CN108205518A - Obtain device, method and the neural network device of functional value - Google Patents
Obtain device, method and the neural network device of functional value Download PDFInfo
- Publication number
- CN108205518A CN108205518A CN201611182655.0A CN201611182655A CN108205518A CN 108205518 A CN108205518 A CN 108205518A CN 201611182655 A CN201611182655 A CN 201611182655A CN 108205518 A CN108205518 A CN 108205518A
- Authority
- CN
- China
- Prior art keywords
- value
- function
- interpolation
- independent variable
- interpolating function
- 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
Abstract
The present invention provides a kind of device and method for obtaining functional value, wherein device includes：I/O modules, for the data after input data and output piecewise interpolation；Memory module, for storing interpolating function and function interpolation table；Searching module, for searching corresponding parameter value in function interpolation table according to the data value；And computing module, for according to interpolating function and parameter value, basic operations to be carried out, and calculate the functional value after interpolation to the data.When the device finds a function value, the range of argument value need to be only directed to, corresponding parameter is selected in function interpolation table and does the operation on basis.Therefore the invention simplifies hardware design, improves arithmetic speed, and reduce the area power dissipation ratio of chip.
Description
Technical field
The invention belongs to computer realms, and in particular to a kind of device for obtaining functional value utilizes piecewise interpolation mode
It obtains, further relates to obtain the method and neural network device of functional value.
Background technology
Arithmetic logic unit (arithmetic logic unit, abridge ALU) is the structure for carrying out integer arithmetic.It is counting
In calculation machine, ALU is the special digital circuit for performing arithmetic sum logical operation.ALU is most important group of central processing unit (CPU)
Also make tally function comprising ALU into part or even small processor.In modern CPU (Central Processing
Unit, central processing unit) and GPU (Graphics Processing Unit, graphics processor) in contained it is powerful and
Complicated ALU；One single element electricity may contain ALU.Most of ALU can complete following operation：Integer arithmetic is transported
Calculate (add, subtract, multiplying and except), position logical operation (with or non-sum exclusive or), shift operation (moves data or floating to the left or to the right
Dynamic certain bits, mobile 1 is equivalent to and is multiplied by 2 or divided by 2).Arithmetic logic unit usually contains only linear operational element, works as arithmetic
Logic unit carries out complicated power operation when operations, it usually needs several execution cycles.
In order to accelerate the arithmetic speed of processor, would generally be integrated in CPU and GPU FPU (Floating-Point Unit,
FPU Float Point Unit).FPU is used exclusively for the processor of floating-point operation, and may support some calculating surmounted function, example
Such as log (x).
The prior art is when calculating complicated function, such as various nonlinear functions, typically by complicated fortune
Simple calculations is disassembled into calculation, using can just obtain result after several execution cycles.This so that arithmetic speed is slow, arithmetic unit
Area is big, power consumption is high.
Invention content
(1) technical problems to be solved
The object of the present invention is to provide a kind of device, method and neural network devices for obtaining functional value.
(2) technical solution
According to an aspect of the present invention, a kind of device for obtaining functional value is provided, is obtained using piecewise interpolation mode, is wrapped
It includes：
I/O modules, for the data after input data and output piecewise interpolation；
Memory module, for storing interpolating function and function interpolation table；
Searching module, for searching corresponding parameter value in function interpolation table according to the data value；And
Computing module, for according to interpolating function and parameter value, basic operations to be carried out, and calculate interpolation to the data
Functional value afterwards.
Further, the interpolating function is linear interpolation function or polynomial interpolating function.
Further, the function interpolation table includes the parameter of the interpolating function.
According to another aspect of the present invention, a kind of method of piecewise interpolation is provided, it is right for obtaining its according to input data
The functional value answered, including step：
S1, input data go to step S2 as independent variable；
S2, the value range of independent variable are divided into N number of big section in advance：A1, A2..., AN, each big section is divided into again
M minizone, N and M are natural number, and independent variable is fallen in big section AiIn, i is obtained, initializes cycle indexed variable a p, p=
0, go to step S3；
S3 stores N sections of interpolation tables, searches the corresponding parameter value of pth section interpolation table, goes to step S4；
S4 goes out the corresponding interpolating function value of independent variable according to the parameter value calculation of independent variable and step S3, and cycle mark becomes
P=p+1 is measured, if judging the value of p --- p ＜ i go to step S5；Otherwise, S6 is gone to step；
S5 transmits operation result, goes to step S3；
Result of calculation is passed to I/O modules, goes to step S7 by S6；
S7, I/O module export result.
Further, in step S2, N number of big section is isometric or Length discrepancy.
Further, the function interpolation table includes the parameter of the interpolating function.
Further, the interpolating function is linear interpolation function or polynomial interpolating function.
In accordance with a further aspect of the present invention, a kind of neural network device is provided, is calculated using piecewise interpolation defeated with neuron
Enter activation primitive value of the inner product of value and weighted value as independent variable, including：
Memory, for storing executable instruction；
Processor, for performing the executable instruction stored in memory, to perform following operating procedure：
One, input data goes to step two as independent variable；
Two, the value range of independent variable is divided into N number of big section in advance：A1, A2..., AN, each big section is divided into again
M minizone, N and M are natural number, and independent variable is fallen in big section AiIn, i is obtained, initializes cycle indexed variable a p, p=
0, go to step three；
Three, according to the N section interpolation tables for being stored in memory, loading pth section interpolation table is searched, and is searched according to independent variable
Go out corresponding parameter value in function interpolation table, go to step four；
Four, corresponding interpolating function value is calculated according to parameter value and independent variable, indexed variable p=p+1 is recycled, judges p
Value --- if p ＜ i go to step three；Otherwise, five are gone to step；
Five, export interpolating function value.
Further, the processor includes CPU or GPU.
Further, the activation primitive is hyperbolic tangent function or Sigmoid functions
(3) advantageous effect
For the present invention by complicated function according to data area, piecewise fitting is simple interpolating function.When finding a function value, only
The range of argument value need to be directed to, corresponding parameter is selected in function interpolation table and does the operation on basis.Therefore the hair
It is bright to simplify hardware design, arithmetic speed is improved, and reduce the area power dissipation ratio of chip；It solves the prior art counting
Hardware configuration is complicated when calculating complicated function, and arithmetic speed is slow, the problems such as arithmetic unit area power dissipation ratio height.
Description of the drawings
Fig. 1 is the overall structure example block diagram according to the device for piecewise interpolation of one embodiment of the invention；
Fig. 2 is the method flow diagram according to the piecewise interpolation of one embodiment of the invention；
Fig. 3 is for the hardware basic circuit of piecewise interpolation and data transmission schematic diagram according to one embodiment of the invention；
Fig. 4 is in the enterprising line piecewise interpolation of fixed interval according to one embodiment of the invention to exponential function exp (x)
Interpolation schematic diagram；
Fig. 5 is the structure diagram according to the neural network device of one embodiment of the invention.
Specific embodiment
To make the objectives, technical solutions, and advantages of the present invention clearer, below in conjunction with specific embodiment, and reference
Attached drawing, the present invention is described in further detail.By described in detail below, other aspects of the invention, advantage and protrusion
Feature will become obvious those skilled in the art.
In the present specification, it is following only to illustrate for describing the various embodiments of the principle of the invention, it should not be with any
Mode is construed to the range of limitation invention.With reference to attached drawing the comprehensive understanding described below that is used to help by claim and its equivalent
The exemplary embodiment of the present invention that object limits.It is described below to help to understand including a variety of details, but these details should
Think to be only exemplary.Therefore, it will be appreciated by those of ordinary skill in the art that without departing substantially from scope and spirit of the present invention
In the case of, embodiment described herein can be made various changes and modifications.In addition, for clarity and brevity,
The description of known function and structure is omitted.In addition, through attached drawing, same reference numbers are used for identity function and operation.
One side according to embodiments of the present invention provides a kind of device for obtaining functional value, can press complicated function
According to data area, piecewise fitting is simple linear function.When finding a function value, searching module load store mould interpolation in the block
Table according to the range of argument value, finds out corresponding slope and intercept, and do basic operations (namely addition and multiplication fortune
It calculates).Which fallen in big section according to independent variable, cycle proceeds as described above, and finally obtains interpolation result, i.e. approximation obtains
Functional value.Therefore the invention simplifies hardware design, improves arithmetic speed, and reduce the area power dissipation ratio of chip.
Fig. 1 is the overall structure example block diagram according to the device for linear segmented interpolation of one embodiment of the invention.Such as
Shown in Fig. 1, which includes I/O modules As, searching module C, memory module B and computing module D.All modules can be by hard
The realization of part circuit, as shown in Figure 3.
I/O modules As, i.e. input/output module, the module are used for input data (independent variable) x1And pass it to lookup
It module C and final calculation result y and is output it for being received from computing module D.Here it mentions, x1It can be direct
It is initial data or initial data x0By pretreated data, in order to describe simplicity, do not refer to pretreated
Journey.
In memory module B, the interpolating function f of storage calculating process needs1, f2..., fN, fpCorresponding pth section interpolation letter
Number.Data area is divided into N number of big section A in advance1, A2..., AN, big section ApLeft and right endpoint respectively use inf Ap、sup
ApIt represents.Each big section ApIt is divided into as M minizonefpIt is defined as follows：
The module stores all slopesAnd interceptP=1,2 ..., N, q=1,2 ..., M+2.The value of M
To be determined according to data precision, M is bigger, and precision is higher, i.e., the functional value obtained by interpolation result approximation is closer to actual value.
In searching module C, data area is divided into N number of big section A in advance1, A2..., AN, i is obtained first so that from change
Measure x1It falls in AiIn.Later, the pth section interpolation table in (load) memory module, 1≤p≤i-1, to being used for pth time are loaded successively
The independent variable x of lookupp, search corresponding slopeAnd interceptAnd with independent variable xpIt is passed in computing module D together.And
The result of calculation x that module reception is obtained from computing module D afterwardsp+1As independent variable, the lookup of pth+1 time is done.Finally, load
I-th section of interpolation table in memory module does last time and searches.
In computing module D, the independent variable x obtained from searching module is receivedp, slopeAnd interceptIt calculatesIf 1≤p≤i-1, by result of calculation xp+1Searching module C is passed to do next time
It searches；If p=i, by result of calculation xi+1I/O modules are passed to as final output result y, i.e. y=xi+1。
Another aspect according to embodiments of the present invention provides a kind of method flow diagram for carrying out functional value acquisition.Fig. 2 is root
According to the carry out piecewise interpolation flow chart of one embodiment of the invention.This method can be applied to devices discussed above.Wherein, specific number
It is as shown in Figure 3 according to transfer process.
S1, I/O modules A are by data x1(independent variable) inputs, and passes it to searching module C, goes to step S2；
In S2, searching module C, i is obtained first so that independent variable x1It falls in big section AiIn.One cycle mark of initialization
Variable p, p=0.Go to step S3；
N sections of interpolation tables are stored in S3, memory module B.Pth section interpolation table in searching module C load memory modules B
It is searched, and by lookup result, i.e. corresponding slope in function interpolation tableAnd interceptWith independent variable xpIt passes together
Computing module D is passed, goes to step S4；
S4, computing module D calculate corresponding interpolating function value：
Indexed variable p=p+1 is recycled, if judging the value of p --- p ＜ i go to step S5；Otherwise, S6 is gone to step；
S5, by result of calculation Xp+1Pass to searching module C (at this time the result as independent variable participate in after lookup and
Calculate), go to step S3；
S6, by result of calculation xi+1I/O modules As are passed to, go to step S7；
S7, I/O modules A output result y=xi+1。
Certainly, interpolating function is not limited only to linear function in the above method, can also be multinomial into row interpolation, as long as
Complicated function is converted into the function of simple operation, including but not limited to linear function and multinomial letter by interpolation method
Number.
Specific embodiment is exemplified below to illustrate：
Embodiment 1
Linear segmented interpolation is carried out above in [0,18] to function F (x)=exp (x)：
First, data area is divided into 3 big section (i.e. N=3), A1=[0,10), A2=[10,15), A3=[15,18].It needs to refer to
Go out, there is no equably go to point 3 big sections here.Because the derived function of the bigger curve of argument value is also bigger, in other words curve
It is steeper, it is approximate fine in order to ensure, it is a little bit smaller in the precipitous local section of curve point, and that not needed in the local section of curve gently
It is small.Each big section is evenly divided into 10 minizones again：
Such as
Then, interpolating function f is provided1(x), f2(x), f3(x) definition：
Wherein, slopeAnd interceptValue rule be：In sectionLeft and right endpoint on, fp(x) value and F
(x) value of=exp (x) is equal.For example, in big section A2On interpolation it is as shown in Figure 4.
Finally, for given independent variable x1, according to shown in Fig. 2 and above-mentioned method and step carries out successively.
Embodiment 2
To being applied to the neural network of image classification, to activation primitive F (x)=sigmoid (x) in [0,255] (image ash
Spend range) linear segmented interpolation is carried out above：
First, data area is divided into 8 big section (i.e. N=8), A1=[0,10) [and 0,31), A2=[32,63), A3=
[64,95) ..., A8=[224,255].It should be pointed out that there is no equably go to point 8 big sections here.Because from becoming
The derived function of the bigger curve of magnitude is also bigger, and curve is steeper in other words, approximate fine in order to ensure, in the place that curve is precipitous
Section point is a little bit smaller, and the local section of curve gently need not be so small.Each big section can be evenly divided into 32 again
Or 64 minizones (determine, can also be the minizone of other quantity) interpolating function and embodiment 1 according to required precision
It is similar, wherein, slopeAnd interceptValue rule be：In sectionLeft and right endpoint on, fp(x) value and F (x)
The value of=sigmoid (x) is equal.
Finally, for given independent variable x1, carried out successively according to above-mentioned method and step.
Based on same inventive concept, the present invention also provides a kind of dedicated neural network device, using piecewise interpolation in people
In the feed forward operation of artificial neural networks, calculate using neuron input value and the inner product of weighted value as the activation primitive of independent variable.
Fig. 5 is the neural network device structure diagram according to one embodiment of the invention.The neural network device 100, profit
It is calculated by the use of piecewise interpolation using neuron input value and the inner product of weighted value as the activation primitive value of independent variable, including：
Memory 101, for storing executable instruction；
Processor 102, for performing the executable instruction stored in memory, to perform following operating procedure：
One, input data goes to step two as independent variable；
Two, the value range of independent variable is divided into N number of big section in advance：A1, A2..., AN, each big section is divided into again
M minizone, N and M are natural number, and independent variable is fallen in big section AiIn, i is obtained, initializes cycle indexed variable a p, p=
0, go to step three；
Three, according to the N section interpolation tables for being stored in memory, loading pth section interpolation table is searched, and is searched according to independent variable
Go out corresponding parameter value in function interpolation table, go to step four；
Four, corresponding interpolating function value is calculated according to parameter value and independent variable, indexed variable p=p+1 is recycled, judges p
Value --- if p ＜ i go to step three；Otherwise, five are gone to step；
Five, export interpolating function value.
The processor can include general purpose microprocessor, instruction set processor and/or related chip group and/or specially
With microprocessor (for example, application-specific integrated circuit (ASIC)).Processor can also include the onboard storage device for caching purposes.
Preferably, using dedicated neural network processor.
Processor be used to perform the different actions of the flow of the present embodiment description single treatment unit (such as CPU or
) or multiple processing units GPU.
For the operating procedure of execution, the method flow diagram that can refer to the piecewise interpolation described in Fig. 2 carries out.Wherein,
The activation primitive is hyperbolic tangent function or Sigmoid functions.
The device of the present embodiment can also include input-output unit 103, original or after pretreatment to input
Data, and export the functional value after interpolated operation.
Particular embodiments described above has carried out the purpose of the present invention, technical solution and advantageous effect further in detail
Describe in detail bright, it should be understood that the above is only a specific embodiment of the present invention, is not intended to restrict the invention, it is all
Within the spirit and principles in the present invention, any modification, equivalent substitution, improvement and etc. done should be included in the protection of the present invention
Within the scope of.
Claims (10)
1. a kind of device for obtaining functional value, is obtained using piecewise interpolation mode, it is characterised in that including：
I/O modules, for the data after input data and output piecewise interpolation；
Memory module, for storing interpolating function and function interpolation table；
Searching module, for searching corresponding parameter value in function interpolation table according to the data value；And
Computing module, for according to interpolating function and parameter value, basic operations to be carried out to the data, and after calculating interpolation
Functional value.
2. the apparatus according to claim 1, which is characterized in that the interpolating function is linear interpolation function or multinomial
Interpolating function.
3. the apparatus according to claim 1, which is characterized in that the function interpolation table includes the ginseng of the interpolating function
Number.
4. a kind of method for obtaining functional value, for obtaining its corresponding function by piecewise interpolation mode according to input data
Value, it is characterised in that including step：
S1, input data go to step S2 as independent variable；
S2, the value range of independent variable are divided into N number of big section in advance：A1, A2..., AN, each big section is divided into M again
Minizone, N and M are natural number, and independent variable is fallen in big section AiIn, i is obtained, initializes cycle indexed variable p, a p=0,
Go to step S3；
S3 stores N sections of interpolation tables, searches the corresponding parameter value of pth section interpolation table, goes to step S4；
S4 goes out the corresponding interpolating function value of independent variable, cycle indexed variable p=according to the parameter value calculation of independent variable and step S3
P+1, if judging the value of p --- p ＜ i go to step S5；Otherwise, S6 is gone to step；
S5 transmits operation result, goes to step S3；
Result of calculation is passed to I/O modules, goes to step S7 by S6；
S7, I/O module export result.
5. according to the method described in claim 4, it is characterized in that, in step S2, N number of big section is isometric or differs
It is long.
6. according to the method described in claim 4, it is characterized in that, the function interpolation table includes the ginseng of the interpolating function
Number.
7. according to the method described in claim 4, it is characterized in that, the interpolating function is linear interpolation function or multinomial
Interpolating function.
8. a kind of neural network device, calculated by the use of piecewise interpolation using neuron input value and the inner product of weighted value as independent variable
Activation primitive value, including：
Memory, for storing executable instruction；
Processor, for performing the executable instruction stored in memory, to perform following operating procedure：
One, input data goes to step two as independent variable；
Two, the value range of independent variable is divided into N number of big section in advance：A1, A2..., AN, each big section is divided into M again
Minizone, N and M are natural number, and independent variable is fallen in big section AiIn, i is obtained, initializes cycle indexed variable p, a p=0,
Go to step three；
Three, according to the N section interpolation tables for being stored in memory, loading pth section interpolation table is searched, and letter is found out according to independent variable
Corresponding parameter value in number interpolation table, goes to step four；
Four, corresponding interpolating function value is calculated according to parameter value and independent variable, indexed variable p=p+1 is recycled, judges p's
Value --- if p ＜ i go to step three；Otherwise, five are gone to step；
Five, export interpolating function value.
9. device according to claim 8, which is characterized in that the processor includes CPU either GPU or dedicated
Neural network processor.
10. device according to claim 8, which is characterized in that the activation primitive for hyperbolic tangent function or
Sigmoid functions.
Priority Applications (2)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201611182655.0A CN108205518A (en) | 2016-12-19 | 2016-12-19 | Obtain device, method and the neural network device of functional value |
| US16/446,564 US20190311264A1 (en) | 2016-12-19 | 2019-06-19 | Device and method for obtaining functional value, and neural network device |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201611182655.0A CN108205518A (en) | 2016-12-19 | 2016-12-19 | Obtain device, method and the neural network device of functional value |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108205518A true CN108205518A (en) | 2018-06-26 |
Family
ID=62601996
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201611182655.0A Pending CN108205518A (en) | 2016-12-19 | 2016-12-19 | Obtain device, method and the neural network device of functional value |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108205518A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| 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 |
| CN116185126A (en) * | 2022-09-08 | 2023-05-30 | 重庆位图信息技术有限公司 | Complex function output method and device based on lookup table |
Citations (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2001126058A (en) * | 1999-10-27 | 2001-05-11 | Nanao Corp | Interpolation arithmetic unit |
| CN1305612A (en) * | 1998-06-17 | 2001-07-25 | 新泻精密株式会社 | Data interpolation method |
| US20040103132A1 (en) * | 2002-11-25 | 2004-05-27 | Zakrzewski Radoslaw Romuald | Method and apparatus for fast interpolation of multi-dimensional functions with non-rectangular data sets |
| CN101030295A (en) * | 2006-02-27 | 2007-09-05 | 精工爱普生株式会社 | Data conversion method for processor, texture producing method, program, recording medium, and projector |
| CN101266463A (en) * | 2008-04-29 | 2008-09-17 | 江南大学 | Fuzzy controller based on FPGA |
| CN102096844A (en) * | 2011-03-07 | 2011-06-15 | 哈尔滨工业大学 | FPGA-based RBF function calculation module |
| US20140313946A1 (en) * | 2013-04-17 | 2014-10-23 | Lsi Corporation | Non-Linear Interference Cancellation For Wireless Transceivers |
| CN104484703A (en) * | 2014-12-30 | 2015-04-01 | 合肥工业大学 | Sigmoid function fitting hardware circuit based on Remez approximating algorithm |
| CN105354006A (en) * | 2015-11-27 | 2016-02-24 | 中国科学院计算技术研究所 | Quick operation device and method of nonlinear function |
| CN105468335A (en) * | 2015-11-24 | 2016-04-06 | 中国科学院计算技术研究所 | Pipeline-level operation device, data processing method and network-on-chip chip |
| CN106227291A (en) * | 2016-07-26 | 2016-12-14 | 中国科学院自动化研究所 | The implementation method of arctan function based on stagewise look-up table and realize device |
-
2016
- 2016-12-19 CN CN201611182655.0A patent/CN108205518A/en active Pending
Patent Citations (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN1305612A (en) * | 1998-06-17 | 2001-07-25 | 新泻精密株式会社 | Data interpolation method |
| JP2001126058A (en) * | 1999-10-27 | 2001-05-11 | Nanao Corp | Interpolation arithmetic unit |
| US20040103132A1 (en) * | 2002-11-25 | 2004-05-27 | Zakrzewski Radoslaw Romuald | Method and apparatus for fast interpolation of multi-dimensional functions with non-rectangular data sets |
| CN101030295A (en) * | 2006-02-27 | 2007-09-05 | 精工爱普生株式会社 | Data conversion method for processor, texture producing method, program, recording medium, and projector |
| CN101266463A (en) * | 2008-04-29 | 2008-09-17 | 江南大学 | Fuzzy controller based on FPGA |
| CN102096844A (en) * | 2011-03-07 | 2011-06-15 | 哈尔滨工业大学 | FPGA-based RBF function calculation module |
| US20140313946A1 (en) * | 2013-04-17 | 2014-10-23 | Lsi Corporation | Non-Linear Interference Cancellation For Wireless Transceivers |
| CN104484703A (en) * | 2014-12-30 | 2015-04-01 | 合肥工业大学 | Sigmoid function fitting hardware circuit based on Remez approximating algorithm |
| CN105468335A (en) * | 2015-11-24 | 2016-04-06 | 中国科学院计算技术研究所 | Pipeline-level operation device, data processing method and network-on-chip chip |
| CN105354006A (en) * | 2015-11-27 | 2016-02-24 | 中国科学院计算技术研究所 | Quick operation device and method of nonlinear function |
| CN106227291A (en) * | 2016-07-26 | 2016-12-14 | 中国科学院自动化研究所 | The implementation method of arctan function based on stagewise look-up table and realize device |
Non-Patent Citations (3)
| Title |
|---|
| AMIN H ET AL.: ""Piecewise linear approximation applied to nonlinear function of a neural network"", 《IEE PROCEEDINGS-CIRCUITS DEVICES AND SYSTEMS》 * |
| CHARLES VAN LOAN: ""Piecewise Polynomial Interpolation"", 《LECTURE NOTES》 * |
| KOLDO BASTERREXTEA ET AL.: ""Approximation of Sigmoid Function and the Derivative for Artificial Neurons"", 《IEE PROCEEDINGS-CIRCUITS DEVICES AND SYSTEMS》 * |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| 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 |
| CN116185126A (en) * | 2022-09-08 | 2023-05-30 | 重庆位图信息技术有限公司 | Complex function output method and device based on lookup table |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN105354006B (en) | A kind of rapid computations devices and methods therefor of nonlinear function | |
| Khan et al. | NPE: An FPGA-based overlay processor for natural language processing | |
| Gembris et al. | Correlation analysis on GPU systems using NVIDIA’s CUDA | |
| CN109858623A (en) | Device and method for executing artificial neural network forward operation | |
| Agarwal et al. | High-throughput implementation of a million-point sparse Fourier transform | |
| CN110163356A (en) | A kind of computing device and method | |
| US20210256357A1 (en) | Embedded stochastic-computing accelerator architecture and method for convolutional neural networks | |
| CN107305484A (en) | A kind of nonlinear function arithmetic unit and method | |
| CN107967132A (en) | A kind of adder and multiplier for neural network processor | |
| Wu et al. | Efficient dynamic fixed-point quantization of CNN inference accelerators for edge devices | |
| CN108205518A (en) | Obtain device, method and the neural network device of functional value | |
| WO2018112692A1 (en) | Device and method for obtaining functional value, and neural network device | |
| Noh et al. | FlexBlock: A flexible DNN training accelerator with multi-mode block floating point support | |
| Lu et al. | Redesk: A reconfigurable dataflow engine for sparse kernels on heterogeneous platforms | |
| Song et al. | Accelerating kNN search in high dimensional datasets on FPGA by reducing external memory access | |
| Isupov et al. | Fast power-of-two RNS scaling algorithm for large dynamic ranges | |
| US20210056399A1 (en) | Asynchronous accumulator using logarithmic-based arithmetic | |
| CN111178492B (en) | Computing device, related product and computing method for executing artificial neural network model | |
| Struharik et al. | Hardware implementation of decision tree ensembles | |
| Prasad et al. | Review on Different Types of Multipliers and Its Performance Comparisons | |
| Ly et al. | A multi-FPGA architecture for stochastic restricted Boltzmann machines | |
| CN110059814A (en) | A kind of lookup tabular form convolution algorithm hardware configuration based on FPGA | |
| Mayannavar et al. | Performance comparison of serial and parallel multipliers in massively parallel environment | |
| Xu et al. | Design and implementation of an efficient CNN accelerator for low-cost FPGAs | |
| Alawad et al. | Stochastic-based convolutional networks with reconfigurable logic fabric |
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 |