CN109117884A - A kind of image-recognizing method based on improvement supervised learning algorithm - Google Patents
A kind of image-recognizing method based on improvement supervised learning algorithm Download PDFInfo
- Publication number
- CN109117884A CN109117884A CN201810933344.6A CN201810933344A CN109117884A CN 109117884 A CN109117884 A CN 109117884A CN 201810933344 A CN201810933344 A CN 201810933344A CN 109117884 A CN109117884 A CN 109117884A
- Authority
- CN
- China
- Prior art keywords
- image
- pulse sequence
- ignition
- output
- delay
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
-
- 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
-
- 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)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Health & Medical Sciences (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- Computational Linguistics (AREA)
- Health & Medical Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of based on the image-recognizing method for improving supervised learning algorithm, includes the following steps: S1, image is inputted to multilayer impulsive neural networks, coding obtains input pulse sequence；S2, introducing delay policy and noise threshold improve supervised learning algorithm, and learn to input pulse sequence, obtain output pulse sequence；S3, output pulse sequence and target pulse sequence are compared, obtains error signal；Whether S4, error in judgement signal meet the requirements；S5, each layer weight of adjustment multilayer impulsive neural networks model and delay；The output pulse sequence that S6, basis obtain, realizes the identification of image；The learning efficiency that the present invention solves supervised learning algorithm of the existing technology is low, causes to be unable to satisfy requirement to the recognition efficiency of image, and noise resisting ability is low, leads to accuracy rate and low efficiency under noisy environment, to the problem of the identification inaccuracy of image.
Description
Technical field
The invention belongs to technical field of computer vision, and in particular to a kind of based on the image for improving supervised learning algorithm
Recognition methods.
Background technique
Since the 1980s, artificial neural network becomes the hot spot of artificial intelligence field research.Its purpose studied
It is that the neuron of abstract human brain establishes model, by comparing the feature of true human brain to the process of information processing with difference
Connection type form different artificial neural networks.Artificial neural network is a computation model, passes through different connection types
Neuron composition.Traditional neural network generallys use threshold value form, although flexibility and practicability are relatively high,
It is that it does not consider temporal information in neuron, is not enough to express brain information expression in true.Impulsive neural networks (SNN)
Referred to as third generation artificial neural network, maximum feature are to be proved to be to locate using sequence form transmitting and expressing information
Manage the most suitable tool of time and spatial information.
In recent years, artificial neural network becomes the hot spot of artificial intelligence field research.It has recently been demonstrated that pulse
Neural network (SNNs) can simulate the processing of the complex information in brain.There is biological evidence to prove that the neuron of brain uses essence
The pulse signal of true time is encoded.However, the study mechanism of precise time training neuron is still one undecided
The problem of.It is that brain obtains fundamental characteristics necessary to new knowledge and the new technical ability of exploitation from the study that instructs or demonstrate, although base
Many years are had been investigated in the concept of instruction study, but realize that the definite neuronal mechanism of this process is not taken off so far
Dew.It is existing at present from instructing or demonstration most of learning method is to be adjusted based on weight, and respectively have its advantage and disadvantage,
Good accuracy rate and efficiency can be still maintained under noisy environment there is no a kind of algorithm.And noise is widely present in arteries and veins
It rushes in neural network (SNN), therefore proposes that a kind of with the algorithm of good learning ability is still very under noise conditions
Significant.
In conclusion the prior art has the following problems:
(1) learning efficiency of the supervised learning algorithm of existing impulsive neural networks is low, cause to the recognition efficiency of image without
Method is met the requirements；
(2) noise is widely present in impulsive neural networks, and existing supervised learning algorithm noise resisting ability is low, is caused
Accuracy rate and low efficiency under noisy environment, thus to the identification inaccuracy of image.
Summary of the invention
For above-mentioned deficiency in the prior art, a kind of recognition efficiency provided by the invention is high and accuracy rate is high base
In the image-recognizing method for improving supervised learning algorithm, the learning efficiency of supervised learning algorithm is improved, and in noisy environment
Under accuracy rate and efficiency, the learning efficiency for solving supervised learning algorithm of the existing technology is low, leads to the knowledge to image
Other efficiency is unable to satisfy requirement, and noise resisting ability is low, leads to accuracy rate and low efficiency under noisy environment, the knowledge to image
Not inaccurate problem.
In order to achieve the above object of the invention, the technical solution adopted by the present invention are as follows:
A kind of image-recognizing method based on improvement supervised learning algorithm, includes the following steps:
S1: according to target pulse sequence definition image type；
S2: pretreated image input multilayer impulsive neural networks are trained, will be schemed using delay phase code
As information is converted into pulse firing mode, coding obtains input pulse sequence；
S3: introducing delay policy and noise threshold improves supervised learning algorithm, and is calculated using supervised learning is improved
Method learns input pulse sequence, obtains output pulse sequence；
S4: output pulse sequence and target pulse sequence are compared, error signal is obtained；
S5: whether error in judgement signal meets the requirements, if then entering step S7, otherwise enters step S6；
S6: according to error signal, each layer weight of multilayer impulsive neural networks model and delay are adjusted, and enters step S3；
S7: corresponding image type in output step S1 realizes the identification of image.
Further, the calculation formula of pulse train are as follows:
S (t)=∑fδ(t-tf)
In formula, S (t) is pulse train, including input pulse sequence Si(t), output pulse sequence So(t) and target arteries and veins
Rush sequence Sd(t)；F={ 1,2 .., N } is the label of pulse, and N is the quantity of current PRF；δ(t-tf) it is Dirac function；tf
For the duration of ignition in past；T is the current duration of ignition.
Further, the learning model of multilayer impulsive neural networks includes input layer, hidden layer and output layer, adjacent
Using full connection between the neuron of layer.
Further, in the delay policy of step S2, the calculation formula of neuron delay are as follows:
In formula, dtmFor neuron delay；For the late ignition time；T is the current duration of ignition；xmIt (t) is late ignition
The track local variable of time；τxFor delay time constant；A is amplitude.
Further, the formula of noise threshold are as follows:
In formula, n_thr is noise threshold；Thr is former fixed threshold；η1、η2It is noise threshold parameter with a；T is current point
The fiery time；tdIt (i) is the target ignition moment；For the desired output time；For the undesirable output time.
Further, be added noise threshold after output neuron postsynaptic potential calculation formula are as follows:
In formula, V (t) is output neuron postsynaptic potential；K(t-ti-dtm) it is film potential influence function；VrestIt is quiet
Cease current potential；θ (t) is dynamic threshold, including noise threshold n_thr；T is the current duration of ignition；For the duration of ignition in postsynaptic；
dtmFor delay；wiFor the weight of i-th of presynaptic neuron；τmFor current time constant.
Further, in step S3, error signal is instantaneous network error, and is defined as all output neurons
Practical instantaneous ignition rate and expectation instantaneous ignition rate between otherness, the formula of error signal function are as follows:
In formula,For instantaneous network error；For practical instantaneous ignition rate；It is instantaneous it is expected
Lighting rate；O is output neuron quantity.
Further, in step S5, the more new formula for adjusting weight is as follows:
Weight more new formula of the input layer to hidden layer are as follows:
In formula, Δ whiIt (t) is the weight renewal function of input layer to hidden layer；To learn
Practise the convolution of window and input layer i pulse train；S is integration variable；T is the current duration of ignition；dhiIt is arrived for input layer
The delay of hidden layer；Si(t-dhi- s) it is the input pulse sequence that delay and integration variable influence is added；For reality output
Pulse train；For target output pulse sequence；A is the weight factor unrelated with activity；apost(s) it is kernel, that is, learns
Window；nhFor hidden layer neuron number；niFor input layer number；wohFor the weight of hidden layer to output layer；
Weight more new formula of the hidden layer to output layer are as follows:
In formula, Δ wohIt (t) is the weight renewal function of hidden layer to output layer；To learn
Practise the convolution of window and input layer i pulse train；S is integration variable；T is the current duration of ignition；dohIt is arrived for hidden layer
The delay of output layer；Si(t-dhi- s) it is the input pulse sequence that delay and integration variable influence is added；dohIt is hidden layer to defeated
The delay of layer out；For reality output pulse train；For target output pulse sequence；A is the power unrelated with activity
Repeated factor；apostIt (s) is kernel, i.e. study window；nhFor hidden layer neuron number.
Further, in step S5, the local variable for being adjusted to update track of delay, more new formula are as follows:
In formula, xiIt (t) is the track local variable renewal function of the current duration of ignition；τxFor delay time constant；A is vibration
Width；To update the duration of ignition；T is the current duration of ignition.
This programme has the beneficial effect that
(1) the improvement supervised learning algorithm that this method uses, according to the difference of output pulse sequence and target pulse sequence
Property update weight, with reach make really export target pulse sequence is gradually approached in learning process.With conventional multilayer ReSuMe
Algorithm is compared, and improved algorithm is obviously improved in convergence, and improves learning efficiency, to improve image
Recognition efficiency；
(2) the improvement supervised learning algorithm that this method uses, very good solution anti-noise problem, by noise threshold application
Into algorithm, to generate new algorithm, model is caused to still ensure that the accuracy of its igniting in the case where noise is added.
In addition the part of former fixed threshold is had more in the desired duration of ignition and the unexpected time lower than former fixed value part is adjustable
, thus it is more flexible in terms of noise immunity, therefore the algorithm will have stronger robustness, to improve image recognition
Accuracy.
Detailed description of the invention
Fig. 1 is based on the image-recognizing method flow chart for improving supervised learning algorithm；
Fig. 2 is study front and back membrane voltage comparison diagram；
Fig. 3 is neuron length of delay histogram after study；
Fig. 4 is study front and back neuron weight comparison diagram；
Fig. 5 is the result figure of image recognition.
Specific embodiment
A specific embodiment of the invention is described below, in order to facilitate understanding by those skilled in the art originally
Invention, it should be apparent that coming the present invention is not limited to the range of specific embodiment to those skilled in the art
It says, as long as various change is in the spirit and scope of the present invention that the attached claims limit and determine, these variations are aobvious
And be clear to, all are using the innovation and creation of present inventive concept in the column of protection.
It is a kind of based on the image-recognizing method for improving supervised learning algorithm in the embodiment of the present invention, as shown in Figure 1, including
Following steps:
S1: according to target pulse sequence definition image type；
S2: pretreated image input multilayer impulsive neural networks are trained, will be schemed using delay phase code
As information is converted into pulse firing mode, coding obtains input pulse sequence；
S3: introducing delay policy and noise threshold improves supervised learning algorithm, and is calculated using supervised learning is improved
Method learns input pulse sequence, obtains output pulse sequence；
S4: output pulse sequence and target pulse sequence are compared, and obtain error signal, and error signal is instantaneous
Network error, and the difference being defined as between the practical instantaneous ignition rate of all output neurons and expectation instantaneous ignition rate
The opposite sex, the formula of error signal function are as follows:
In formula,For instantaneous network error；For practical instantaneous ignition rate；It is instantaneous it is expected
Lighting rate；O is output neuron quantity；
The calculation formula of instantaneous ignition rate are as follows:
In formula, R (t) is instantaneous ignition rate；<S (t)>is the expectation of pulse train；SjIt (t) is to test true arteries and veins every time
Rush sequence；M is experiment number；J is experiment indicatrix；
S5: whether error in judgement signal meets the requirements, if then entering step S7, otherwise enters step S6；
S6: according to error signal, each layer weight of multilayer impulsive neural networks model and delay are adjusted, and enters step S3；
The more new formula for adjusting weight is as follows:
Weight more new formula of the input layer to hidden layer are as follows:
In formula, Δ whiIt (t) is the weight renewal function of input layer to hidden layer；To learn
Practise the convolution of window and input layer i pulse train；S is integration variable；T is the current duration of ignition；dhiIt is arrived for input layer
The delay of hidden layer；Si(t-dhi- s) it is the input pulse sequence that delay and integration variable influence is added；It is practical defeated
Pulse train out；For target output pulse sequence；A is the weight factor unrelated with activity；apost(s) it is kernel, that is, learns
Practise window；nhFor hidden layer neuron number；niFor input layer number；wohFor the weight of hidden layer to output layer；
Weight more new formula of the hidden layer to output layer are as follows:
In formula, Δ wohIt (t) is the weight renewal function of hidden layer to output layer；To learn
Practise the convolution of window and input layer i pulse train；S is integration variable；T is the current duration of ignition；dohIt is arrived for hidden layer
The delay of output layer；Si(t-dhi- s) it is the input pulse sequence that delay and integration variable influence is added；dohIt is hidden layer to defeated
The delay of layer out；For reality output pulse train；For target output pulse sequence；A is the power unrelated with activity
Repeated factor；apostIt (s) is kernel, i.e. study window；nhFor hidden layer neuron number；
The calculation formula of study window are as follows:
In formula, apostIt (s) is kernel, i.e. study window；τ is exponential function damping time constant；A is amplitude；S is product
Variation per minute；
The local variable for being adjusted to update track of delay, more new formula are as follows:
In formula, xiIt (t) is the track local variable renewal function of the current duration of ignition；τxFor delay time constant；A is vibration
Width；To update the duration of ignition；T is the current duration of ignition；
S7: corresponding image type in output step S1 realizes the identification of image.
In the present embodiment, the calculation formula of pulse train are as follows:
S (t)=∑fδ(t-tf)
In formula, S (t) is pulse train, including input pulse sequence Si(t), output pulse sequence So(t) and target arteries and veins
Rush sequence Sd(t)；F={ 1,2 .., N } is the label of pulse, and N is the quantity of current PRF；δ(t-tf) it is Dirac function；tf
For the duration of ignition in past；T is the current duration of ignition.
In the present embodiment, the learning model of multilayer impulsive neural networks includes input layer, hidden layer and output layer, phase
Using full connection between the neuron of adjacent bed.
In the present embodiment, delay policy is the expectation pulse time in no any reality output pulse, is not prolonged from previously
Slow excitatory neuron and current time does not have to select track local variable in the neuron of pulse, and calculates selected
Neuron postpones, in the delay policy of step S2, the calculation formula of neuron delay are as follows:
In formula, dtmFor neuron delay；For the late ignition time；T is the current duration of ignition；xmIt (t) is late ignition
The track local variable of time；τxFor delay time constant；A is amplitude.
In the present embodiment, noise threshold strategy be at the desired duration of ignition membrane voltage will be above it is set before training
Fixed value (typically much deeper than former fixed threshold), and the membrane voltage in the unexpected duration of ignition is below the noise threshold of setting
(typically well below former fixed threshold), the formula of noise threshold are as follows:
In formula, n_thr is noise threshold；Thr is former fixed threshold；η1、η2It is noise threshold parameter with a；T is current point
The fiery time；tdIt (i) is the target ignition moment；For the desired output time；For the undesirable output time；
The formula of desired output time are as follows:
In formula, δ is constant, and value determinesLength；
The formula of undesirable output time are as follows:
In formula, δ is constant, and value determinesLength；
The calculation formula of output neuron postsynaptic potential after addition noise threshold are as follows:
In formula, V (t) is output neuron postsynaptic potential；K(t-ti-dtm) it is film potential influence function；VrestIt is quiet
Cease current potential；θ (t) is dynamic threshold, including noise threshold n_thr；T is the current duration of ignition；For the duration of ignition in postsynaptic；
dtmFor delay；wiFor the weight of i-th of presynaptic neuron；τmFor current time constant.
Experimental result and analysis:
The neural network for constructing 100*100*1, is both 100 input layers, 100 middle layers, 1 output layer, is taken complete
Connection type connection.If the target pulse time is [40,80], input layer is random sequence (range is in 0-100ms), ωhiInitially
For the random number of 0.05-0.2, ωoh20 negative weights are initially, 80 positive weights, range is between 0.05-0.2.Nerve
Meta-model parameter a=0.001, A+=1.2, τm=5ms, τs=20ms, taking time interval is 0.1, noise threshold parameter: thr
=1.0, δ=5ms, η1=0.4mV, η2=0.1mV, a=0.01.
Front and back membrane voltage comparison diagram is learnt it is found that Multi-DLN-ReSuMe algorithm can perfectly learn to mesh by Fig. 2
Mark pulse train.Fig. 3 is that each neuron postpones histogram after study, due to it is initial when each neuron delay be 0, and
Each iteration at most only processes the delay of a neuron, and the length of delay of the neuron after delay no longer changes
Become, so it is also both the number of iterations that learning algorithm study stops that histogram delay, which changes number,.
Fig. 4 is study front and back neuron weight comparison diagram, according to comparison as can be seen that the property of neuron changes
It is few, we discussed that the change of neuron property was the advantage in a kind of calculating in neural network model to last chapter, here
Also in the case where illustrating actually in learning process if existing simultaneously inhibition and excitatory synapse, property changes
Neuron is simultaneously few.
The input of three different images, for each image be arranged a target pulse sequence, length 3, image one
Target pulse sequence is 30ms, 60ms, 90ms；The target pulse sequence of image two is 40ms, 70ms, 100ms；Image three
Target pulse sequence is 50ms, 80ms, 110ms, and the result of identification is illustrated in fig. 5 shown below, and wherein left figure is neuron membrane before training
Voltage change figure, the right are neuron membrane voltage change figure after training.
As seen from Figure 5, it does not train the pulse train of first three picture in random distribution, learns by learning algorithm
After, model can light a fire to the corresponding spiking neuron of three pictures in corresponding object time, can either be to different defeated
Enter mode and generates corresponding output pulse sequence.Above the experimental results showed that the impulsive neural networks learning algorithm can succeed
Identify three kinds of different pictures.
The improvement supervised learning algorithm that this method uses, according to the otherness of output pulse sequence and target pulse sequence
Weight is updated, gradually approaches target pulse sequence in learning process to reach to make really to export.It is calculated with conventional multilayer ReSuMe
Method is compared, and improved algorithm is obviously improved in convergence, and improves learning efficiency, to improve image
Recognition efficiency；
The improvement supervised learning algorithm that this method uses, very good solution anti-noise problem, by noise threshold is applied to calculation
In method, to generate new algorithm, model is caused to still ensure that the accuracy of its igniting in the case where noise is added.In addition
The desired duration of ignition have more the part of former fixed threshold and the unexpected time lower than former fixed value part be it is adjustable,
To more flexible in terms of noise immunity, therefore the algorithm will have stronger robustness, to improve the standard of image recognition
True property.
A kind of recognition efficiency provided by the invention is high and accuracy rate is high based on the image knowledge for improving supervised learning algorithm
Other method improves the learning efficiency of supervised learning algorithm, and accuracy rate and efficiency under noisy environment, solves existing
There is the learning efficiency of supervised learning algorithm existing for technology low, causes to be unable to satisfy requirement, antinoise to the recognition efficiency of image
Ability is low, leads to accuracy rate and low efficiency under noisy environment, to the problem of the identification inaccuracy of image.
Claims (9)
1. a kind of based on the image-recognizing method for improving supervised learning algorithm, which comprises the steps of:
S1: according to target pulse sequence definition image type；
S2: pretreated image input multilayer impulsive neural networks are trained, and are believed image using delay phase code
Breath is converted into pulse firing mode, and coding obtains input pulse sequence；
S3: introducing delay policy and noise threshold improves supervised learning algorithm, and uses improvement supervised learning algorithm pair
Input pulse sequence is learnt, and output pulse sequence is obtained；
S4: output pulse sequence and target pulse sequence are compared, error signal is obtained；
S5: whether error in judgement signal meets the requirements, if then entering step S7, otherwise enters step S6；
S6: according to error signal, each layer weight of multilayer impulsive neural networks model and delay are adjusted, and enters step S3；
S7: corresponding image type in output step S1 realizes the identification of image.
2. according to claim 1 based on the image-recognizing method for improving supervised learning algorithm, which is characterized in that the arteries and veins
Rush the calculation formula of sequence are as follows:
S (t)=∑fδ(t-tf)
In formula, S (t) is pulse train, including input pulse sequence Si(t), output pulse sequence So(t) and target pulse sequence
Arrange Sd(t)；F={ 1,2 .., N } is the label of pulse, and N is the quantity of current PRF；δ(t-tf) it is Dirac function；tfFor mistake
Go the duration of ignition；T is the current duration of ignition.
3. according to claim 1 based on the image-recognizing method for improving supervised learning algorithm, which is characterized in that described more
The learning model of layer impulsive neural networks includes input layer, hidden layer and output layer, is used between the neuron of adjacent layer complete
Connection.
4. according to claim 1 based on the image-recognizing method for improving supervised learning algorithm, which is characterized in that the step
In the delay policy of rapid S2, the calculation formula of neuron delay are as follows:
In formula, dtmFor neuron delay；For the late ignition time；T is the current duration of ignition；xmIt (t) is the late ignition time
Track local variable；τxFor delay time constant；A is amplitude.
5. according to claim 1 based on the image-recognizing method for improving supervised learning algorithm, which is characterized in that described to make an uproar
The formula of sound threshold value are as follows:
In formula, n_thr is noise threshold；Thr is former fixed threshold；η1、η2It is noise threshold parameter with a；When t is current igniting
Between；tdIt (i) is the target ignition moment；For the desired output time；For the undesirable output time.
6. according to claim 5 based on the image-recognizing method for improving supervised learning algorithm, which is characterized in that addition is made an uproar
The calculation formula of output neuron postsynaptic potential after sound threshold value are as follows:
In formula, V (t) is output neuron postsynaptic potential；K(t-ti-dtm) it is film potential influence function；VrestFor resting potential；
θ (t) is dynamic threshold, including noise threshold n_thr；T is the current duration of ignition；For the duration of ignition in postsynaptic；dtmTo prolong
Late；wiFor the weight of i-th of presynaptic neuron；τmFor current time constant.
7. according to claim 1 based on the image-recognizing method for improving supervised learning algorithm, which is characterized in that the step
In rapid S3, error signal is instantaneous network error, and is defined as practical instantaneous ignition rate and the phase of all output neurons
Hope the otherness between instantaneous ignition rate, the formula of error signal function are as follows:
In formula,For instantaneous network error；For practical instantaneous ignition rate；It is expected instantaneous ignition
Rate；O is output neuron quantity.
8. according to claim 1 based on the image-recognizing method for improving supervised learning algorithm, which is characterized in that the step
In rapid S5, the more new formula for adjusting weight is as follows:
Weight more new formula of the input layer to hidden layer are as follows:
In formula, Δ whiIt (t) is the weight renewal function of input layer to hidden layer；For study window
With the convolution of input layer i pulse train；S is integration variable；T is the current duration of ignition；dhiFor input layer to hidden layer
Delay；Si(t-dhi- s) it is the input pulse sequence that delay and integration variable influence is added；For reality output pulse sequence
Column；For target output pulse sequence；A is the weight factor unrelated with activity；apostIt (s) is kernel, i.e. study window；nh
For hidden layer neuron number；niFor input layer number；wohFor the weight of hidden layer to output layer；
Weight more new formula of the hidden layer to output layer are as follows:
In formula, Δ wohIt (t) is the weight renewal function of hidden layer to output layer；For study window
With the convolution of input layer i pulse train；S is integration variable；T is the current duration of ignition；dohFor hidden layer to output layer
Delay；Si(t-dhi- s) it is the input pulse sequence that delay and integration variable influence is added；dohOutput layer is arrived for hidden layer
Delay；For reality output pulse train；For target output pulse sequence；A is the weight factor unrelated with activity；
apostIt (s) is kernel, i.e. study window；nhFor hidden layer neuron number.
9. according to claim 1 based on the image-recognizing method for improving supervised learning algorithm, which is characterized in that the step
In rapid S5, the local variable for being adjusted to update track of delay, more new formula are as follows:
In formula, xiIt (t) is the track local variable renewal function of the current duration of ignition；τxFor delay time constant；A is amplitude；
To update the duration of ignition；T is the current duration of ignition.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810933344.6A CN109117884A (en) | 2018-08-16 | 2018-08-16 | A kind of image-recognizing method based on improvement supervised learning algorithm |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810933344.6A CN109117884A (en) | 2018-08-16 | 2018-08-16 | A kind of image-recognizing method based on improvement supervised learning algorithm |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109117884A true CN109117884A (en) | 2019-01-01 |
Family
ID=64853311
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810933344.6A Pending CN109117884A (en) | 2018-08-16 | 2018-08-16 | A kind of image-recognizing method based on improvement supervised learning algorithm |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109117884A (en) |
Cited By (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110659666A (en) * | 2019-08-06 | 2020-01-07 | 广东工业大学 | Image classification method of multilayer pulse neural network based on interaction |
| CN110796231A (en) * | 2019-09-09 | 2020-02-14 | 珠海格力电器股份有限公司 | Data processing method, data processing device, computer equipment and storage medium |
| CN111046954A (en) * | 2019-12-12 | 2020-04-21 | 电子科技大学 | Image classification method of spiking learning model based on dynamic threshold |
| CN111488908A (en) * | 2020-03-10 | 2020-08-04 | 天津大学 | Brain-imitating image identification method based on enhanced pulse |
| CN113255436A (en) * | 2021-04-09 | 2021-08-13 | 东南大学 | Method and equipment for extracting features of nerve pulse signals |
| CN113392911A (en) * | 2021-06-18 | 2021-09-14 | 电子科技大学 | DW-ReSuMe algorithm-based image classification method |
| CN113408611A (en) * | 2021-06-18 | 2021-09-17 | 电子科技大学 | Multilayer image classification method based on delay mechanism |
| CN113408618A (en) * | 2021-06-18 | 2021-09-17 | 电子科技大学 | Image classification method based on R-Multi-parameter PBSNLR model |
| WO2023178737A1 (en) * | 2022-03-24 | 2023-09-28 | 中国科学院深圳先进技术研究院 | Spiking neural network-based data enhancement method and apparatus |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105760930A (en) * | 2016-02-18 | 2016-07-13 | 天津大学 | Multilayer spiking neural network recognition system for AER |
| US20160210552A1 (en) * | 2013-08-26 | 2016-07-21 | Auckland University Of Technology | Improved Method And System For Predicting Outcomes Based On Spatio/Spectro-Temporal Data |
| CN107194426A (en) * | 2017-05-23 | 2017-09-22 | 电子科技大学 | A kind of image-recognizing method based on Spiking neutral nets |
-
2018
- 2018-08-16 CN CN201810933344.6A patent/CN109117884A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20160210552A1 (en) * | 2013-08-26 | 2016-07-21 | Auckland University Of Technology | Improved Method And System For Predicting Outcomes Based On Spatio/Spectro-Temporal Data |
| CN105760930A (en) * | 2016-02-18 | 2016-07-13 | 天津大学 | Multilayer spiking neural network recognition system for AER |
| CN107194426A (en) * | 2017-05-23 | 2017-09-22 | 电子科技大学 | A kind of image-recognizing method based on Spiking neutral nets |
Non-Patent Citations (3)
| Title |
|---|
| MALU ZHANG等: "《Supervised learning in spiking neural networks with noise-threshold》", 《NEUROCOMPUTING》 * |
| XIAOLIANG XU等: "《A Hierarchical Visual Recognition Model with Precise-Spike-Driven Synaptic Plasticity》", 《2016 IEEE SYMPOSIUM ON COMPUTATIONAL INTELLIGENCE (SSCI)》 * |
| 王阳阳: "《基于脉冲驱动的神经网络学习算法研究》", 《中国优秀硕士学位论文全文数据库》 * |
Cited By (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110659666B (en) * | 2019-08-06 | 2022-05-13 | 广东工业大学 | Image classification method of multilayer pulse neural network based on interaction |
| CN110659666A (en) * | 2019-08-06 | 2020-01-07 | 广东工业大学 | Image classification method of multilayer pulse neural network based on interaction |
| CN110796231A (en) * | 2019-09-09 | 2020-02-14 | 珠海格力电器股份有限公司 | Data processing method, data processing device, computer equipment and storage medium |
| CN111046954A (en) * | 2019-12-12 | 2020-04-21 | 电子科技大学 | Image classification method of spiking learning model based on dynamic threshold |
| CN111488908A (en) * | 2020-03-10 | 2020-08-04 | 天津大学 | Brain-imitating image identification method based on enhanced pulse |
| CN113255436A (en) * | 2021-04-09 | 2021-08-13 | 东南大学 | Method and equipment for extracting features of nerve pulse signals |
| CN113392911A (en) * | 2021-06-18 | 2021-09-14 | 电子科技大学 | DW-ReSuMe algorithm-based image classification method |
| CN113408611A (en) * | 2021-06-18 | 2021-09-17 | 电子科技大学 | Multilayer image classification method based on delay mechanism |
| CN113408618A (en) * | 2021-06-18 | 2021-09-17 | 电子科技大学 | Image classification method based on R-Multi-parameter PBSNLR model |
| CN113408611B (en) * | 2021-06-18 | 2022-05-10 | 电子科技大学 | Multilayer image classification method based on delay mechanism |
| CN113392911B (en) * | 2021-06-18 | 2023-04-18 | 电子科技大学 | DW-ReSuMe algorithm-based image classification method |
| CN113408618B (en) * | 2021-06-18 | 2023-04-18 | 电子科技大学 | Image classification method based on R-Multi-parameter PBSNLR model |
| WO2023178737A1 (en) * | 2022-03-24 | 2023-09-28 | 中国科学院深圳先进技术研究院 | Spiking neural network-based data enhancement method and apparatus |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109117884A (en) | A kind of image-recognizing method based on improvement supervised learning algorithm | |
| CN107203134B (en) | Front vehicle following method based on deep convolutional neural network | |
| Bouganis et al. | Training a spiking neural network to control a 4-dof robotic arm based on spike timing-dependent plasticity | |
| CN109711544A (en) | Method, apparatus, electronic equipment and the computer storage medium of model compression | |
| CN103279958B (en) | A kind of image partition method based on Spiking neural network | |
| KR20160136364A (en) | Cold neuron spike timing back propagation | |
| Wang et al. | Complex dynamic neurons improved spiking transformer network for efficient automatic speech recognition | |
| US20230367934A1 (en) | Method and apparatus for constructing vehicle dynamics model and method and apparatus for predicting vehicle state information | |
| CN111047085A (en) | Hybrid vehicle working condition prediction method based on meta-learning | |
| CN112651374B (en) | Future trajectory prediction method based on social information and automatic driving system | |
| CN109800517B (en) | Improved reverse modeling method for magnetorheological damper | |
| CN112101535A (en) | Signal processing method of pulse neuron and related device | |
| CN111724796B (en) | Musical instrument sound identification method and system based on deep pulse neural network | |
| CN111198966A (en) | Natural language video clip retrieval method based on multi-agent boundary perception network | |
| CN117008620A (en) | Unmanned self-adaptive path planning method, system, equipment and medium | |
| JPWO2022024211A5 (en) | ||
| US20240257943A1 (en) | Training method of brain activity state classification model, brain activity state classification method, device, and storage medium | |
| CN111091815A (en) | Voice recognition method of aggregation label learning model based on membrane voltage driving | |
| Yamagata et al. | Reinforcement learning with feedback from multiple humans with diverse skills | |
| CN117523359A (en) | Image comparison and identification method and device based on reinforcement learning | |
| CN108453737A (en) | Robot action track acquisition system and method based on neural network | |
| Gardner et al. | Learning temporally precise spiking patterns through reward modulated spike-timing-dependent plasticity | |
| CN110874629A (en) | Structure optimization method of reserve pool network based on excitability and inhibition STDP | |
| Kunkle et al. | Pulsed neural networks and their application | |
| CN114548497B (en) | Crowd motion path planning method and system for realizing scene self-adaption |
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: 20190101 |
|
| RJ01 | Rejection of invention patent application after publication |