CN109657791A - It is a kind of based on cerebral nerve cynapse memory mechanism towards open world successive learning method - Google Patents
It is a kind of based on cerebral nerve cynapse memory mechanism towards open world successive learning method Download PDFInfo
- Publication number
- CN109657791A CN109657791A CN201811532220.3A CN201811532220A CN109657791A CN 109657791 A CN109657791 A CN 109657791A CN 201811532220 A CN201811532220 A CN 201811532220A CN 109657791 A CN109657791 A CN 109657791A
- Authority
- CN
- China
- Prior art keywords
- task
- parameter
- learning
- model
- cynapse
- 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
- 238000000034 method Methods 0.000 title claims abstract description 44
- 230000007246 mechanism Effects 0.000 title claims abstract description 17
- 230000002490 cerebral effect Effects 0.000 title claims abstract description 13
- 210000005036 nerve Anatomy 0.000 title claims abstract description 11
- 230000004044 response Effects 0.000 claims abstract description 31
- 238000013527 convolutional neural network Methods 0.000 claims abstract description 10
- 238000012549 training Methods 0.000 claims abstract description 10
- 238000009966 trimming Methods 0.000 claims abstract description 10
- 230000003014 reinforcing effect Effects 0.000 claims abstract description 5
- 230000006870 function Effects 0.000 claims description 15
- 239000011159 matrix material Substances 0.000 claims description 13
- 230000008569 process Effects 0.000 claims description 8
- 230000000717 retained effect Effects 0.000 claims description 6
- 230000008859 change Effects 0.000 claims description 5
- 238000005457 optimization Methods 0.000 claims description 4
- 230000000007 visual effect Effects 0.000 abstract description 2
- 238000005215 recombination Methods 0.000 abstract 1
- 230000006798 recombination Effects 0.000 abstract 1
- 238000005728 strengthening Methods 0.000 abstract 1
- 238000013528 artificial neural network Methods 0.000 description 6
- 230000000694 effects Effects 0.000 description 4
- 238000005516 engineering process Methods 0.000 description 4
- 238000013473 artificial intelligence Methods 0.000 description 3
- 210000004556 brain Anatomy 0.000 description 3
- 238000013135 deep learning Methods 0.000 description 3
- 238000013507 mapping Methods 0.000 description 3
- 238000005259 measurement Methods 0.000 description 3
- 238000011160 research Methods 0.000 description 3
- 230000003956 synaptic plasticity Effects 0.000 description 3
- 230000006978 adaptation Effects 0.000 description 2
- 230000003044 adaptive effect Effects 0.000 description 2
- 239000013039 cover film Substances 0.000 description 2
- 238000013461 design Methods 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 238000002474 experimental method Methods 0.000 description 2
- 230000003585 interneuronal effect Effects 0.000 description 2
- 210000002569 neuron Anatomy 0.000 description 2
- 230000009471 action Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000008901 benefit Effects 0.000 description 1
- 238000005094 computer simulation Methods 0.000 description 1
- 238000007596 consolidation process Methods 0.000 description 1
- 238000013136 deep learning model Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 230000004927 fusion Effects 0.000 description 1
- 238000009499 grossing Methods 0.000 description 1
- 238000013138 pruning Methods 0.000 description 1
- 238000005086 pumping Methods 0.000 description 1
- 230000003252 repetitive effect Effects 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
- 230000009466 transformation Effects 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
Abstract
The invention discloses a kind of based on cerebral nerve cynapse memory mechanism towards open world successive learning method, establishes sensing module first with depth convolutional neural networks, learns to Current vision task;Secondly, building network reconfiguration module, carries out the trimming and reinforcing of cynapse on trained model, to subtract the knowledge that oligosynaptic connection is arrived with intensified learning;Finally, consolidating when learning new visual task to front memory, interference of the information of new task to front task is protected by it.For the present invention after the completion of training in a task data, recombination module can reorganize network, trim unessential parameter, promote the plasticity of parameter, while the performance in order to guarantee current task, consolidate current knowledge by strengthening important parameter, and update network;When learning new task, new task is instructed to learn by retaining the peak response figure formerly held the post of under business parameter space, so that averting a calamity property is forgotten.
Description
Technical field
The present invention relates to a kind of based on cerebral nerve cynapse memory mechanism towards open world successive learning method, belongs to
Artificial intelligence field.
Background technique
The mankind just constantly perceive ambient enviroment since baby and learn step by step, grasp various technical ability.For example, first learning
On foot, then association's running, cycles, this gradually to obtain and finely tune existing knowledge, and remains to while fusion new knowledge
It is enough to retain the experience learnt before, referred to as successive learning, it is the key that lead to intelligence.
The fast development of artificial intelligence be unable to do without the breakthrough of depth learning technology.In recent years, deep learning is in every field
Achieve the achievement to attract people's attention, from speech recognition to computer vision, and in conjunction with nitrification enhancement after " Alpha
Go " starts to roll the mankind in many projects.
However, existing depth model is based on closure it is assumed that specifically including that
It is sample space closing first it is assumed that deep learning method is usually firstly the need of preparatory building magnanimity, closed
Library is trained, then the learning characteristic from training library.There are a root problems for the incompleteness and closure of data space: test
Trained library covering power can be depended critically upon, the generalization ability of model will receive strong influence.
The followed by closure of task.The method of deep learning is mainly directed towards single particular task, constructs particular model.Base
Model is established with each data set in each task.
However, the information flow that the computing system run in real world touches is continuously, to need from dynamic data
Learn and remember multiple tasks in distribution.For example, needing it can be from past during agent is interacted with ambient enviroment
Learn in experience, then agent learn each class from data flow, data are gradually introduced in model, and algorithm is needed to connect
Continuous ground learns new class from sequence data stream.
Although traditional certain methods can solve this problem to a certain degree, there are still many limitations.Moreover,
In some scenarios, such as automatic Pilot, model is needed rapidly to make a response in learning process, traditional certain methods mistake
In study that is inefficient, and will affect new task.
The mankind inspire us based on the successive learning mechanism of synaptic plasticity: first is that the connection of cerebral neuron was parameter
Change, less connection can learn a certain task, this is the adequate condition that brain keeps successive learning;Second is that brain is by dashing forward
The plasticity of touching handles the conflict of new and old task, and the interneuronal connection of the weaker explanation of plasticity is more important to current task.
Existing research also indicates that different parameter configuration neural networks can obtain comparable performance, it was confirmed that model is crossed parametrization and asked
Topic.Based on this research, how cerebral neuron synaptic plasticity mechanism to be introduced into neural network and overcome catastrophic forgetting tool
There is important meaning.
Summary of the invention
The object of the present invention is to provide a kind of based on cerebral nerve cynapse memory mechanism towards open world successive learning
Method, to construct quick, flexible learning system based on depth learning technology, the artificial intelligence adapted under open true environment is answered
With needs.
To achieve the goals above, the present invention provide it is a kind of based on cerebral nerve cynapse memory mechanism towards open world
Successive learning method, includes the following steps:
(1) sensing module is established using depth convolutional neural networks, Current vision task is learnt;
(2) network reconfiguration module is constructed, the trimming and reinforcing of cynapse are carried out on trained model, to reduce prominent
The knowledge that the connection of touching and intensified learning arrive;
(3) when learning new task, new task is instructed by retaining the peak response figure formerly held the post of under business parameter space
Study, so that averting a calamity property is forgotten.
Further, the step (1) the following steps are included:
(11) according to current task type and data, building is based on depth convolutional neural networks identification model, including basis
Network and classifier;
(12) in training next task, the additional classifier of the task is directed to for task distribution.
Further, the step (2) the following steps are included:
(21) in decision model parameter importance, be also equivalent to model parameter carry out equal extent disturbance, that
On the contrary influencing the bigger parameter of model result degree of accuracy is exactly important parameter, then be inessential parameter;
(22) inessential parameter is trimmed, according to parameter importance, least part and parcel connection is found out, generates corresponding cover
Film, the exposure mask are the mask matrixes constituted by 0,1, and 0, which indicates inessential, to be trimmed to about, and 1 indicates that the connection can be retained;
(23) mask matrix and parameter are subjected to multiplying, to trim unessential connection;Then, using instructing again
Experienced mode reinforces the intensity between the connection retained.
Further, the step (3) the following steps are included:
(31) after having learnt a task, peak response figure, stent are generated based on basic model and corresponding classifier
Parameter θ after type training initializes an output X, specifies the value h of the i-th-th of j layer neuronsij, find objective function X*Energy
Enough so that its value tends to be maximum, objective function are as follows:
X*=argXs.t. | | X | |=ρmaxhij(θ,X)
(32) before learning new task, corresponding each task corresponds to specific output module, is generated not by output layer
After the peak response figure of output node, as the corresponding standard output of predecessor's business parameter space;It is current defeated by optimizing
Out at a distance from standard output, to control the path of the new task optimization of successive learning, with enable model in learning process always
The memory about peak response is kept, realizes that formula is as follows:
lossAM=minD (argXs.t. | | X | |=ρmaxhij(θ,X)-X*)
Wherein, D indicate the output of same node normal response with current output response at a distance from, by X-direction optimization θ come
Minimize distance;Learn new task TKWhen, complete loss function form are as follows:
loss(TK) be current task objective function, XkIt is the corresponding peak response figure generated of k-th of front task.
Through the above technical solutions, following beneficial technical effect may be implemented:
1) theoretically guarantee, this research is introduced into depth model method from brain memory mechanism, and is obtained
Certain effect;
2) relationship during successive learning between neural network capacity and network expansion threshold value is considered, is proposed adaptive
The network evolution method answered;
3) from the mechanism of mankind's successive learning, respectively come the company of discussion in terms of synaptic plasticity and cynapse trim two
The modeling method of continuous study;
4) coding expressed from invariance inquires into the study mechanism of neural network under the conditions of limited capacity, from feature space
Knowledge Representation Method is established, in successive learning;
5) semi parallel computational model is constructed based on this method, is effectively improved the successive learning energy of current depth model
Power.
The other feature and advantage of the embodiment of the present invention will the following detailed description will be given in the detailed implementation section.
Detailed description of the invention
Attached drawing is to further understand for providing to the embodiment of the present invention, and constitute part of specification, under
The specific embodiment in face is used to explain the present invention embodiment together, but does not constitute the limitation to the embodiment of the present invention.Attached
In figure:
Fig. 1 shows the flow chart of one embodiment of the invention;
Fig. 2 shows the multi-categorizers in the image recognition tasks based on depth convolutional neural networks in the present invention to strain machine
Principle processed;
Fig. 3 shows the pseudocode that cynapse is trimmed in model adaptation restructuring procedure in the present invention;
Fig. 4, which is shown in the present invention, overcomes the catastrophic schematic diagram forgotten based on peak response figure;
Fig. 5 shows the instance graph of the peak response generation figure in the present invention on depth convolutional neural networks.
Specific embodiment
It is described in detail below in conjunction with specific embodiment of the attached drawing to the embodiment of the present invention.It should be understood that this
Locate described specific embodiment and be merely to illustrate and explain the present invention embodiment, is not intended to restrict the invention embodiment.
In one embodiment of the invention, based on cerebral nerve cynapse memory mechanism towards open world successive learning
Method includes the following steps:
Step 1: be directed to current task, establish sensing module using depth convolutional neural networks, to Current vision task into
Row study, is illustrated by taking natural image identification mission as an example here;
Step 2: building network reconfiguration module carries out the trimming and reinforcing of cynapse, to subtract on trained model
The knowledge that oligosynaptic connection and intensified learning arrive;
Step 3: when learning new task, instructing new post by retaining the peak response figure formerly held the post of under business parameter space
Business study, so that averting a calamity property is forgotten.
The method, utilization depth convolutional neural networks described in step 1 establish sensing module, including following interior
Hold:
Since the difference between different task may be larger, high level has stronger semantic information, so that particular task is in height
The expression and task of layer have stronger specificity, therefore can be relatively difficult using same group of parameter coding different task.Moreover, this
Class method assumes that learning every time for task is identical with the output node number of predecessor's business, it is difficult to be extended to the learning tasks of increment class
On.For these reasons, it is optimized in structure, different classifiers is distributed for different task, to guarantee different pairs
As corresponding specific mapping output, tangling for different task high-level characteristic is avoided, not only makes model is opposite in the study to hold
Easily, and make more flexible in different scenes tasking learning.
The method, network reconfiguration described in step 2 realize the trimming and reinforcing to parameter adaptive, including following
Content:
It is the measurement of parameter importance first, the learning process of neural network is the function F (X, W) in fitting X → Y,
Middle W is the parameter of study.In order to allow the connection and weight according to self adjusting parameter of task of model adaptation, need to parameter
It is trimmed.Its matter of utmost importance is which determining parameter connection should be trimmed to about, which should be retained.Assuming that only trimming every time
It one in current all parameter connections, gradually trims, so all trimming influences the smallest connection to network each time.In order to every
It is secondary to find most unessential connection, it is modeled and just obtains following form --- to the mapping function F after increase disturbed value
Carry out second order Taylor expansion, then derive increase disturbed value δ w mapping function F (x, w, δ w) with reflecting for disturbed value is not added
Penetrate the difference between function F (x, w, 0) are as follows:
Wherein, H isThat is Hessian matrix, when model convergence after, model learning to y be intended to model
True value y_, at this time
Problem is converted into solution:
The parameter of minimized target can be allowed in order to find, and introduce Lagrange multiplier, above formula is changed into solving condition
The problem of extreme value:
Finally solve:
However, solving Hessian inverse of a matrix, to will cause operand extremely complex, for this purpose, providing two kinds of methods to force
The value of nearly Hessian matrix.
Obtaining from described above can be usedAs the measurement to parameter importance.Hessian square
Battle array can be write as following form:
Wherein, P is training sample sum, inkFor k-th of training sample of input
From the point of view of the overall situation,As the output F of model is to the gradient of parameter w, therefore, Hessian matrix
Regard the output of model as to square of the gradient of parameter, i.e. parameter importance measures matrix are as follows:
WhereinGradient of the output of representative model to each layer of parameter.
It then, is parameter trimming strategy.In order to which model can be connected according to work transformation matrix adjusting parameter, and according to new
Task generate new connection.After having learnt a task, interneuronal connection can be adjusted automatically, unessential connection
It can disappear, important bonding strength can be reinforced.
Including the following steps:
1) in decision model parameter importance, be also equivalent to model parameter carry out equal extent disturbance, then
On the contrary influencing the bigger parameter of model result degree of accuracy is exactly important parameter, then be inessential parameter;
2) inessential parameter is trimmed, according to parameter importance, least part and parcel connection is found out, generates corresponding cover
Film, the exposure mask are the mask matrixes constituted by 0,1, and 0, which indicates inessential, to be trimmed to about, and 1 indicates that the connection can be retained;
3) mask matrix and parameter are subjected to multiplying, to trim unessential connection;Then, using retraining
Mode, reinforce retain connection between intensity;
It should be noted that need to save this mask matrix, while after learning current task, the mask matrix of this task
All matrixes are done and operation before, as long as parameter important to any one task in such mask matrix all can be in mask square
It is embodied in battle array, so that avoiding current task inessential during trimming and the parameter important to front task is trimmed to about
Fall, ensure that the accuracy of trimming.
The method instructs new task by the peak response figure under the task parameters space of front described in step 3
Study, so that averting a calamity property is forgotten, including the following contents:
Parameter θ after fixed model training, initializes an output X, specifies the value h of the i-th-th of j layer neuronsij, wish
X is found in prestige*Enable to its value as big as possible.By the above problem is defined as:
X*=argXs.t. | | X | |=ρmaxhij(θ,X)
Before learning new task, corresponding each task corresponds to specific output module, is generated by output layer different defeated
The peak response figure of egress, as the corresponding standard output of predecessor's business parameter space, with the new task of study, parameter
The peak response that changing causes it to generate can also change.Wish model can be remained in learning process about
The memory of peak response, it is therefore, excellent to control the new task of successive learning by optimizing current output at a distance from standard output
The path of change.Its formula is as follows:
lossAM=minD (argXs.t. | | X | |=ρmaxhij(θ,X)-X*)
Wherein, D indicates the same node normal response output at a distance from current output response, it is desirable to excellent by X-direction
Change θ to minimize distance.Learn new task TKWhen, complete loss function form are as follows:
loss(TK) be current task objective function, XkIt is the corresponding peak response figure generated of k-th of front task.
It can still can be very good to keep the performance of old task while learning new task to verify this technology, into
It has gone two groups of experiments, and has been compared with tri- kinds of methods of SGD, EWC, MAS, to be based on mnist and Fashionmnist two
For the successive learning of data set.
Data set and network parameter setting:
Data set is identified using mnist Handwritten Digit Recognition data set and fashion-mnist dress ornament.Mnist data set
There are 10 classifications with fashion-mnist data set, can be used as a benchmark dataset of deep learning model.Respectively
Nist data set and fashion-mnist data set upset pixel, obtain two new data sets with this.Allow model in order
Successively learn four tasks, the sequence of this four tasks are as follows: A task (fashion_mnist), B task (mnist), C task
(shufflefashion_mnist), D task (shuffle mnist).It is depth convolutional neural networks architecture design first, makes
Classifier framework is added with the basic convolutional network of Fig. 2.To prevent over-fitting, joined behind the full articulamentum of the second layer
Dropout, dropout is set as 0.5 in experiment, and learning rate is set as 1e-3, and repetitive exercise 10000 times.
Performance Evaluation index:
The BWT index defined using following formula carrys out the performance of assessment models, which is learning current task t
Afterwards, the influence to task before has much, and BWT value is negative, shows the predecessor that model can have been forgotten after having learnt current task
The performance of business, numerical value is bigger, shows that forgetting degree is higher.
Wherein, T is task quantity, Ri,jLearning task t for modeliAfterwards, in task t beforejOn measuring accuracy
Then, calculating parameter importance, and beta pruning and retraining, detailed process such as Fig. 3 are carried out to parameter.
It is trained jointly finally, generating peak response figure and being put into model with new task data, whole flow process such as Fig. 4.?
When learning first task, the peak response of classification layer has been used as a result, the sample image generated can most express convolution mind
Some the class another characteristic arrived through e-learning is defined as the abstract knowledge of visual expression of the network to some class.?
On the basis of this, using the sample come is generated, is mixed in new task with new training data, re-start training, controlled
The optimization process of loss function, so that neural network still remains the pumping to classes some in the task of front when learning new task
As indicating.In order to allow peak response figure more can interpretation of images indicate essence, used certain smoothing action to guarantee to generate
The effect of response diagram out, makes it visually have more interpretation.Fig. 5 comes from the last one convolutional layer, respectively represents preceding four
The peak response result of a convolution kernel.
As a result with analysis:
It can be seen that, four kinds of methods all produce forgetting for task before from the result that table 1 obtains.SGD for
The ability of task is worst before reservation, this is because SGD does not introduce any mechanism for overcoming forgetting, although EWC is to power
Joined elastic consolidation again, but generally, for the ability of task before reservation be not it is very good, forgetting degree reaches
12.81%.MAS protects important parameter, so that the model parameter renewal speed important to task before slows down, from
As a result from the point of view of, the method effect of MAS is better than EWC and SGD, forgets degree and also there was only 8.73%.And method phase of the invention
Than be in the forgetting degree of SGD, EWC and MAS for task before it is minimum, only 6.97%.From the point of view of numerically, the present invention
Method ratio SGD improve 11.11%, improve 5.84% than EWC, improve 1.76% than MAS.Illustrate method of the invention
For parameter importance measurement ratio MAS it is more more accurate, also preliminary identification the method for the present invention is effective for this.
1 the method for the present invention of table and other methods performance comparison
The optional embodiment of the embodiment of the present invention is described in detail in conjunction with attached drawing above, still, the embodiment of the present invention is simultaneously
The detail being not limited in above embodiment can be to of the invention real in the range of the technology design of the embodiment of the present invention
The technical solution for applying example carries out a variety of simple variants, these simple variants belong to the protection scope of the embodiment of the present invention.
It is further to note that specific technical features described in the above specific embodiments, in not lance
In the case where shield, it can be combined in any appropriate way.In order to avoid unnecessary repetition, the embodiment of the present invention pair
No further explanation will be given for various combinations of possible ways.
In addition, any combination can also be carried out between a variety of different embodiments of the embodiment of the present invention, as long as it is not
The thought of the embodiment of the present invention is violated, equally should be considered as disclosure of that of the embodiment of the present invention.
Claims (4)
1. it is a kind of based on cerebral nerve cynapse memory mechanism towards open world successive learning method, which is characterized in that including
Following steps:
(1) sensing module is established using depth convolutional neural networks, Current vision task is learnt;
(2) network reconfiguration module is constructed, the trimming and reinforcing of cynapse are carried out on trained model, to subtract oligosynaptic
The knowledge that connection and intensified learning arrive;
(3) when learning new task, new task is instructed to learn by retaining the peak response figure formerly held the post of under business parameter space,
To which averting a calamity property is forgotten.
2. it is according to claim 1 based on cerebral nerve cynapse memory mechanism towards open world successive learning method,
It is characterized in that, the step (1) the following steps are included:
(11) according to current task type and data, building is based on depth convolutional neural networks identification model, including basic network
And classifier;
(12) in training next task, the additional classifier of the task is directed to for task distribution.
3. it is according to claim 1 based on cerebral nerve cynapse memory mechanism towards open world successive learning method,
It is characterized in that, the step (2) the following steps are included:
(21) in decision model parameter importance, be also equivalent to model parameter carry out equal extent disturbance, then shadow
On the contrary ringing the bigger parameter of model result degree of accuracy is exactly important parameter, then be inessential parameter;
(22) inessential parameter is trimmed, according to parameter importance, least part and parcel connection is found out, generates corresponding exposure mask,
The exposure mask is the mask matrix constituted by 0,1, and 0, which indicates inessential, to be trimmed to about, and 1 indicates that the connection can be retained;
(23) mask matrix and parameter are subjected to multiplying, to trim unessential connection;Then, using retraining
Mode reinforces the intensity between the connection retained.
4. it is according to claim 1 based on cerebral nerve cynapse memory mechanism towards open world successive learning method,
It is characterized in that, the step (3) the following steps are included:
(31) after having learnt a task, peak response figure, fixed model instruction are generated based on basic model and corresponding classifier
Parameter θ after white silk initializes an output X, specifies the value h of the i-th-th of j layer neuronsij, find objective function X*It can make
It obtains its value and tends to be maximum, objective function are as follows:
X*=argXs.t. | | X | |=ρmaxhij(θ,X)
(32) before learning new task, corresponding each task corresponds to specific output module, is generated by output layer different defeated
After the peak response figure of egress, as the corresponding standard output of predecessor's business parameter space;By optimize current output with
The distance of standard output, to control the path of the new task optimization of successive learning, to enable model remain in learning process
About the memory of peak response, realize that formula is as follows:
lossAM=minD (argXs.t. | | X | |=ρmaxhij(θ,X)-X*)
Wherein, D indicates that the same node normal response output at a distance from current output response, optimizes θ by X-direction come minimum
Change distance;Learn new task TKWhen, complete loss function form are as follows:
loss(TK) be current task objective function, XkIt is the corresponding peak response figure generated of k-th of front task.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811532220.3A CN109657791A (en) | 2018-12-14 | 2018-12-14 | It is a kind of based on cerebral nerve cynapse memory mechanism towards open world successive learning method |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811532220.3A CN109657791A (en) | 2018-12-14 | 2018-12-14 | It is a kind of based on cerebral nerve cynapse memory mechanism towards open world successive learning method |
Publications (1)
Publication Number | Publication Date |
---|---|
CN109657791A true CN109657791A (en) | 2019-04-19 |
Family
ID=66114584
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201811532220.3A Pending CN109657791A (en) | 2018-12-14 | 2018-12-14 | It is a kind of based on cerebral nerve cynapse memory mechanism towards open world successive learning method |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109657791A (en) |
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110837856A (en) * | 2019-10-31 | 2020-02-25 | 深圳市商汤科技有限公司 | Neural network training and target detection method, device, equipment and storage medium |
CN111160562A (en) * | 2019-11-18 | 2020-05-15 | 清华大学 | Continuous learning method and device based on meta-learning optimization method |
CN112257785A (en) * | 2020-10-23 | 2021-01-22 | 中科院合肥技术创新工程院 | Serialized task completion method and system based on memory consolidation mechanism and GAN model |
WO2021217282A1 (en) * | 2020-04-30 | 2021-11-04 | Chen Yongcong | Method for implementing universal artificial intelligence |
-
2018
- 2018-12-14 CN CN201811532220.3A patent/CN109657791A/en active Pending
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110837856A (en) * | 2019-10-31 | 2020-02-25 | 深圳市商汤科技有限公司 | Neural network training and target detection method, device, equipment and storage medium |
CN111160562A (en) * | 2019-11-18 | 2020-05-15 | 清华大学 | Continuous learning method and device based on meta-learning optimization method |
WO2021217282A1 (en) * | 2020-04-30 | 2021-11-04 | Chen Yongcong | Method for implementing universal artificial intelligence |
CN112257785A (en) * | 2020-10-23 | 2021-01-22 | 中科院合肥技术创新工程院 | Serialized task completion method and system based on memory consolidation mechanism and GAN model |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109657791A (en) | It is a kind of based on cerebral nerve cynapse memory mechanism towards open world successive learning method | |
CN112163465B (en) | Fine-grained image classification method, fine-grained image classification system, computer equipment and storage medium | |
CN112052948B (en) | Network model compression method and device, storage medium and electronic equipment | |
CN106503654A (en) | A kind of face emotion identification method based on the sparse autoencoder network of depth | |
CN111507993A (en) | Image segmentation method and device based on generation countermeasure network and storage medium | |
Kong et al. | Hexpo: A vanishing-proof activation function | |
CN111416797B (en) | Intrusion detection method for optimizing regularization extreme learning machine by improving longicorn herd algorithm | |
Zeng et al. | CNN model design of gesture recognition based on tensorflow framework | |
CN106503723A (en) | A kind of video classification methods and device | |
Panigrahi et al. | Deep learning approach for image classification | |
CN113537365B (en) | Information entropy dynamic weighting-based multi-task learning self-adaptive balancing method | |
CN113902131B (en) | Updating method of node model for resisting discrimination propagation in federal learning | |
CN113128671B (en) | Service demand dynamic prediction method and system based on multi-mode machine learning | |
CN106897744A (en) | A kind of self adaptation sets the method and system of depth confidence network parameter | |
CN113128478A (en) | Model training method, pedestrian analysis method, device, equipment and storage medium | |
CN111598213A (en) | Network training method, data identification method, device, equipment and medium | |
Gao et al. | Knowledge concentration: Learning 100k object classifiers in a single CNN | |
CN108229536A (en) | Optimization method, device and the terminal device of classification prediction model | |
CN113434699A (en) | Pre-training method of BERT model, computer device and storage medium | |
Sahu et al. | Dynamic routing using inter capsule routing protocol between capsules | |
CN110414560A (en) | A kind of autonomous Subspace clustering method for high dimensional image | |
CN111831356B (en) | Weight precision configuration method, device, equipment and storage medium | |
CN111445024A (en) | Medical image recognition training method | |
CN113591988B (en) | Knowledge cognitive structure analysis method, system, computer equipment, medium and terminal | |
Cong et al. | On Adam trained models and a parallel method to improve the generalization performance |
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 | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190419 |