CN108717570A - A kind of impulsive neural networks parameter quantification method - Google Patents
A kind of impulsive neural networks parameter quantification method Download PDFInfo
- Publication number
- CN108717570A CN108717570A CN201810501442.2A CN201810501442A CN108717570A CN 108717570 A CN108717570 A CN 108717570A CN 201810501442 A CN201810501442 A CN 201810501442A CN 108717570 A CN108717570 A CN 108717570A
- Authority
- CN
- China
- Prior art keywords
- parameter
- neural network
- training
- neural networks
- impulsive
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—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
- G06N3/06—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons
- G06N3/061—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons using biological neurons, e.g. biological neurons connected to an integrated circuit
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/06—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons
- G06N3/063—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons using electronic means
Abstract
The present invention relates to nerual network technique field more particularly to a kind of impulsive neural networks parameter quantification methods.The present invention method by map offline or on-line training obtain training complete original pulse neural network, the parameters such as impulsive neural networks weights, threshold value, leakage constant, set voltage, refractory period, the synaptic delay completed to training quantify, and all layers of neural network can share same group of quantization parameter or respectively one group of quantization parameter.Impulsive neural networks after parameter quantization only need a small amount of parameter that high-precision pulse neural network function can be realized.This method is high-precision simultaneously in holding, and effectively save impulsive neural networks parameter storage space improves arithmetic speed, reduces operation power consumption.
Description
Technical field
The present invention relates to nerual network technique field more particularly to a kind of impulsive neural networks parameter quantification methods.
Background technology
Impulsive neural networks (abbreviation SNN) are referred to as third generation neural network, it handles the side of information closer to human brain
Formula is the developing direction of the following nerual network technique.SNN receives information based on pulse train, has many coding modes can be with
Pulse train is construed to an actual number, common coding mode has pulse code and frequency coding.Between neuron
Communication be also to be carried out by pulse, when the film potential of a neuron is more than its threshold value, it will produce pulse letter
Number other neurons are passed to, increases or decreases its film potential.The hardware platform of SNN is referred to as neuromorphic chip or class brain core
Piece, has overturned traditional von Neumann framework completely, and this kind of chip has the features such as low-power consumption, resource consumption is few, in classification and
The performance in the classes human brain fields such as identification will be significantly better than traditional die.There are mainly two types of the training methods of SNN, one is by
The corresponding artificial neural network (abbreviation ANN) of training is mapped to trained parameter in SNN again under specified conditions, but is reflecting
It often may require that during penetrating and transmit a large amount of parameter;Another is the direct on-line study for carrying out SNN, equally also along with
Generate a large amount of parameter.Huge memory space is needed if storing parameter according to legacy memory (such as SRAM, DRAM etc.), if
Parameter is stored using new devices such as memristors, then is difficult to accurately and stably realize numerous parameters;Meanwhile huge parameter amount meeting
It reduces arithmetic speed, increase operation power consumption.There is presently no a kind of methods that can be compressed to the quantity of parameters in SNN.
Invention content
In order to solve problems in the prior art, the present invention provides a kind of methods that can reduce SNN parameter storage spaces.
Technical scheme is as follows:
A kind of impulsive neural networks parameter quantification method, which is characterized in that include the following steps:
Obtain the original SNN that training is completed.Neuron is spiking neuron (such as LIF neurons) in SNN, has input arteries and veins
Accumulation function is integrated in punching and function is provided in pulse, and for SNN using pulse train as input, the major parameter of SNN includes weights, threshold
Value, leakage constant, set voltage, refractory period, synaptic delay etc..The SNN that training is completed has the work(such as high-precision classification, identification
Energy.Obtaining the neural network that training is completed, there are mainly two types of methods:One is by map offline obtain training complete SNN,
By the methods of the training common stochastic gradient descents of ANN training ANN (including MLP, CNN, RNN, LSTM etc.), obtains satisfaction and refer to
Training process is completed after the ANN of mark requirement (such as classification, accuracy of identification etc.), then the parameter for the ANN that training is completed is mapped to
In the identical SNN of topological structure, the input of ANN is encoded into (such as pulse train of Poisson distribution) conduct afterwards using pulse train
The input of SNN, to obtain the SNN of training completion;One is the SNN completed by on-line training acquisition training, establish as certainly
The SNN or SNN of other structures is organized, the learning rules such as plasticity (STDP) are relied on using cynapse pulse sequence, using pulse sequence
It arranges (such as Poisson distribution pulse train, time encoding pulse train etc.) and SNN is trained by on-line study, adjusted in training process
The parameters such as weights, threshold value, leakage constant, set voltage, refractory period, the synaptic delay of SNN, acquisition meet index request (such as
Classification, accuracy of identification etc.) training process is completed after SNN, the weights of SNN, threshold value, leakage constant, set are fixed after training
The parameters such as voltage, refractory period, synaptic delay, to obtain the SNN of training completion;
Choose the one or more parameters for needing to quantify.The parameter that can quantify includes weights, threshold value, leakage constant, sets
Position voltage, refractory period, synaptic delay etc..
A certain layer or certain several layers of or all layer parameter distribution situation are counted respectively;
It attempts to carry out interval division to parameter.Selecting All Parameters interval division method and section number, the division methods in section
Equal point-score, non-equal point-score, confidence interval partitioning etc. can be used, the method and number of demarcation interval can be according to specific nerves
Network structure and task type and index request are attempted and are adjusted by parameter adjustment experience;
Parameter is quantified according to section.The parameter in all sections is traversed, is distributed in all in same section
Parameter is quantified as the same value (i.e. a quantized value), the size of quantized value and the specific neural network structure of positive and negative basis and appoints
Service type and index request are attempted and are adjusted by parameter adjustment experience;
It is replaced with the parameter after quantization and corresponds to parameter in original SNN, obtained parameter and quantify SNN;
Using the input of original SNN as input, parameter quantization SNN is tested, if test result meets index and wants
It asks, terminates, otherwise return to Selecting All Parameters interval division method again and section number, and interval division is carried out with after to parameter
Continuous process.
Beneficial effects of the present invention are that method of the invention can be converted into ANN SNN, and realize the parameter amount of SNN
Change, the quantization method is easy to operate flexibly to may be implemented the quantization of diversified forms, and to the performance of neural network almost without
Any influence can save storage resource, improve calculating speed.Especially when needing to realize SNN Hardwares, this method can
To reduce the consumption of the Resources on Chip such as RAM and computation complexity, hardware calculating speed and performance are improved.
Description of the drawings
Fig. 1 is a kind of SNN parameter quantification methods schematic diagram that present example provides;
Fig. 2 is one of ANN examples MLP schematic diagrames in Fig. 1;
Fig. 3 is one of ANN examples CNN schematic diagrames in Fig. 1;
Fig. 4 is one of ANN examples RNN schematic diagrames in Fig. 1;
Fig. 5 is one of ANN examples LSTM schematic diagrames in Fig. 1;
Fig. 6 is one of SNN examples self-organizing network schematic diagram in Fig. 1;
Fig. 7 is one of quantization parameter example weights distribution map and interval division situation in Fig. 1;
Fig. 8 is one of quantization parameter example threshold value distribution map and interval division situation in Fig. 2;
Fig. 9 is one of quantization parameter example leakage constant distribution map and interval division situation in Fig. 2;
Figure 10 is one of quantization parameter example set voltage distribution map and interval division situation in Fig. 2;
Figure 11 is one of quantization parameter example refractory period distribution map and interval division situation in Fig. 2;
Figure 12 is one of quantization parameter example synaptic delay distribution map and interval division situation in Fig. 2.
Figure 13 is a kind of exemplary method schematic diagram that SNN realizes parameter quantization in Fig. 1.
Specific implementation mode
The present invention is described in detail below in conjunction with the accompanying drawings, so that those skilled in the art more fully understands this hair
It is bright.Requiring particular attention is that in the following description, when perhaps known function and the detailed description of design can desalinate this
When the main contents of invention, these descriptions will be ignored herein.
As shown in Figure 1, a kind of SNN parameter quantification methods, include the following steps:
S1:Obtain the original SNN that training is completed.
Neuron is spiking neuron (such as LIF neurons) in SNN, has the function of that input pulse integrates accumulation and pulse
Provide function, SNN using pulse train as inputting, the major parameter of SNN include weights, threshold value, leakage constant, set voltage,
Refractory period, synaptic delay etc..The SNN that training is completed has the function of high-precision classification, identification etc..Obtain the nerve that training is completed
There are mainly two types of methods for network:One is the SNN that training completion is obtained by mapping offline, commonly random by training ANN
The ANN such as the methods of gradient decline training MLP (refering to Fig. 2), CNN (refering to Fig. 3), RNN (refering to Fig. 4), LSTM (refering to Fig. 5),
Acquisition completes training process after meeting the ANN of index request (such as classification, accuracy of identification etc.), then the ANN that training is completed
Parameter is mapped in the identical SNN of topological structure, by the input of ANN using pulse train coding (such as the pulse of Poisson distribution
Sequence) afterwards as the input of SNN, to obtain the SNN of training completion;One is obtain what training was completed by on-line training
SNN, establishes the SNN such as self-organizing SNN (refering to Fig. 6) or other structures, and plasticity (STDP) is relied on using cynapse pulse sequence
Equal learning rules, pass through on-line study using pulse train (such as Poisson distribution pulse train, time encoding pulse train etc.)
SNN is trained, the parameters such as weights, threshold value, leakage constant, set voltage, refractory period, the synaptic delay of SNN are adjusted in training process,
Acquisition completes training process after meeting index request (such as classification, accuracy of identification etc.) SNN, and the power of SNN is fixed after training
The parameters such as value, threshold value, leakage constant, set voltage, refractory period, synaptic delay, to obtain the SNN of training completion.
S2:Choose the one or more parameters for needing to quantify.
A parameter can once be quantified, can also once quantify multiple parameters.The parameter that can quantify includes weights, threshold
Value, leakage constant, set voltage, refractory period, synaptic delay etc..
S3:A certain layer or certain several layers of or all layer parameter distribution situation are counted respectively.
For the parameter that some needs quantifies, counts a certain layers of SNN or certain several layers of or all layers parameter and draw ginseng
Number distribution map.
S4:It attempts to carry out interval division to parameter.
The division methods of Selecting All Parameters interval division method and section number, section can be used equal point-score, non-equal point-score, set
Believe interval division method etc., the method and number of demarcation interval according to specific neural network structure and task type and can refer to
Mark requires, and is attempted and is adjusted by parameter adjustment experience.Described in S3 on the basis of parameter distribution figure, using equal point-score to power
It is worth (refering to Fig. 7), threshold value (refering to Fig. 8), leakage constant (refering to Fig. 9), set voltage (refering to fig. 1 0), refractory period (refering to figure
11), the result after the parameters such as synaptic delay (refering to fig. 1 2) progress interval division provides in the accompanying drawings.
S5:Parameter is quantified according to section.
The parameter in all sections is traversed, all parameters being distributed in same section are quantified as the same value (i.e. one
A quantized value), the size of quantized value and the specific neural network structure of positive and negative basis and task type and index request pass through
Parameter adjustment experience is attempted and is adjusted.
S6:It obtains parameter and quantifies SNN.
It is replaced with the parameter after quantization and corresponds to parameter in original SNN, obtained parameter and quantify SNN.
S7:Test parameter quantifies SNN.
Using the input of original SNN as input, parameter quantization SNN is tested, if test result meets index and wants
It asks, terminates, otherwise return to Selecting All Parameters interval division method again and section number, and interval division is carried out with after to parameter
Continuous process.
Refering to fig. 13, the quantization method is carried out for realizing MNIST Handwritten Digit Recognition tasks using MLP below
Further explanation includes the following steps:
S8:Set target identification accuracy.
That is index request described in SI, target identification accuracy of the setting nerve network system to MNIST test sets.
S9:MLP is trained using BP algorithm and trained weights are mapped directly into SNN.
That is off-line training method described in S1 needs to meet some specific conditions when training MLP in this example:(1)
All units of MLP all should use Relu functions as activation primitive;(2) in the training process, the deviation of neuron is fixed
It is 0.
S10:The threshold value and maximum frequency of SNN are set.
This is the peculiar parameter of SNN.The input of SNN needs to change into impulse form, so needing to carry out input picture
Coding, in this example encodes each pixel of picture using the mode of Poisson distribution frequency coding, the frequency of pulse
It is proportionate with the size of input pixel.The maximum frequency of pulse and the threshold value of LIF neurons according to the sizes of mapping parameters and
The discrimination of subsequent feedback is attempted and is adjusted by parameter adjustment experience.
S11:Test SNN discriminations.
It carries out in next step, otherwise returning to S8 re -trainings MLP and follow-up mistake if meeting target identification accuracy described in S8
Journey.
S12:Obtain all layers of weights distribution map and the respectively section of weights.
That is it attempts to carry out interval division to parameter described in all layer parameter distribution situations of statistics and S4 described in S3.In this example
Statistics and interval division only are carried out to weights, attempt to be divided using equal point-score, section number is 4.
S13:Weights are quantified and (attempt to be quantified using interval midpoint) according to section.
That is parameter is quantified according to section described in S5.It attempts to be quantified using interval midpoint in this example.
S14:Traversal maximizing wmax and minimum value wmin is carried out to all weights of SNN, takes putting down for wmax and wmin
Mean value is denoted as w0.
The step of being quantified using interval midpoint.
S15:The average value of wmin and w0 is taken to be denoted as w-1 again;The average value of wmax and w0 is taken to be denoted as w1.
The step of being quantified using interval midpoint.
S16:All weights are traversed again, if weights between wmin and w-1, enable weights be equal to the flat of the two
Mean value x1;If between w-1 and w0, enables weights be equal to the average value x2 of the two, similarly obtain x3 and x4.
The step of being quantified using interval midpoint.
S17:Test SNN discriminations.
That is test parameter described in S7 quantifies SNN.Terminate if meeting performance indicator, otherwise returns to S12 and choose area again
Between division methods and subsequent process.
Claims (4)
1. a kind of impulsive neural networks parameter quantification method, which is characterized in that include the following steps:
S1, the original pulse neural network that training is completed is obtained, the parameter of original pulse neural network includes weights, threshold value, lets out
Leak constant, set voltage, refractory period, synaptic delay;
S2, one or more parameter for needing to quantify is chosen;
Distribution situation of the parameter in neural network selected by S3, statistics;
S4, interval division is carried out to selected parameter;
S5, parameter is quantified according to section, that is, will be distributed over the parameter in same section and is quantified as the same value;
S6, parameter quantification impulse neural network is obtained, that is, the quantized value obtained in step 5 is used to substitute original pulse neural network
It can only strange corresponding initial parameter;
S7, the step S6 parameter quantification impulse neural networks obtained are tested:Using the defeated of primitivation impulsive neural networks
Enter as input, parameter quantization neural network is tested, meets if desired indicator requires if test result and terminate, otherwise return
Return step S3.
2. a kind of impulsive neural networks parameter quantification method according to claim 1, which is characterized in that described in step S1
Obtaining the specific method of original pulse neural network that training is completed is:
By the corresponding artificial neural network of training, the artificial neural network is multilayer perceptron, convolutional neural networks, cycle
One kind in neural network, shot and long term memory network, then training parameter is mapped to the identical impulsive neural networks of topological structure,
Input data is encoded using pulse train, obtains original pulse neural network;
Alternatively,
Impulsive neural networks are established, using pulse train as inputting, the learning machine of plasticity is relied on using cynapse pulse sequence
On-line training impulsive neural networks processed, the parameter of fixed pulse neural network after training, to obtain the original of training completion
Initial pulse neural network.
3. a kind of impulsive neural networks parameter quantification method according to claim 2, which is characterized in that the step S3's
Specific method is:
Parameter distribution situation of the selected parameter of statistics in a certain layer network;
Alternatively,
The parameter that Selecting All Parameters are counted in a few layer networks distinguishes situation;
Again alternatively,
The parameter that Selecting All Parameters are counted in each layer network distinguishes situation.
4. a kind of impulsive neural networks parameter quantification method according to claim 3, which is characterized in that the step S4's
Specific method is:
Parameter section is divided using a kind of in equal point-score, non-equal point-score, confidence interval partitioning, and obtains interval number
Mesh.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810501442.2A CN108717570A (en) | 2018-05-23 | 2018-05-23 | A kind of impulsive neural networks parameter quantification method |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810501442.2A CN108717570A (en) | 2018-05-23 | 2018-05-23 | A kind of impulsive neural networks parameter quantification method |
Publications (1)
Publication Number | Publication Date |
---|---|
CN108717570A true CN108717570A (en) | 2018-10-30 |
Family
ID=63900490
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810501442.2A Pending CN108717570A (en) | 2018-05-23 | 2018-05-23 | A kind of impulsive neural networks parameter quantification method |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108717570A (en) |
Cited By (25)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109635938A (en) * | 2018-12-29 | 2019-04-16 | 电子科技大学 | A kind of autonomous learning impulsive neural networks weight quantization method |
CN110059822A (en) * | 2019-04-24 | 2019-07-26 | 苏州浪潮智能科技有限公司 | One kind compressing quantization method based on channel packet low bit neural network parameter |
CN110364232A (en) * | 2019-07-08 | 2019-10-22 | 河海大学 | It is a kind of based on memristor-gradient descent method neural network Strength of High Performance Concrete prediction technique |
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 |
WO2020155741A1 (en) * | 2019-01-29 | 2020-08-06 | 清华大学 | Fusion structure and method of convolutional neural network and pulse neural network |
CN112085190A (en) * | 2019-06-12 | 2020-12-15 | 上海寒武纪信息科技有限公司 | Neural network quantitative parameter determination method and related product |
WO2020253692A1 (en) * | 2019-06-17 | 2020-12-24 | 浙江大学 | Quantification method for deep learning network parameters |
WO2021036890A1 (en) * | 2019-08-23 | 2021-03-04 | 安徽寒武纪信息科技有限公司 | Data processing method and apparatus, computer device, and storage medium |
WO2021036908A1 (en) * | 2019-08-23 | 2021-03-04 | 安徽寒武纪信息科技有限公司 | Data processing method and apparatus, computer equipment and storage medium |
CN113111758A (en) * | 2021-04-06 | 2021-07-13 | 中山大学 | SAR image ship target identification method based on pulse neural network |
CN113111997A (en) * | 2020-01-13 | 2021-07-13 | 中科寒武纪科技股份有限公司 | Method, apparatus and computer-readable storage medium for neural network data quantization |
CN113974607A (en) * | 2021-11-17 | 2022-01-28 | 杭州电子科技大学 | Sleep snore detecting system based on impulse neural network |
US11397579B2 (en) | 2018-02-13 | 2022-07-26 | Shanghai Cambricon Information Technology Co., Ltd | Computing device and method |
US11437032B2 (en) | 2017-09-29 | 2022-09-06 | Shanghai Cambricon Information Technology Co., Ltd | Image processing apparatus and method |
US11442785B2 (en) | 2018-05-18 | 2022-09-13 | Shanghai Cambricon Information Technology Co., Ltd | Computation method and product thereof |
US11513586B2 (en) | 2018-02-14 | 2022-11-29 | Shanghai Cambricon Information Technology Co., Ltd | Control device, method and equipment for processor |
US11544059B2 (en) | 2018-12-28 | 2023-01-03 | Cambricon (Xi'an) Semiconductor Co., Ltd. | Signal processing device, signal processing method and related products |
US11609760B2 (en) | 2018-02-13 | 2023-03-21 | Shanghai Cambricon Information Technology Co., Ltd | Computing device and method |
US11630666B2 (en) | 2018-02-13 | 2023-04-18 | Shanghai Cambricon Information Technology Co., Ltd | Computing device and method |
US11676028B2 (en) | 2019-06-12 | 2023-06-13 | Shanghai Cambricon Information Technology Co., Ltd | Neural network quantization parameter determination method and related products |
US11703939B2 (en) | 2018-09-28 | 2023-07-18 | Shanghai Cambricon Information Technology Co., Ltd | Signal processing device and related products |
US11762690B2 (en) | 2019-04-18 | 2023-09-19 | Cambricon Technologies Corporation Limited | Data processing method and related products |
US11789847B2 (en) | 2018-06-27 | 2023-10-17 | Shanghai Cambricon Information Technology Co., Ltd | On-chip code breakpoint debugging method, on-chip processor, and chip breakpoint debugging system |
US11847554B2 (en) | 2019-04-18 | 2023-12-19 | Cambricon Technologies Corporation Limited | Data processing method and related products |
US11966583B2 (en) | 2018-08-28 | 2024-04-23 | Cambricon Technologies Corporation Limited | Data pre-processing method and device, and related computer device and storage medium |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106022614A (en) * | 2016-05-22 | 2016-10-12 | 广州供电局有限公司 | Data mining method of neural network based on nearest neighbor clustering |
CN107704917A (en) * | 2017-08-24 | 2018-02-16 | 北京理工大学 | A kind of method of effectively training depth convolutional neural networks |
-
2018
- 2018-05-23 CN CN201810501442.2A patent/CN108717570A/en active Pending
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106022614A (en) * | 2016-05-22 | 2016-10-12 | 广州供电局有限公司 | Data mining method of neural network based on nearest neighbor clustering |
CN107704917A (en) * | 2017-08-24 | 2018-02-16 | 北京理工大学 | A kind of method of effectively training depth convolutional neural networks |
Cited By (42)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US11437032B2 (en) | 2017-09-29 | 2022-09-06 | Shanghai Cambricon Information Technology Co., Ltd | Image processing apparatus and method |
US11663002B2 (en) | 2018-02-13 | 2023-05-30 | Shanghai Cambricon Information Technology Co., Ltd | Computing device and method |
US11609760B2 (en) | 2018-02-13 | 2023-03-21 | Shanghai Cambricon Information Technology Co., Ltd | Computing device and method |
US11740898B2 (en) | 2018-02-13 | 2023-08-29 | Shanghai Cambricon Information Technology Co., Ltd | Computing device and method |
US11720357B2 (en) | 2018-02-13 | 2023-08-08 | Shanghai Cambricon Information Technology Co., Ltd | Computing device and method |
US11630666B2 (en) | 2018-02-13 | 2023-04-18 | Shanghai Cambricon Information Technology Co., Ltd | Computing device and method |
US11704125B2 (en) | 2018-02-13 | 2023-07-18 | Cambricon (Xi'an) Semiconductor Co., Ltd. | Computing device and method |
US11397579B2 (en) | 2018-02-13 | 2022-07-26 | Shanghai Cambricon Information Technology Co., Ltd | Computing device and method |
US11507370B2 (en) | 2018-02-13 | 2022-11-22 | Cambricon (Xi'an) Semiconductor Co., Ltd. | Method and device for dynamically adjusting decimal point positions in neural network computations |
US11709672B2 (en) | 2018-02-13 | 2023-07-25 | Shanghai Cambricon Information Technology Co., Ltd | Computing device and method |
US11620130B2 (en) | 2018-02-13 | 2023-04-04 | Shanghai Cambricon Information Technology Co., Ltd | Computing device and method |
US11513586B2 (en) | 2018-02-14 | 2022-11-29 | Shanghai Cambricon Information Technology Co., Ltd | Control device, method and equipment for processor |
US11442786B2 (en) | 2018-05-18 | 2022-09-13 | Shanghai Cambricon Information Technology Co., Ltd | Computation method and product thereof |
US11442785B2 (en) | 2018-05-18 | 2022-09-13 | Shanghai Cambricon Information Technology Co., Ltd | Computation method and product thereof |
US11789847B2 (en) | 2018-06-27 | 2023-10-17 | Shanghai Cambricon Information Technology Co., Ltd | On-chip code breakpoint debugging method, on-chip processor, and chip breakpoint debugging system |
US11966583B2 (en) | 2018-08-28 | 2024-04-23 | Cambricon Technologies Corporation Limited | Data pre-processing method and device, and related computer device and storage medium |
US11703939B2 (en) | 2018-09-28 | 2023-07-18 | Shanghai Cambricon Information Technology Co., Ltd | Signal processing device and related products |
US11544059B2 (en) | 2018-12-28 | 2023-01-03 | Cambricon (Xi'an) Semiconductor Co., Ltd. | Signal processing device, signal processing method and related products |
CN109635938A (en) * | 2018-12-29 | 2019-04-16 | 电子科技大学 | A kind of autonomous learning impulsive neural networks weight quantization method |
CN109635938B (en) * | 2018-12-29 | 2022-05-17 | 电子科技大学 | Weight quantization method for autonomous learning impulse neural network |
WO2020155741A1 (en) * | 2019-01-29 | 2020-08-06 | 清华大学 | Fusion structure and method of convolutional neural network and pulse neural network |
US11847554B2 (en) | 2019-04-18 | 2023-12-19 | Cambricon Technologies Corporation Limited | Data processing method and related products |
US11762690B2 (en) | 2019-04-18 | 2023-09-19 | Cambricon Technologies Corporation Limited | Data processing method and related products |
US11934940B2 (en) | 2019-04-18 | 2024-03-19 | Cambricon Technologies Corporation Limited | AI processor simulation |
CN110059822A (en) * | 2019-04-24 | 2019-07-26 | 苏州浪潮智能科技有限公司 | One kind compressing quantization method based on channel packet low bit neural network parameter |
CN112085190B (en) * | 2019-06-12 | 2024-04-02 | 上海寒武纪信息科技有限公司 | Method for determining quantization parameter of neural network and related product |
US11675676B2 (en) | 2019-06-12 | 2023-06-13 | Shanghai Cambricon Information Technology Co., Ltd | Neural network quantization parameter determination method and related products |
US11676029B2 (en) | 2019-06-12 | 2023-06-13 | Shanghai Cambricon Information Technology Co., Ltd | Neural network quantization parameter determination method and related products |
US11676028B2 (en) | 2019-06-12 | 2023-06-13 | Shanghai Cambricon Information Technology Co., Ltd | Neural network quantization parameter determination method and related products |
CN112085190A (en) * | 2019-06-12 | 2020-12-15 | 上海寒武纪信息科技有限公司 | Neural network quantitative parameter determination method and related product |
WO2020253692A1 (en) * | 2019-06-17 | 2020-12-24 | 浙江大学 | Quantification method for deep learning network parameters |
CN110364232B (en) * | 2019-07-08 | 2021-06-11 | 河海大学 | High-performance concrete strength prediction method based on memristor-gradient descent method neural network |
CN110364232A (en) * | 2019-07-08 | 2019-10-22 | 河海大学 | It is a kind of based on memristor-gradient descent method neural network Strength of High Performance Concrete prediction technique |
WO2021036908A1 (en) * | 2019-08-23 | 2021-03-04 | 安徽寒武纪信息科技有限公司 | Data processing method and apparatus, computer equipment and storage medium |
WO2021036890A1 (en) * | 2019-08-23 | 2021-03-04 | 安徽寒武纪信息科技有限公司 | Data processing method and apparatus, computer device, and storage medium |
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 |
CN113111997B (en) * | 2020-01-13 | 2024-03-22 | 中科寒武纪科技股份有限公司 | Method, apparatus and related products for neural network data quantization |
CN113111997A (en) * | 2020-01-13 | 2021-07-13 | 中科寒武纪科技股份有限公司 | Method, apparatus and computer-readable storage medium for neural network data quantization |
CN113111758B (en) * | 2021-04-06 | 2024-01-12 | 中山大学 | SAR image ship target recognition method based on impulse neural network |
CN113111758A (en) * | 2021-04-06 | 2021-07-13 | 中山大学 | SAR image ship target identification method based on pulse neural network |
CN113974607A (en) * | 2021-11-17 | 2022-01-28 | 杭州电子科技大学 | Sleep snore detecting system based on impulse neural network |
CN113974607B (en) * | 2021-11-17 | 2024-04-26 | 杭州电子科技大学 | Sleep snore detecting system based on pulse neural network |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108717570A (en) | A kind of impulsive neural networks parameter quantification method | |
CN109635917B (en) | Multi-agent cooperation decision and training method | |
Stromatias et al. | Scalable energy-efficient, low-latency implementations of trained spiking deep belief networks on spinnaker | |
CN105095961B (en) | A kind of hybrid system of artificial neural network and impulsive neural networks | |
CN107247989A (en) | A kind of neural network training method and device | |
CN107358293A (en) | A kind of neural network training method and device | |
Cheng et al. | Predicting productivity loss caused by change orders using the evolutionary fuzzy support vector machine inference model | |
CN110222760B (en) | Quick image processing method based on winograd algorithm | |
CN106982359A (en) | A kind of binocular video monitoring method, system and computer-readable recording medium | |
CN105095967A (en) | Multi-mode neural morphological network core | |
CN109165730B (en) | State quantization network implementation method in cross array neuromorphic hardware | |
CN110223515B (en) | Vehicle track generation method | |
Zambrano et al. | Efficient computation in adaptive artificial spiking neural networks | |
KR970008532B1 (en) | Neural metwork | |
CN113706151A (en) | Data processing method and device, computer equipment and storage medium | |
CN107609637A (en) | A kind of combination data represent the method with the raising pattern-recognition precision of pseudo- reversal learning self-encoding encoder | |
CN110084371A (en) | Model iteration update method, device and computer equipment based on machine learning | |
CN115018039A (en) | Neural network distillation method, target detection method and device | |
Gao et al. | Road Traffic Freight Volume Forecast Using Support Vector Machine Combining Forecasting. | |
CN112446462A (en) | Generation method and device of target neural network model | |
AU2021100614A4 (en) | A novel regression prediction method for electronic nose based on broad learning system | |
CN108073985A (en) | A kind of importing ultra-deep study method for voice recognition of artificial intelligence | |
CN106357437A (en) | Web Service Qos prediction method based on multivariate time series | |
Lee et al. | Semi-supervised learning for spiking neural networks based on spike-timing-dependent plasticity | |
KR102535635B1 (en) | Neuromorphic computing device |
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 | ||
WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20181030 |
|
WD01 | Invention patent application deemed withdrawn after publication |