network
The number of plies of middle hidden layer returns and training input data is inputted updated work as again to update current deep learning neural network
Preceding deep learning neural network is trained, and whether meets preset condition with replicate analysis；If training result information meets default
Condition, expression is trained successfully, using current deep learning neural network as the deep learning neural network after training.It is in this way, logical
The adjustment for carrying out hidden layers numbers according to training result information in the training process is crossed, until training result information meets default item
Part, it can be ensured that deep learning neural network is trained successfully.Compared to traditional training deep learning neural network fashion, this Shen
Please number of plies adjusting apparatus it is high to the training success rate of deep learning neural network.
In one embodiment, training input data includes no label input data, tape label input data and each band mark
Sign the corresponding expected label of input data.Network training module 130 includes unsupervised learning unit (not shown), supervised learning
Unit (not shown) and Information Statistics unit (not shown).Unsupervised learning unit will be without label input data as current depth
The input data of learning neural network carries out unsupervised learning to current deep learning neural network, it is initially neural to obtain parameter
Network.Supervised learning unit is using tape label input data and the corresponding expected label of tape label input data as parameter
The input data and anticipated output of initial neural network, to parameter, initial neural network carries out supervised learning, is exported
The physical tags of corresponding tape label input data.The number of Information Statistics unit statistics supervised learning obtains training total degree,
The quantity of statistics tape label input data obtains total sample number, by training total degree, total sample number and each tape label input data
Corresponding physical tags are as training result information.
At the beginning of it will obtain parameter without label input data input current deep learning neural network progress unsupervised learning
On the basis of beginning neural network, further according to tape label input data and the corresponding expected label of tape label input data at the beginning of parameter
Beginning neural network carries out supervised learning, is trained with reference to unsupervised learning and supervised learning, closer to global optimum, instruction
It is good to practice effect.
Specifically, in the present embodiment, preset condition includes：Training total degree is less than or equal to default times of total sample number
Number, and the average value of the residual error of the expection label and physical tags of all tape label input datas is less than or equal to default residual error threshold
Value.By judging whether to meet according to the average value of training total degree, total sample number and the residual error of expected label and physical tags
Preset condition can accurately analyse whether successfully to be fitted.It is appreciated that in other embodiments, preset condition can also be it
He.
Wherein, preset multiple can be pre-set according to actual needs；Specifically, preset multiple can be more than 1, can also
It is less than 1 more than zero.Average value passes through the expection label for calculating each tape label input data respectively and corresponding physical tags
Residual error, which obtains multiple residual errors, calculates being averaged for multiple residual errors again is worth to.Specifically, the expection of each tape label input data is chosen
Label calculates residual error with the physical tags that the initial neural network of corresponding tape label input data last time input parameter obtains.
In one embodiment, above-mentioned number of plies adjusting apparatus further includes network creation module (not shown), has for creating
There are one output layer, a grader, a hidden layer, an input layer and the deep learning nerve nets of top-down arrangement successively
Network obtains current deep learning neural network.To be used as initial god there are one the deep learning neural network of hidden layer by creating tool
Through network, based on the hidden layer of minimum number of layers, subsequently to increase hidden layer.Specifically, network creation module is in network training
Module 130 creates before performing corresponding function and obtains current deep learning neural network.
Specifically, in the present embodiment, number of plies adjustment module 150 adjusts the number of plies of hidden layer in current deep learning neural network
It is one hidden layer of insertion in current deep learning neural network to update current deep learning neural network, obtains new work as
Preceding deep learning neural network.By being stepped up the number of plies of hidden layer, and the analytical judgment for passing through preset condition checks fitting
Situation, enabling the state being just fully fitted is progressivelyed reach, without because blindly increasing the number of plies leads to over-fitting.
Specifically, number of plies adjustment module 150 creates the hidden layer being newly inserted into, the last one in deep learning neural network is hidden
The output of layer is connected with the input of hidden layer being newly inserted by encoding and decoding network, using the output for the hidden layer being newly inserted into as grader
Input, to update current deep learning neural network.
In the present embodiment, the number of nodes for the hidden layer being newly inserted into is less than or equal to the section of hidden layer that the hidden layer being newly inserted into is connected
Points.The hidden layer that the hidden layer being newly inserted into is connected is the last one hidden layer of link sort device before updating.It is new slotting by allowing
The number of nodes rather than the node more than the last one hidden layer that the number of nodes of the hidden layer entered is less than or equal to the last one hidden layer
Number can so that the top layer concept for inputting grader is more abstract, neglect the spy that can not be mapped with output label in this way
Sign, taking out can be with output label energy fully corresponding feature.
It is appreciated that in other embodiments, network creation module can also be first establish multiple hidden layer current
Deep learning neural network.The number of plies adjusts module 150 when training result information does not meet preset condition, reduces by a hidden layer.
In one embodiment, above-mentioned number of plies adjusting apparatus further includes network output module (not shown), for that will train
Deep learning neural network output afterwards.Deep learning neural network after training is the deep learning nerve net of successfully fitting
Network, the deep learning neural network after output training can be used for Machine self-learning.
In one embodiment, a kind of medium is provided, is stored with computer program, the computer program of storage is by processor
The step of above-mentioned deep learning neural network training method is realized during execution.Specifically, medium can be computer-readable storage
Medium.
In one embodiment, a kind of robot system is provided, including memory, processor and is stored on a memory simultaneously
The computer program that can be run on a processor, processor realize above-mentioned deep learning neural network instruction when performing computer program
The step of practicing method.
Above-mentioned medium and robot system, similarly can be true due to realizing above-mentioned deep learning neural network training method
Training successfully for deep learning neural network is protected, training success rate is high.
Each technical characteristic of embodiment described above can be combined arbitrarily, to make description succinct, not to above-mentioned reality
It applies all possible combination of each technical characteristic in example to be all described, as long as however, the combination of these technical characteristics is not deposited
In contradiction, it is all considered to be the range of this specification record.
Embodiment described above only expresses the several embodiments of the present invention, and description is more specific and detailed, but simultaneously
It cannot therefore be construed as limiting the scope of the patent.It should be pointed out that those of ordinary skill in the art are come
It says, without departing from the inventive concept of the premise, various modifications and improvements can be made, these belong to the protection of the present invention
Range.Therefore, the protection domain of patent of the present invention should be determined by the appended claims.
Claims (10)
1. a kind of deep learning neural network training method, which is characterized in that including：
Obtain training input data；
The trained input data input current deep learning neural network is trained, obtains training result information, it is described
Current deep learning neural network includes input layer, hidden layer, grader and output layer；
If the training result information does not meet preset condition, the layer of hidden layer in the current deep learning neural network is adjusted
Number is returned described by the trained input data input current depth study nerve with updating current deep learning neural network
Network is trained, the step of obtaining training result information；
If the training result information meets the preset condition, using the current deep learning neural network as training after
Deep learning neural network.
2. deep learning neural network training method according to claim 1, which is characterized in that the trained input data
It is described by described in including no label input data, tape label input data and the corresponding expected label of each tape label input data
Training input data input current deep learning neural network is trained, and obtains training result information, including：
Using the no label input data as the input data of current deep learning neural network, the current depth is learnt
Neural network carries out unsupervised learning, obtains the initial neural network of parameter；
At the beginning of the tape label input data and the corresponding expected label of the tape label input data as the parameter
The input data and anticipated output of beginning neural network carry out supervised learning to the initial neural network of the parameter, are exported
Correspondence tape label input data physical tags；
The number of statistics supervised learning obtains training total degree, and counting the quantity of the tape label input data, to obtain sample total
Number, using the trained total degree, the total sample number and the corresponding physical tags of each tape label input data as the training
Result information.
3. deep learning neural network training method according to claim 2, which is characterized in that the preset condition packet
It includes：The trained total degree is less than or equal to the preset multiple of the total sample number, and the expection of all tape label input datas
The average value of the residual error of label and physical tags is less than or equal to default threshold residual value.
4. according to claim 1-3 any one of them deep learning neural network training methods, which is characterized in that described by institute
It states trained input data input current deep learning neural network to be trained, before obtaining training result information, further include：
Creating tool, there are one output layer, a grader, a hidden layer, an input layer and the depth of top-down arrangement successively
Learning neural network obtains the current deep learning neural network.
5. deep learning neural network training method according to claim 4, which is characterized in that the adjustment is described current
The number of plies of hidden layer to be to update current deep learning neural network in deep learning neural network, including：
The hidden layer being newly inserted into is created, by the output of the last one hidden layer in the deep learning neural network and the hidden layer being newly inserted into
Input be connected by encoding and decoding network, using the hidden layer being newly inserted into output as the grader input, with update currently
Deep learning neural network.
6. deep learning neural network training method according to claim 5, which is characterized in that the section for the hidden layer being newly inserted into
Points are less than or equal to the number of nodes of hidden layer that the hidden layer being newly inserted into is connected.
7. a kind of number of plies adjusting apparatus, which is characterized in that including：
Data acquisition module, for obtaining trained input data；
Network training module for the trained input data input current deep learning neural network to be trained, obtains
Training result information, the current deep learning neural network include input layer, hidden layer, grader and output layer；
The number of plies adjusts module, for when the training result information does not meet preset condition, adjusting the current depth study
The number of plies of hidden layer is to update current deep learning neural network in neural network, and returns to the network training module again by institute
It states trained input data input current deep learning neural network to be trained, obtains training result information；
Training terminate module, for when the training result information meets the preset condition, the current depth to be learnt
Neural network is as the deep learning neural network after training.
8. number of plies adjusting apparatus according to claim 7, which is characterized in that network creation module is further included, for creating
There are one output layer, a grader, a hidden layer, an input layer and the deep learning of top-down arrangement successively nerves for tool
Network obtains the current deep learning neural network.
9. a kind of medium, is stored with computer program, which is characterized in that is realized when the computer program of storage is executed by processor
Such as the step of any one of claim 1-6 the methods.
10. a kind of robot system including memory, processor and stores the meter that can be run on a memory and on a processor
Calculation machine program, which is characterized in that the processor is realized when performing the computer program as described in claim any one of 1-6
The step of method.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201711330260.5A CN108171329A (en) | 2017-12-13 | 2017-12-13 | Deep learning neural network training method, number of plies adjusting apparatus and robot system |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201711330260.5A CN108171329A (en) | 2017-12-13 | 2017-12-13 | Deep learning neural network training method, number of plies adjusting apparatus and robot system |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108171329A true CN108171329A (en) | 2018-06-15 |
Family
ID=62525851
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201711330260.5A Pending CN108171329A (en) | 2017-12-13 | 2017-12-13 | Deep learning neural network training method, number of plies adjusting apparatus and robot system |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108171329A (en) |
Cited By (15)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108922521A (en) * | 2018-08-15 | 2018-11-30 | 合肥讯飞数码科技有限公司 | A kind of voice keyword retrieval method, apparatus, equipment and storage medium |
| CN108985463A (en) * | 2018-07-15 | 2018-12-11 | 大国创新智能科技(东莞)有限公司 | The artificial intelligence method of fighting and robot system in knowledge based library and deep learning |
| CN108985456A (en) * | 2018-07-25 | 2018-12-11 | 大国创新智能科技(东莞)有限公司 | The number of plies increases and decreases deep learning neural network training method, system, medium and equipment |
| CN109242089A (en) * | 2018-07-27 | 2019-01-18 | 大国创新智能科技(东莞)有限公司 | Progressive supervision deep learning neural network training method, system, medium and equipment |
| CN109291052A (en) * | 2018-10-26 | 2019-02-01 | 山东师范大学 | A kind of massaging manipulator training method based on deeply study |
| CN109784490A (en) * | 2019-02-02 | 2019-05-21 | 北京地平线机器人技术研发有限公司 | Training method, device and the electronic equipment of neural network |
| CN110263949A (en) * | 2019-06-21 | 2019-09-20 | 安徽智寰科技有限公司 | Merge the data processing method and system of machine mechanism and intelligent algorithm system |
| CN110532318A (en) * | 2019-09-02 | 2019-12-03 | 安徽三马信息科技有限公司 | A kind of injection molding machine operating condition data analysis system based on more hidden layer neural networks |
| CN111060221A (en) * | 2019-12-31 | 2020-04-24 | 云领电气智能科技(苏州)有限公司 | Transformer overheating fault early warning method based on cyclic neural network |
| CN111246056A (en) * | 2018-11-29 | 2020-06-05 | 佳能株式会社 | Information processing apparatus, control method therefor, image capturing apparatus, and storage medium |
| CN112133426A (en) * | 2020-09-11 | 2020-12-25 | 上海朔茂网络科技有限公司 | Respiratory system disease auxiliary diagnosis method based on deep learning |
| CN112734308A (en) * | 2021-03-10 | 2021-04-30 | 张怡然 | Data collaboration system and method based on neural network |
| CN112819152A (en) * | 2020-08-14 | 2021-05-18 | 腾讯科技（深圳）有限公司 | Neural network training method and device |
| WO2021120719A1 (en) * | 2019-12-19 | 2021-06-24 | 华为技术有限公司 | Neural network model update method, and image processing method and device |
| TWI746038B (en) * | 2020-07-02 | 2021-11-11 | 阿證科技股份有限公司 | Neural network-like artificial intelligence decision-making core system |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103945533A (en) * | 2014-05-15 | 2014-07-23 | 济南嘉科电子技术有限公司 | Big data based wireless real-time position positioning method |
| US20140310218A1 (en) * | 2013-04-11 | 2014-10-16 | Nec Laboratories America, Inc. | High-Order Semi-RBMs and Deep Gated Neural Networks for Feature Interaction Identification and Non-Linear Semantic Indexing |
| CN105787557A (en) * | 2016-02-23 | 2016-07-20 | 北京工业大学 | Design method of deep nerve network structure for computer intelligent identification |
-
2017
- 2017-12-13 CN CN201711330260.5A patent/CN108171329A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20140310218A1 (en) * | 2013-04-11 | 2014-10-16 | Nec Laboratories America, Inc. | High-Order Semi-RBMs and Deep Gated Neural Networks for Feature Interaction Identification and Non-Linear Semantic Indexing |
| CN103945533A (en) * | 2014-05-15 | 2014-07-23 | 济南嘉科电子技术有限公司 | Big data based wireless real-time position positioning method |
| CN105787557A (en) * | 2016-02-23 | 2016-07-20 | 北京工业大学 | Design method of deep nerve network structure for computer intelligent identification |
Non-Patent Citations (2)
| Title |
|---|
| 奚雪峰等: "面向自然语言处理的深度学习研究", 《自动化学报》 * |
| 杨铁军: "《产业专利分析报告 第33册 智能识别》", 30 June 2015, 知识产权出版社 * |
Cited By (25)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108985463A (en) * | 2018-07-15 | 2018-12-11 | 大国创新智能科技(东莞)有限公司 | The artificial intelligence method of fighting and robot system in knowledge based library and deep learning |
| CN108985463B (en) * | 2018-07-15 | 2021-10-26 | 大国创新智能科技(东莞)有限公司 | Artificial intelligence combat method and system based on knowledge base and deep learning |
| CN108985456A (en) * | 2018-07-25 | 2018-12-11 | 大国创新智能科技(东莞)有限公司 | The number of plies increases and decreases deep learning neural network training method, system, medium and equipment |
| CN108985456B (en) * | 2018-07-25 | 2021-06-22 | 大国创新智能科技(东莞)有限公司 | Number-of-layers-increasing deep learning neural network training method, system, medium, and device |
| CN109242089B (en) * | 2018-07-27 | 2021-03-19 | 大国创新智能科技(东莞)有限公司 | Progressive supervised deep learning neural network training method, system, medium and device |
| CN109242089A (en) * | 2018-07-27 | 2019-01-18 | 大国创新智能科技(东莞)有限公司 | Progressive supervision deep learning neural network training method, system, medium and equipment |
| CN108922521A (en) * | 2018-08-15 | 2018-11-30 | 合肥讯飞数码科技有限公司 | A kind of voice keyword retrieval method, apparatus, equipment and storage medium |
| CN109291052A (en) * | 2018-10-26 | 2019-02-01 | 山东师范大学 | A kind of massaging manipulator training method based on deeply study |
| CN109291052B (en) * | 2018-10-26 | 2021-11-09 | 山东师范大学 | Massage manipulator training method based on deep reinforcement learning |
| US11165952B2 (en) | 2018-11-29 | 2021-11-02 | Canon Kabushiki Kaisha | Information processing apparatus, image capturing apparatus, method for controlling information processing apparatus, and non-transitory storage medium |
| CN111246056A (en) * | 2018-11-29 | 2020-06-05 | 佳能株式会社 | Information processing apparatus, control method therefor, image capturing apparatus, and storage medium |
| CN111246056B (en) * | 2018-11-29 | 2022-05-17 | 佳能株式会社 | Information processing apparatus, control method therefor, image capturing apparatus, and storage medium |
| US11645537B2 (en) | 2019-02-02 | 2023-05-09 | Beijing Horizon Robotics Technology Research And Development Co., Ltd. | Neural network training method, neural network training apparatus and electronic device |
| CN109784490A (en) * | 2019-02-02 | 2019-05-21 | 北京地平线机器人技术研发有限公司 | Training method, device and the electronic equipment of neural network |
| CN109784490B (en) * | 2019-02-02 | 2020-07-03 | 北京地平线机器人技术研发有限公司 | Neural network training method and device and electronic equipment |
| CN110263949A (en) * | 2019-06-21 | 2019-09-20 | 安徽智寰科技有限公司 | Merge the data processing method and system of machine mechanism and intelligent algorithm system |
| CN110532318A (en) * | 2019-09-02 | 2019-12-03 | 安徽三马信息科技有限公司 | A kind of injection molding machine operating condition data analysis system based on more hidden layer neural networks |
| WO2021120719A1 (en) * | 2019-12-19 | 2021-06-24 | 华为技术有限公司 | Neural network model update method, and image processing method and device |
| CN111060221A (en) * | 2019-12-31 | 2020-04-24 | 云领电气智能科技(苏州)有限公司 | Transformer overheating fault early warning method based on cyclic neural network |
| TWI746038B (en) * | 2020-07-02 | 2021-11-11 | 阿證科技股份有限公司 | Neural network-like artificial intelligence decision-making core system |
| CN112819152A (en) * | 2020-08-14 | 2021-05-18 | 腾讯科技（深圳）有限公司 | Neural network training method and device |
| CN112819152B (en) * | 2020-08-14 | 2024-03-01 | 腾讯科技（深圳）有限公司 | Neural network training method and device |
| CN112133426A (en) * | 2020-09-11 | 2020-12-25 | 上海朔茂网络科技有限公司 | Respiratory system disease auxiliary diagnosis method based on deep learning |
| CN112734308A (en) * | 2021-03-10 | 2021-04-30 | 张怡然 | Data collaboration system and method based on neural network |
| CN112734308B (en) * | 2021-03-10 | 2023-06-27 | 张怡然 | Data collaboration system and method based on neural network |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108171329A (en) | Deep learning neural network training method, number of plies adjusting apparatus and robot system | |
| CN111199550B (en) | Training method, segmentation method, device and storage medium of image segmentation network | |
| CN111754596B (en) | Editing model generation method, device, equipment and medium for editing face image | |
| CN108182389A (en) | User data processing method, robot system based on big data and deep learning | |
| CN110147732A (en) | Refer to vein identification method, device, computer equipment and storage medium | |
| CN111914644A (en) | Dual-mode cooperation based weak supervision time sequence action positioning method and system | |
| CN108846314A (en) | A kind of food materials identification system and food materials discrimination method based on deep learning | |
| CN109948647A (en) | A kind of electrocardiogram classification method and system based on depth residual error network | |
| CN106068514A (en) | For identifying the system and method for face in free media | |
| CN110148318A (en) | A kind of number assiatant system, information interacting method and information processing method | |
| CN110215216B (en) | Behavior identification method and system based on skeletal joint point regional and hierarchical level | |
| CN108108807A (en) | Learning-oriented image processing method, system and server | |
| US20110150301A1 (en) | Face Identification Method and System Using Thereof | |
| CN113961705A (en) | Text classification method and server | |
| CN116994709B (en) | Personalized diet and exercise recommendation method and system and electronic equipment | |
| KR20200074958A (en) | Neural network learning method and device | |
| CN110503155A (en) | A kind of method and relevant apparatus, server of information classification | |
| CN112418302A (en) | Task prediction method and device | |
| CN109242089B (en) | Progressive supervised deep learning neural network training method, system, medium and device | |
| CN113591988B (en) | Knowledge cognitive structure analysis method, system, computer equipment, medium and terminal | |
| CN111046655A (en) | Data processing method and device and computer readable storage medium | |
| KR102251704B1 (en) | Method and Apparatus for Detecting Object Using Relational Query | |
| CN108009636A (en) | Deep learning ANN Evolutionary method, apparatus, medium and computer equipment | |
| WO2021006779A1 (en) | Network status classification | |
| CN113255765B (en) | Cognitive learning method based on brain mechanism |
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: 20180615 |
|
| RJ01 | Rejection of invention patent application after publication |