CN107861916A - A kind of method and apparatus for being used to perform nonlinear operation for neutral net - Google Patents
A kind of method and apparatus for being used to perform nonlinear operation for neutral net Download PDFInfo
- Publication number
- CN107861916A CN107861916A CN201711103463.0A CN201711103463A CN107861916A CN 107861916 A CN107861916 A CN 107861916A CN 201711103463 A CN201711103463 A CN 201711103463A CN 107861916 A CN107861916 A CN 107861916A
- Authority
- CN
- China
- Prior art keywords
- piecewise interval
- function
- input value
- intercept
- nonlinear
- 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
- G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
- G06F17/10—Complex mathematical operations
- G06F17/15—Correlation function computation including computation of convolution operations
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Computational Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Mathematical Analysis (AREA)
- Mathematical Optimization (AREA)
- Pure & Applied Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Algebra (AREA)
- Databases & Information Systems (AREA)
- Software Systems (AREA)
- General Engineering & Computer Science (AREA)
- Computing Systems (AREA)
- Complex Calculations (AREA)
Abstract
The present invention provides a kind of method for being used to perform nonlinear function computing in neutral net, including：1) according to the scope of the input value of the nonlinear function, multiple piecewise intervals are divided；2) each in the multiple piecewise interval is directed to, the independent variable using two endpoint value as the nonlinear function, calculates and obtains two dependent variables corresponding with the piecewise interval；3) using the described two dependent variables corresponding with the piecewise interval obtained and the described two independents variable of calculating as 2 points on linear function corresponding with the piecewise interval, the slope and intercept of calculating linear function corresponding with each piecewise interval；4) each in the multiple piecewise interval is directed to, stores corresponding slope and intercept, for performing the nonlinear function computing.
Description
Technical field
The present invention relates to the data processing of neutral net.
Background technology
With the development of artificial intelligence, increasing research begins to focus on what is calculated using neural network model
The improvement of neural network processor.Neural network model imitate animal nervous system, layer by layer to the information of input at
The result that reason it is expected to obtain to level off to.In neural network model, each layer each data are required to by non-linear
Activation primitive handled, such as sigmoid functions, this to include during the complete computation of neural network model
Extremely large amount of nonlinear operation.
Here nonlinear function belongs to a kind of functional form, and it, which is plotted in, shows as curve or broken line on coordinate plane
Form.Generally comprised in the calculating of nonlinear function multiplying, division arithmetic, exponent arithmetic, trigonometric function operation or
Combinations thereof.Computing for such a series of complex usually requires have very high complexity using ALU etc.
The circuit block of degree completes corresponding processing procedure.
For neutral net, when performing sigmoid functions, nonlinear operation occurs with very high frequency, if
That neural network processor still seriatim performs each above-mentioned nonlinear operation using traditional logic circuit, then can be direct
The computational efficiency of neural network processor is restricted, and may require that the substantial amounts of energy consumption of consumption to perform above-mentioned computing.
The content of the invention
Therefore, it is an object of the invention to overcome above-mentioned prior art the defects of, there is provided one kind is used in neutral net
The method for performing nonlinear function computing, including：
1) according to the scope of the input value of the nonlinear function, multiple piecewise intervals are divided；
2) each in the multiple piecewise interval is directed to, using two endpoint value as the nonlinear function
Independent variable, calculate and obtain corresponding with the piecewise interval two dependent variables；
3) the described two dependent variables corresponding with the piecewise interval obtained and described two independents variable are calculated by described
As 2 points on linear function corresponding with the piecewise interval, the oblique of linear function corresponding with each piecewise interval is calculated
Rate and intercept；
4) each in the multiple piecewise interval is directed to, stores corresponding slope and intercept, for performing
The nonlinear function computing.
Preferably, also included according to methods described, wherein step 3)：
If the nonlinear function is concave function, its worst error between the linear function is calculated, and use
Current intercept subtracts the result of the half of the worst error to be used as new intercept.
Preferably, included according to methods described, wherein step 1)：
For each layer network of neutral net, count the scope of the input value of activation primitive in the layer network, using as
For the scope of the need input value to be processed of the layer network.
Preferably, according to methods described, wherein nonlinear function corresponding with the nonlinear operation includes：F (x)=tanh (x), f (x)=max (0, x).
Preferably, according to methods described, wherein also including：
5) when performing the nonlinear function computing, according to the segment identifier residing for the input value of the nonlinear function
Between, the result that execution computing of linear function determined by slope corresponding with the piecewise interval and intercept is obtained is made
For the result of the nonlinear operation.
A kind of method for based on above-mentioned method, performing nonlinear function computing in neutral net, including：
A1) by the input value of the nonlinear function compared with the endpoint value of each piecewise interval stored, it is determined that
Piecewise interval residing for the input value of the nonlinear function；
A2 the oblique of the linear function corresponding with the piecewise interval residing for the input value of the nonlinear operation of storage) is obtained
Rate a and intercept b, the independent variable using the input value of the nonlinear operation as linear function y=ax+b, calculate accordingly because becoming
Measure using the output valve as the nonlinear function computing.
A kind of computing device that nonlinear function computing is performed for neural network processor, including：
Look-up table unit, for storing described in each in the multiple piecewise interval obtained by the above method
The slope and intercept of two endpoint values and linear function corresponding with each piecewise interval；
Matching unit, point stored for the input value according to the nonlinear function and the look-up table unit
The endpoint value in section section, corresponding piecewise interval is matched by the input value of the nonlinear function；
Computing unit, for matching the piecewise interval determined and the look-up table according to by the matching unit
The slope and intercept for the linear function corresponding with the piecewise interval that unit is stored, by the input value of the nonlinear function
As the dependent variable of the linear function, the computing of the linear function is performed.
A kind of method that activation primitive computing is performed in the calculating process of neutral net, including：
B1 the current Internet calculated) is performed for neutral net, determines its used activation primitive, and count
The scope of the input value of the activation primitive；
B2) according to the scope of the input value of the activation primitive obtained and the resolution ratio set, by the activation
The scope of the input value of function divides multiple piecewise intervals；
B3) be directed to the multiple piecewise interval in each, using two endpoint value and as with the activation
Argument of function, two dependent variables corresponding to obtaining are calculated, and described two independents variable and described two dependent variables are made
For 2 points on linear function corresponding with the piecewise interval, the slope and intercept of the linear function are calculated, for described
Piecewise interval storage calculates the slope obtained and the intercept；
B4) by the input value of the activation primitive compared with the endpoint value of each piecewise interval stored, really
Piecewise interval residing for the fixed input value, and determine therefrom that slope corresponding with the piecewise interval and intercept；
B5) using the input value of the activation primitive as linear function corresponding with identified slope and intercept
Independent variable, corresponding dependent variable is calculated using the output valve as the activation primitive.
Preferably, according to methods described, wherein step B3) include：
If the activation primitive is concave function, its worst error between the linear function is calculated, and use and work as
Preceding intercept subtracts the result of the half of the worst error to be used as new intercept.
A kind of computer-readable recording medium, wherein being stored with computer program, the computer program is when executed
For realizing the method as described in above-mentioned any one.
Compared with prior art, the advantage of the invention is that：
The invention provides a kind of nonlinear function computational methods based on piecewise approximation, using the thought of approximate calculation,
Using the method for piecewise linearity, function is divided into some sections, function is calculated by the way of linear approximation in each section
Value, the calculating of the complicated functions such as power operation in complicated function, division arithmetic and trigonometric function operation is eliminated, save energy damage
Consumption, the application scenarios for the low-power consumption high energy efficiency that is particularly suitable for use in.Also, method provided by the present invention, replaced using look-up table multiple
Miscellaneous computing, and dynamic load look-up table, reduce circuit area overhead while computation rate is improved, it is only necessary to set a small amount of
Adder and multiplier can be to complete various nonlinear operations.
Brief description of the drawings
Embodiments of the present invention is further illustrated referring to the drawings, wherein：
Fig. 1 is according to an embodiment of the invention to be realized for nonlinear operation in neutral net based on look-up table
Computational methods flow chart；
Fig. 2 is the calculating electricity according to an embodiment of the invention that nonlinear operation is performed for neural network processor
The module map on road；
Fig. 3 is the structural representation for the computing unit being directed to according to one embodiment of present invention in counting circuit described in Fig. 2
Figure.
Embodiment
The present invention is elaborated with reference to the accompanying drawings and detailed description.
As described in background technology, a large amount of nonlinear functions in neural network model be present, for example, Local Phase should
Normalize calculating, batch processing calculating, activation primitive of layer etc..Wherein, activation primitive is very important composition in neutral net
Part, activation primitive are that neutral net adds non-linear description so that neural network model can better adapt to data
Nonlinear characteristic.Conventional activation primitive includes sigmoid functionsF (x)=tanh (x), ReLU function, f
(x)=max (0, x) etc..For neutral net, each data of each of which Internet are required to by activation primitive
Handled, this causes neural network processor to need to perform substantial amounts of nonlinear operation.
Above-mentioned nonlinear operation is performed using substantial amounts of logic circuit in traditional neural network processor, if energy
Circuit devcie used in enough reductions, such as adder, multiplier, register etc. used in reduction, then can reduce circuit
Area and processor energy consumption.
Based on above-mentioned consideration, inventor proposes approx regard the computing for various nonlinear functions as countless
The individual linear operation compared with minizone, when neural network processor needs to perform nonlinear operation, fallen according to the value of input
The minizone entered performs linear operation to obtain approximate result of calculation to the value of input.It has been recognised by the inventors that above-mentioned will be non-linear
The mode that computing is converted into the linear operation in the respective cell is particularly suitable for calculating for neutral net.This be by
In the algorithm of neural network model has certain fault-tolerance for intermediate result caused by calculating in itself, and it can be to centre
As a result further working process is carried out, also will not shadow even if some operation results in calculating process are not fully accurate
Ring the result of calculation of final output.
With reference to figure 1, according to one embodiment of present invention, there is provided one kind is realized for neutral net based on look-up table
The computational methods of middle nonlinear operation.
The construction method of the look-up table is introduced first, including：
Step S1, the span of the independent variable of nonlinear function is determined, the span is divided into multiple segmentations
Section, and the end points of each piecewise interval is stored in the look-up table.
Here it is possible to the span of the independent variable is estimated based on experience value, can also be according to the service condition of reality
The size of the independent variable is counted.For example, using 6 Sigma's criterions, 99.99966% number in statistics total amount
The scope of value.
When dividing piecewise interval, the granularity of each piecewise interval can be set big to the demand of resolution ratio according to system
It is small, such as less segmentation granularity is set for high-resolution demand.
In demarcation interval, the span of the independent variable equably can be divided into multiple piecewise intervals, such as
Span [0,100] is evenly divided into 100 equal portions, the segmentation granularity of setting is 1, [0,1), [1,2) ... [99,100].
Different size of piecewise interval can also be divided according to the distribution density of argument data, for example, it is close for the distribution of independent variable
The higher part of degree sets less piecewise interval to obtain greater number of piecewise interval.
Step S2, for each piecewise interval, by two end points x0And x1As the independent variable of the nonlinear function,
Calculate dependent variable y corresponding to obtaining0And y1.For example, it is assumed that the nonlinear function isThen
And
Step S3, for each piecewise interval, by (x0,y0) and (x1,y1) it is considered as at 2 points on straight line y=ax+b, meter
Calculate parameter a and b.
Here parameter a and b is respectively linear function y=ax+b slope and intercept, by by 2 points of (x0,y0) and
(x1,y1) be brought into the linear function, it can calculate and obtain the slope and intercept.y0=ax0+ b and y1=ax1+ b,
It is possible thereby to calculate acquisition slopeAnd intercept
Step S4, if the nonlinear function is concave function, calculated for each piecewise interval in the piecewise interval
Worst error E between straight line y=ax+b interior nonlinear function and that step S2 is tried to achievemax, and by (b-Emax/ 2) conduct
New b values.
Here error calculation method can use any appropriate prior art, such as be commonly used most in curve matching
Ultimate range on the small square law calculating nonlinear function and the straight line between each point is to be used as the worst error
Emax。
Step S5, for each piecewise interval, corresponding slope a and intercept b is stored in the look-up table.
When based on look-up table come the calculating of nonlinear operation in realizing for neutral net, stored in the look-up table
Each piecewise interval endpoint value and for each piecewise interval linear function slope and intercept will be used in god
Perform the nonlinear operation through network processing unit, the piecewise interval according to residing for input value, will perform by with the segment identifier
Between result of the result that is obtained of the computing of linear function determined by corresponding slope and intercept as the nonlinear operation.
It can be seen that the lookup that nonlinear operation is performed for neural network processor is established in the above-described embodiments
Table, the scope of input value is divided into multiple piecewise intervals by it, and utilizes the dullness of the nonlinear functions such as activation primitive
Property and concavity and convexity adjust the parameter of the linear function for each piecewise interval, are deposited in a lookup table for each piecewise interval
The parameter of corresponding endpoint value and corresponding linear function is stored up, nonlinear operation is performed for neutral net.
For the above method, present invention also offers a kind of corresponding counting circuit for neural network processor.Ginseng
Fig. 2 is examined, according to one embodiment of present invention, the counting circuit includes：Matching unit, look-up table unit and computing unit.
Wherein, look-up table unit is used to store the endpoint value of each piecewise interval of the nonlinear function and is directed to
The slope a and intercept b of the linear function of each piecewise interval.Here can be stored in the form of parameter is to (a, b), for
Each piecewise interval stores two endpoint value and corresponding parameter to (a, b).
Matching unit is used for the endpoint value according to each piecewise interval stored in the look-up table unit, will be described non-thread
The input value of property function is matched to corresponding piecewise interval, and according to identified corresponding piecewise interval in the look-up table
Searched in unit and obtain parameter corresponding with the piecewise interval to (a, b).It is appreciated that can in look-up table in the present invention
Only to record interval endpoint and its corresponding approximation linear function value, it is therefore desirable to by function input value and corresponding segmentation
Section is matched, further to be searched and the piecewise interval in look-up table unit according to the specific piecewise interval
The slope and intercept of corresponding linear function.
Computing unit is used for the slope and intercept of the linear function according to determined by matching unit, completes the meter of linear function
Calculation process.In one embodiment of the invention, with reference to figure 3, a multiplication unit and an addition are included in computing unit
Unit, wherein, the multiplication operation that multiplication unit is used to complete in linear function calculating process, adder unit is used to complete linear letter
Add operation in number calculating process.Multiplication unit receives the parameter a in the look-up table unit outside computing unit and come
The function input value x of matching unit output outside from computing unit, and parameter a and function input value x as multiplier and are multiplied
Number carries out multiplication operation, obtains product p；Adder unit receives output result p from multiplication unit and outside computing unit
Parameter b in portion's look-up table, and numerical value p is added with numerical value a, obtain result of calculation.By above-mentioned computing unit, can utilize
The input value x for nonlinear function obtained from matching unit piecewise interval, obtained from look-up table unit corresponding
Parameter performs multiplication and add operation, to be calculated closely by computing unit to (a, b) according to the parameter (a, b) of linear function
The functional value of the nonlinear function is similar to, and exports the functional value.
The look-up table configuration for neural network processor provided according to above-described embodiment, computation of table lookup can be used
Method replace NONLINEAR CALCULATION in neutral net, it is real to pass through the dynamic load look-up table content during neural computing
Existing different calculating function, while improve universality of the processor for a variety of calculating functions.
Also, dynamically loaded during the use of neural network processor for performing present invention also offers a kind of
The method of the parameter of nonlinear operation.On the one hand methods described can update the fortune of look-up tables'implementation different functions by dynamic
Calculate, on the other hand can also be directed to heterogeneous networks layer in same neural network model, dynamic updates when calculating heterogeneous networks layer
Different parameters.
According to one embodiment of present invention, include for the workflow of the counting circuit of activation primitive：
Step 1, activation primitive type is determined according to neural network model algorithm.
Step 2, training neutral net obtains parameters in network model, and counts activation primitive in every layer network and input
The scope of value, using the quasi- side of 6 Sigmas, the scope of the numerical value of Ji Lushuojuzongliang 99.99966%.
Step 3, neutral net Internet is determined.
Step 4, some sections will be divided into function computer capacity, interval division method needs to consider look-up table sky
Between the factor such as size and calculating resolution.
Step 5, current layer activation primitive look-up table parameter is loaded.
Step 6, the parameter a of linear function in each section is determined using look-up table section provided by the invention construction method
And b, and be loaded into look-up table.
Step 7, when need carry out activation primitive calculate be to access to function input value x in matching unit, obtain with it is defeated
Enter function section corresponding to value.
Step 8, the linear function parameter a and parameter b in function section are obtained in a lookup table according to function section.
Step 9, function input value, linear function parameter a and linear function parameter b are inputted into computing unit, completed
Calculate.
Step 10, after the calculating of current layer is completed, undated parameter a and b again, into next Internet.
By the above embodiment of the present invention as can be seen that the invention provides a kind of non-linear letter based on piecewise approximation
Number calculating method, using the thought of approximate calculation, using the method for piecewise linearity, function is divided into some sections, in each area
It is interior that functional value is calculated by the way of linear approximation, eliminate power operation in complicated function, division arithmetic and trigonometric function fortune
The complicated functions such as calculation calculate, and save energy loss, the application scenarios for the low-power consumption high energy efficiency that is particularly suitable for use in.Also, the present invention
The method provided, complex calculation, and dynamic load look-up table are replaced using look-up table, reduced while computation rate is improved
Circuit area overhead.
It should be noted that each step introduced in above-described embodiment is all not necessary, those skilled in the art
Appropriate choice, replacement, modification etc. can be carried out according to being actually needed.
It should be noted last that the above embodiments are merely illustrative of the technical solutions of the present invention and it is unrestricted.On although
The present invention is described in detail with reference to embodiment for text, it will be understood by those within the art that, to the skill of the present invention
Art scheme is modified or equivalent substitution, and without departure from the spirit and scope of technical solution of the present invention, it all should cover at this
Among the right of invention.
Claims (10)
1. a kind of method for being used to perform nonlinear function computing in neutral net, including：
1) according to the scope of the input value of the nonlinear function, multiple piecewise intervals are divided；
2) be directed to each in the multiple piecewise interval, using two endpoint value as the nonlinear function from
Variable, calculate and obtain two dependent variables corresponding with the piecewise interval；
3) using it is described calculate obtain described two dependent variables corresponding with the piecewise interval and described two independents variable as
2 points on linear function corresponding with the piecewise interval, calculate corresponding with each piecewise interval linear function slope and
Intercept；
4) each in the multiple piecewise interval is directed to, corresponding slope and intercept are stored, for described in execution
Nonlinear function computing.
2. according to the method for claim 1, wherein step 3) also includes：
If the nonlinear function is concave function, its worst error between the linear function is calculated, and using current
Intercept subtracts the result of the half of the worst error to be used as new intercept.
3. according to the method for claim 1, wherein step 1) includes：
For each layer network of neutral net, count the scope of the input value of activation primitive in the layer network, using as
The scope of the need input value to be processed of the layer network.
4. according to the method described in any one in claim 1-3, wherein non-linear letter corresponding with the nonlinear operation
Number includes：F (x)=tanh (x), f (x)=max (0, x).
5. according to the method described in any one in claim 1-3, wherein also including：
5), will according to the piecewise interval residing for the input value of the nonlinear function when performing the nonlinear function computing
The result that execution computing of linear function determined by slope corresponding with the piecewise interval and intercept is obtained is as institute
State the result of nonlinear operation.
6. a kind of method in 1-4 based on claim described in any one, perform nonlinear function computing in neutral net
Method, including：
A1) by the input value of the nonlinear function compared with the endpoint value of each piecewise interval stored, it is determined that described
Piecewise interval residing for the input value of nonlinear function；
A2 the slope a of the linear function corresponding with the piecewise interval residing for the input value of the nonlinear operation of storage) is obtained
With intercept b, the independent variable using the input value of the nonlinear operation as linear function y=ax+b, corresponding dependent variable is calculated
Using the output valve as the nonlinear function computing.
7. a kind of computing device that nonlinear function computing is performed for neural network processor, including：
Look-up table unit, pass through for storing in the multiple piecewise interval that any one method obtains in claim 1-4
The described two endpoint values of each and linear function corresponding with each piecewise interval slope and intercept；
Matching unit, the segment identifier stored for the input value according to the nonlinear function and the look-up table unit
Between endpoint value, the input value of the nonlinear function is matched into corresponding piecewise interval；
Computing unit, for matching the piecewise interval determined and the look-up table unit according to by the matching unit
The slope and intercept of the linear function corresponding with the piecewise interval stored, using the input value of the nonlinear function as
The dependent variable of the linear function, perform the computing of the linear function.
8. a kind of method that activation primitive computing is performed in the calculating process of neutral net, including：
B1 the current Internet calculated) is performed for neutral net, determines its used activation primitive, and described in statistics
The scope of the input value of activation primitive；
B2) according to the scope of the input value of the activation primitive obtained and the resolution ratio set, by the activation primitive
The scope of input value divide multiple piecewise intervals；
B3) be directed to the multiple piecewise interval in each, using two endpoint value and as with the activation primitive
Independent variable, calculate two dependent variables corresponding to obtaining, and using described two independents variable and described two dependent variables as with
, the slope and intercept of the linear function are calculated, for the segmentation at 2 points corresponding to the piecewise interval on linear function
Section storage calculates the slope obtained and the intercept；
B4 the input value of the activation primitive) is determined into institute compared with the endpoint value of each piecewise interval stored
The piecewise interval residing for input value is stated, and determines therefrom that slope corresponding with the piecewise interval and intercept；
B5) becoming certainly using the input value of the activation primitive as linear function corresponding with identified slope and intercept
Amount, calculates corresponding dependent variable using the output valve as the activation primitive.
9. according to the method for claim 8, wherein step B3) include：
If the activation primitive is concave function, its worst error between the linear function is calculated, and cut using current
Result away from the half for subtracting the worst error is to be used as new intercept.
10. a kind of computer-readable recording medium, wherein being stored with computer program, the computer program is used when executed
In method of the realization as described in any one in claim 1-6,8,9.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201711103463.0A CN107861916A (en) | 2017-11-10 | 2017-11-10 | A kind of method and apparatus for being used to perform nonlinear operation for neutral net |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201711103463.0A CN107861916A (en) | 2017-11-10 | 2017-11-10 | A kind of method and apparatus for being used to perform nonlinear operation for neutral net |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN107861916A true CN107861916A (en) | 2018-03-30 |
Family
ID=61701314
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201711103463.0A Pending CN107861916A (en) | 2017-11-10 | 2017-11-10 | A kind of method and apparatus for being used to perform nonlinear operation for neutral net |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107861916A (en) |
Cited By (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108804883A (en) * | 2018-04-27 | 2018-11-13 | 北京奇艺世纪科技有限公司 | A kind of operation method and device of obfuscated codes |
| CN108898216A (en) * | 2018-05-04 | 2018-11-27 | 中国科学院计算技术研究所 | Activation processing unit applied to neural network |
| CN108921288A (en) * | 2018-05-04 | 2018-11-30 | 中国科学院计算技术研究所 | Neural network activates processing unit and the neural network processor based on the device |
| CN108920748A (en) * | 2018-05-22 | 2018-11-30 | 复旦大学 | A kind of non-linear arithmetic circuit structure with high flexibility and precision |
| CN109542912A (en) * | 2018-12-04 | 2019-03-29 | 北京锐安科技有限公司 | Interval censored data storage method, device, server and storage medium |
| CN109657788A (en) * | 2018-12-18 | 2019-04-19 | 北京中科寒武纪科技有限公司 | Data processing method, device and Related product |
| CN109871941A (en) * | 2019-02-18 | 2019-06-11 | 北京中科寒武纪科技有限公司 | Data processing method, device and Related product |
| CN110058841A (en) * | 2019-04-22 | 2019-07-26 | 南京大学 | Towards nonlinear function general-purpose calculating appts and method with symmetry |
| CN110210612A (en) * | 2019-05-14 | 2019-09-06 | 北京中科汇成科技有限公司 | A kind of integrated circuit accelerated method and system based on dispositif de traitement lineaire adapte approximating curve |
| CN110866595A (en) * | 2018-08-28 | 2020-03-06 | 北京嘉楠捷思信息技术有限公司 | Method, device and circuit for operating activation function in integrated circuit |
| CN113378013A (en) * | 2021-06-30 | 2021-09-10 | 北京思朗科技有限责任公司 | Data processing method, device and system |
| CN114567396A (en) * | 2022-02-28 | 2022-05-31 | 哲库科技(北京)有限公司 | Wireless communication method, fitting method of nonlinear function, terminal and equipment |
| CN114969631A (en) * | 2022-05-26 | 2022-08-30 | Oppo广东移动通信有限公司 | Baseband chip, channel estimation method, data processing method and equipment |
-
2017
- 2017-11-10 CN CN201711103463.0A patent/CN107861916A/en active Pending
Cited By (22)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108804883A (en) * | 2018-04-27 | 2018-11-13 | 北京奇艺世纪科技有限公司 | A kind of operation method and device of obfuscated codes |
| CN108804883B (en) * | 2018-04-27 | 2020-06-19 | 北京奇艺世纪科技有限公司 | Method and device for running obfuscated code |
| CN108898216A (en) * | 2018-05-04 | 2018-11-27 | 中国科学院计算技术研究所 | Activation processing unit applied to neural network |
| CN108921288A (en) * | 2018-05-04 | 2018-11-30 | 中国科学院计算技术研究所 | Neural network activates processing unit and the neural network processor based on the device |
| CN108920748A (en) * | 2018-05-22 | 2018-11-30 | 复旦大学 | A kind of non-linear arithmetic circuit structure with high flexibility and precision |
| CN108920748B (en) * | 2018-05-22 | 2023-01-06 | 复旦大学 | Nonlinear operation circuit structure with high flexibility and precision |
| CN110866595A (en) * | 2018-08-28 | 2020-03-06 | 北京嘉楠捷思信息技术有限公司 | Method, device and circuit for operating activation function in integrated circuit |
| CN110866595B (en) * | 2018-08-28 | 2024-04-26 | 嘉楠明芯(北京)科技有限公司 | Method, device and circuit for operating activation function in integrated circuit |
| CN109542912A (en) * | 2018-12-04 | 2019-03-29 | 北京锐安科技有限公司 | Interval censored data storage method, device, server and storage medium |
| CN109657788A (en) * | 2018-12-18 | 2019-04-19 | 北京中科寒武纪科技有限公司 | Data processing method, device and Related product |
| CN111126581A (en) * | 2018-12-18 | 2020-05-08 | 中科寒武纪科技股份有限公司 | Data processing method and device and related products |
| CN109871941B (en) * | 2019-02-18 | 2020-02-21 | 中科寒武纪科技股份有限公司 | Data processing method and device and related products |
| CN109871941A (en) * | 2019-02-18 | 2019-06-11 | 北京中科寒武纪科技有限公司 | Data processing method, device and Related product |
| CN110058841A (en) * | 2019-04-22 | 2019-07-26 | 南京大学 | Towards nonlinear function general-purpose calculating appts and method with symmetry |
| CN110058841B (en) * | 2019-04-22 | 2023-03-28 | 南京大学 | General computing device and method for nonlinear function with symmetry |
| CN110210612A (en) * | 2019-05-14 | 2019-09-06 | 北京中科汇成科技有限公司 | A kind of integrated circuit accelerated method and system based on dispositif de traitement lineaire adapte approximating curve |
| CN110210612B (en) * | 2019-05-14 | 2021-04-16 | 北京中科汇成科技有限公司 | Integrated circuit acceleration method and system based on self-adaptive piecewise linear approximation curve |
| CN113378013A (en) * | 2021-06-30 | 2021-09-10 | 北京思朗科技有限责任公司 | Data processing method, device and system |
| CN113378013B (en) * | 2021-06-30 | 2023-10-27 | 上海思朗科技有限公司 | Data processing method, device and system |
| CN114567396A (en) * | 2022-02-28 | 2022-05-31 | 哲库科技(北京)有限公司 | Wireless communication method, fitting method of nonlinear function, terminal and equipment |
| CN114969631A (en) * | 2022-05-26 | 2022-08-30 | Oppo广东移动通信有限公司 | Baseband chip, channel estimation method, data processing method and equipment |
| CN114969631B (en) * | 2022-05-26 | 2024-05-10 | Oppo广东移动通信有限公司 | Baseband chip, channel estimation method, data processing method and equipment |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107861916A (en) | A kind of method and apparatus for being used to perform nonlinear operation for neutral net | |
| CN107729984A (en) | A kind of computing device and method suitable for neutral net activation primitive | |
| CN105844330A (en) | Data processing method of neural network processor and neural network processor | |
| CN107862374A (en) | Processing with Neural Network system and processing method based on streamline | |
| CN110688088B (en) | General nonlinear activation function computing device and method for neural network | |
| CN109934331A (en) | Device and method for executing artificial neural network forward operation | |
| CN106126481A (en) | A kind of computing engines and electronic equipment | |
| CN108390393B (en) | Multi-target reactive power optimization method for power distribution network and terminal equipment | |
| CN107479856A (en) | Arctan function data structure and method for building up, function value-acquiring method and device | |
| CN108564592A (en) | Based on a variety of image partition methods for being clustered to differential evolution algorithm of dynamic | |
| CN109615068A (en) | The method and apparatus that feature vector in a kind of pair of model is quantified | |
| CN107589935A (en) | A kind of nonlinear function computing device and method | |
| CN101833691A (en) | Realizing method of least square support vector machine serial structure based on EPGA (Filed Programmable Gate Array) | |
| CN109325530B (en) | Image classification method, storage device and processing device | |
| CN109993279A (en) | A kind of double-deck same or binary neural network compression method calculated based on look-up table | |
| CN108256644A (en) | Microprocessor circuit and the method for performing neural network computing | |
| Ge et al. | BNReLU: Combine batch normalization and rectified linear unit to reduce hardware overhead | |
| CN111383157A (en) | Image processing method and device, vehicle-mounted operation platform, electronic equipment and system | |
| CN110837885B (en) | Sigmoid function fitting method based on probability distribution | |
| CN112668871A (en) | Dynamic assignment method for expert weight in multi-round group decision | |
| CN104933110B (en) | A kind of data prefetching method based on MapReduce | |
| CN113516171B (en) | Image classification method based on Bayes neural network random addition decomposition structure | |
| CN107590105B (en) | Computing device and method towards nonlinear function | |
| CN114298319A (en) | Method and device for determining joint learning contribution value, electronic equipment and storage medium | |
| CN108123893A (en) | A kind of multiple target bandwidth allocation methods of real time virtual machine migration |
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: 20180330 |