CN1331092C - Special purpose neural net computer system for pattern recognition and application method - Google Patents
Special purpose neural net computer system for pattern recognition and application method Download PDFInfo
- Publication number
- CN1331092C CN1331092C CNB2004100379654A CN200410037965A CN1331092C CN 1331092 C CN1331092 C CN 1331092C CN B2004100379654 A CNB2004100379654 A CN B2004100379654A CN 200410037965 A CN200410037965 A CN 200410037965A CN 1331092 C CN1331092 C CN 1331092C
- Authority
- CN
- China
- Prior art keywords
- neural network
- arithmetic
- pattern
- logic
- recognition
- 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.)
- Expired - Lifetime
Links
- 238000003909 pattern recognition Methods 0.000 title claims abstract description 56
- 238000000034 method Methods 0.000 title abstract description 29
- 230000001537 neural effect Effects 0.000 title description 4
- 238000013528 artificial neural network Methods 0.000 claims abstract description 82
- 210000002569 neuron Anatomy 0.000 claims description 12
- 230000007613 environmental effect Effects 0.000 claims description 8
- 238000005303 weighing Methods 0.000 claims description 8
- 238000013178 mathematical model Methods 0.000 claims description 3
- 210000000225 synapse Anatomy 0.000 claims description 3
- 238000004364 calculation method Methods 0.000 abstract 1
- 238000013497 data interchange Methods 0.000 abstract 1
- 238000013500 data storage Methods 0.000 abstract 1
- 238000012549 training Methods 0.000 description 16
- 238000013473 artificial intelligence Methods 0.000 description 6
- 238000000605 extraction Methods 0.000 description 5
- 238000010586 diagram Methods 0.000 description 3
- 235000013580 sausages Nutrition 0.000 description 3
- 230000000875 corresponding effect Effects 0.000 description 2
- 238000013461 design Methods 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 230000011218 segmentation Effects 0.000 description 2
- OYPRJOBELJOOCE-UHFFFAOYSA-N Calcium Chemical compound [Ca] OYPRJOBELJOOCE-UHFFFAOYSA-N 0.000 description 1
- 241000282326 Felis catus Species 0.000 description 1
- 244000287680 Garcinia dulcis Species 0.000 description 1
- 241000282320 Panthera leo Species 0.000 description 1
- 241000282376 Panthera tigris Species 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000018109 developmental process Effects 0.000 description 1
- 230000009977 dual effect Effects 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 239000011664 nicotinic acid Substances 0.000 description 1
- 238000012567 pattern recognition method Methods 0.000 description 1
- 238000005070 sampling Methods 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
Images
Landscapes
- Image Analysis (AREA)
Abstract
The present invention relates to a special pattern recognition neural network computer system and an application method. The special pattern recognition neural network computer system comprises a bus, a memory component, an arithmetic / logical operation and control component, neural network hardware and an environment interface component, wherein the memory component provides data storage space in the system; the arithmetic / logical operation and control component undertakes the task of arithmetic / logical operation and is used for controlling the operation and the data interchange of other components in the system by the bus; the neural network hardware receives data from the memory component to carry out neural network calculation and store a result to the memory component by the bus according to the instruction of the arithmetic / logical operation and control component; the environment interface component acquires information from environment according to the instruction of the arithmetic / logical operation and control component, or expresses the operation result of the system by voice or other modes.
Description
Technical field
The invention belongs to computer realm, the special-purpose neural network computer of a kind of pattern-recognition system more specifically says so.
Background technology
Be the history that the development of the pattern-recognition of representative has had decades with Fisher, Vapnik, obtained significant achievement.But these traditional mode identification methods are only paid attention to " division " between the various different sample types, and think little of the feature of similar sample itself.This mode identification method based on " division " thought can not overcome bigger not training sample misclassification rate problem, and when increasing a kind of new type sample, all needs all types of samples are trained again.
In recent years, the king keeps and feels that the academician has proposed a kind of new bionical pattern recognition theory, this theory has fundamentally broken through the thought of traditional mode identification " division ", utilization topology and higher dimensional space geometry, with the new network mode recognition system of multiple weighing value cynapse neuron structure, overcome traditional mode recognition methods not misclassification rate height, the overall problem of training again of newly-increased sample type needs of training sample, significantly improved the performance of pattern recognition system.
Based on bionical pattern recognition theory, the present invention has realized the special-purpose neural network computer of a kind of pattern-recognition system, and has proposed a kind ofly to utilize the special-purpose neural network computer of this pattern-recognition system to learn and discern the pattern-recognition demonstration application method of separating.This demo system ratio of performance to price height, the demonstration recognition effect is good, and simple to operate, have stronger interest, can be applied to large, medium and small student's artificial intelligence teaching demonstration teaching aid, to improve the interest of student's Scientific exploration knowledge.
Summary of the invention
The objective of the invention is to, the special-purpose neural network computer of a kind of pattern-recognition system is provided, put into practice bionical pattern recognition theory, improve the sample discrimination of pattern recognition system, it is higher and add the problem that new samples need be trained again to all samples to overcome in the traditional mode identification training sample misclassification rate not.Because the pattern-recognition application process that adopts neural network learning to separate with identification, the special-purpose neural network computer of pattern-recognition of the present invention system has higher performance and ease for operation, can be applicable to artificial intelligence teaching demonstration teaching aid.
The special-purpose neural network computer of a kind of pattern-recognition of the present invention system is characterized in that, comprising:
One bus
One memory member provides data space in system;
One arithmetic/logic and control assembly, this arithmetic/logic and control assembly are born the arithmetic/logic task in system, and pass through the operation and the exchanges data of other each parts of bus control system;
One neural network hardware, this neural network hardware carry out neural network by bus from memory member reception data according to the instruction of arithmetic/logic and control assembly and calculate, and the result is saved in memory member;
One environmental interface parts, these environmental interface parts obtain information according to the instruction of arithmetic/logic and control assembly from environment, or The results of running is expressed by voice or alternate manner.
Neural network hardware is wherein born the distributed parallel neural network and is calculated in system, this neural network hardware is the essential structure unit with the general neuron of multiple weighing value cynapse, and wherein the general neuronic mathematical model basic calculating formula of multiple weighing value cynapse is as follows:
Y=f[Φ (W
1, W
2..., W
m, X)], wherein Y is neuronic output, W
i(i=1...m) be the synapse weights, X is the neuron input.
Wherein arithmetic/logic and control assembly are finished arithmetic/logic in system, and the operation by total line traffic control neural network hardware, environmental interface parts, and the data access operation of memory member.
Description of drawings
For further specifying technology contents of the present invention, the invention will be further described below in conjunction with drawings and Examples, wherein:
The system construction drawing of the special-purpose neural network computer of the pattern-recognition among Fig. 1 the present invention system;
The neural network learning process flow diagram that the special-purpose neural network computer system applies of pattern-recognition among Fig. 2 the present invention microcomputer is assisted;
The special-purpose neural network computer of pattern-recognition among Fig. 3 the present invention system breaks away from the network mode identification process figure that microcomputer is independently finished.
Embodiment
The invention provides the special-purpose neural network computer of pattern-recognition system among the present invention of a kind of disengaging microcomputer.The embodiment of the special-purpose neural network computer of the pattern-recognition among the present invention system sees also Fig. 1, is described below:
The special-purpose neural network computer of a kind of pattern-recognition system is characterized in that, comprising:
One bus 50;
One memory member 10, this memory member 10 comprises nonvolatile memory 11 and volatile memory RAM 12, is perhaps all realized by nonvolatile memory, and data space is provided in system;
One arithmetic/logic and control assembly 30, this arithmetic/logic and control assembly are born the arithmetic/logic task in system, and pass through the operation and the exchanges data of bus 50 other each parts of control system;
One neural network hardware 20, this neural network hardware 20 carry out neural network by bus 50 from memory member 10 reception data according to the instruction of arithmetic/logic and control assembly 30 and calculate, and the result is saved in memory member 10;
One interface unit 40, these environmental interface parts 40 obtain information according to the instruction of arithmetic/logic and control assembly 30 from environment, or The results of running is expressed by voice or alternate manner.
Wherein neural network hardware 20 is born the neural calculating of distributed parallel in system, this neural network hardware 20 adopts the multiplexing mode of hardware neuron to realize, hardware neuron wherein is the general neuron of a kind of multiple weighing value cynapse, and the general neuronic mathematical model basic calculating formula of multiple weighing value cynapse is as follows:
Y=f[Φ (W
1, W
2..., W
m, X)], wherein Y is neuronic output, W
i(i=1...m) be the synapse weights, X is the neuron input.
Wherein arithmetic/logic and control assembly 30 adopt MCU 31 (MicrocontrolerUnit single-chip microcomputer) and utilize association's controller 32 of programmable logic device (PLD) exploitation to realize, also can adopt a powerful general processor to realize.This arithmetic/logic and control assembly 30 are finished arithmetic/logic in system, and pass through the operation of bus 50 control neural network hardwares 20, environmental interface parts 40, and the data access operation of memory member 10.
The application process of the special-purpose neural network computer of a kind of pattern-recognition of the present invention system, it is characterized in that, be a kind of based on bionical pattern recognition theory, realize the artificial intelligence demonstration application method that neural network learning separates with identification with the special-purpose neural network computer of pattern-recognition of the present invention system.Comprise the steps:
1) the neural network learning process of microcomputer assistance; The neural network learning that this microcomputer is assisted is meant and utilizes microcomputer to assist the special-purpose neuro-computer of pattern-recognition of the present invention that training sample set is learnt, and makes it have " identification " ability to the things of having learnt.Wherein the microcomputer neural network learning process of assisting is meant by microcomputer and gathers sample and feature extraction, and according to the distance between the sample point in the feature space, and sample is selected and sorted; The special-purpose neural network computer of pattern-recognition of the present invention then system is learnt the sample of selecting ordering, constructing neural network, and neural network structure is saved in the nonvolatile memory.
2) break away from microcomputer, the neural network identification presentation process that the special-purpose neural network computer of pattern-recognition system independently finishes.Wherein break away from the neural network identification demonstration that the special-purpose neural network computer of microcomputer pattern-recognition system independently finishes, be meant that the special-purpose neural network computer of pattern-recognition of the present invention breaks away from microcomputer, independently finish network mode identification demonstration work.
Wherein break away from the neural network identification presentation process that the special-purpose neural network computer of microcomputer pattern-recognition system independently finishes, be meant the pattern-recognition demo system of the special-purpose neural network computer of pattern-recognition of the present invention system as an independent operation, independently finish following work: gather pattern sample to be identified and this sample is carried out feature extraction, calculate and judgement by neural network, the output mode recognition result or (with) carry out corresponding actions.
The special-purpose neural network computer system application method of pattern-recognition among the present invention is a kind of bionical pattern recognition theory based on non-division, realize with the special-purpose neural network computer of pattern-recognition of the present invention system, the artificial intelligence demonstration application method that neural network learning separates with identification, its concrete implementation step sees also Fig. 2, Fig. 3.Be described below:
1, the neural network learning process of microcomputer assistance, its process flow diagram as shown in Figure 2.Gather sample and feature extraction by microcomputer, and according to the distance between the sample point in the feature space, sample is selected and sorted.The special-purpose neural network computer of pattern-recognition of the present invention then system constructs the sample distribution subspace of such sample according to the complicated geometirc physique that similar sample point distributed in the feature space with the method for higher dimensional space complicated geometirc physique piecewise approximation covering.For example the actual body of sample is distributed as a hypercurve, then can adopt the straight line of segmentation to be similar to the sample distribution subspace that such sample is constructed in covering.Repeat said process until the structure of finishing all types sample distribution subspace, thereby finish the study of neural network and network architecture parameters is stored in the nonvolatile memory 11.
2, break away from the neural network identification presentation process that the special-purpose neural network computer of pattern-recognition system independently finishes among microcomputer the present invention, its process flow diagram as shown in Figure 3.The special-purpose neural network computer of pattern-recognition among the present invention system is as the pattern-recognition demo system of an independent operation, independently finish following work: gather pattern sample to be identified and this sample is carried out feature extraction, calculate and judgement by neural network, the output mode recognition result or (with) carry out corresponding actions.
Embodiment
The application example of the special-purpose neural network computer of pattern-recognition of the present invention system is that requirement is discerned demonstration to the mock-up of different angles on the surface level.The collection of sample be utilize microcomputer from different directions the Bmp file that collects of observation post screen and sort composing training sample set S={S
i' | (in the formula, i is the sequence number of training sample to i=0...j}, and j is the number of training sample, S
i' be i training sample).The sample of training sample being concentrated by the special-purpose neural network computer of pattern-recognition of the present invention system carries out feature extraction then, and each sample image obtains the proper vector of one 256 dimension.Because direction of observation all is a level, we can say that the change of direction has only a variable, thereby the distribution of sample point is one-dimensional manifold distribution Pa in the feature space.Therefore we adopt the approximate geometry body of such sample subspace of the approximate structure of mode of super sausage segmentation covering, and it is super sausage neuron that multiple weighing value cynapse hardware neuron is set, and finishes the training of neural network.Super sausage approximate geometry body covering method is described below:
Cover Pa approx with j neuron Pi (i=0...j-1), wherein Pi is:
P
i＝{x|ρ(x，y)≤K，y∈B
i，x∈R
n}
B
i＝{x|x＝αS
i′+(1-α)S
i+1′，α∈[0，1]}
Wherein (x y) is x to ρ, the distance between 2 of the y, B
iFor with sample point S
i', S
I+1' be the line segment of end points.
The sample subspace Pa that then obtains is:
In experiment, as training sample model, gather 6400 in sample with 8 models such as lion, tiger, tank altogether, wherein every class sample is selected 26 to 50 real training samples of conduct, and 8 classes have 338.In addition with 6 models such as cat, doggie, 400 test sample books that conduct is not trained of respectively sampling.Drawing that the correct recognition rata of training sample is 99.75%, is 0.25% according to the knowledge rate, and misclassification rate is 0; The training sample misclassification rate is not 0.From top result as can be seen, the artificial intelligence demo system that the present invention realizes can not known training objects not by mistake, has demonstrated fully the superiority of bionical pattern recognition theory with respect to the traditional mode recognition methods.
The system architecture of the special-purpose neural network computer of the pattern-recognition system in the special-purpose neural network computer of the pattern-recognition of the present invention system as shown in Figure 1, solve the speciality of thinking in images problem and the arithmetic/logic unit ability in the logical thinking field in conjunction with neural network hardware, this system adopts the institutional framework of dual processor shared storage.
The special-purpose neural network computer system application method of pattern-recognition in the special-purpose neural network computer of the pattern-recognition of the present invention system, be a kind of based on bionical pattern recognition theory, realize the artificial intelligence demonstration application method that neural network learning separates with identification with the special-purpose neural network computer of pattern-recognition of the present invention system.Because the technological means that has adopted study to separate with identification, greatly reduce the cost of system and the complicacy of system operation, the participation presentation process is simple to operate, the result is accurate, interesting strong, therefore, be suitable as very much large, medium and small student's teaching demonstration teaching aid, to improve the interest of student's Scientific exploration knowledge.
The present invention compared with prior art has the following advantages:
1, adopts the mode of shared storage spare parts 10, finish with arithmetic/logic and control But the two-processor system structure centered by parts 30 and the variable topological structure neural network hardware 20 Design, and the hardware design of realization special purpose neural net computer system for pattern recognition of the present invention.
2. based on the theoretical system applies of bionic pattern identification, obtain more than traditional mode recognition methods Superior performance.
3. propose the application of pattern recognition method that study separates with identification, make pattern-recognition of the present invention special Has the higher ratio of performance to price with neural computer.
Claims (3)
1, the special-purpose neural network computer of a kind of pattern-recognition system is characterized in that, comprising:
One bus;
One memory member provides data space in system;
One arithmetic/logic and control assembly, this arithmetic/logic and control assembly are born the arithmetic/logic task in system, and pass through the operation and the exchanges data of other each parts of bus control system;
One neural network hardware, this neural network hardware carry out neural network by bus from memory member reception data according to the instruction of arithmetic/logic and control assembly and calculate, and the result is saved in memory member;
One environmental interface parts, these environmental interface parts obtain information according to the instruction of arithmetic/logic and control assembly from environment, or The results of running is expressed by voice mode.
2. the special-purpose neural network computer of pattern-recognition according to claim 1 system, it is characterized in that: neural network hardware wherein, bearing the distributed parallel neural network in system calculates, this neural network hardware is the essential structure unit with the general neuron of multiple weighing value cynapse, and wherein the general neuronic mathematical model basic calculating formula of multiple weighing value cynapse is as follows:
Y=f[Φ (W
1, W
2..., W
m, X)], wherein Y is neuronic output, W
i(i=1...m) be the synapse weights, X is the neuron input.
3. the special-purpose neural network computer of pattern-recognition according to claim 1 system, it is characterized in that: wherein arithmetic/logic and control assembly, in system, finish arithmetic/logic, and the operation by total line traffic control neural network hardware, environmental interface parts, and the data access operation of memory member.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CNB2004100379654A CN1331092C (en) | 2004-05-17 | 2004-05-17 | Special purpose neural net computer system for pattern recognition and application method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CNB2004100379654A CN1331092C (en) | 2004-05-17 | 2004-05-17 | Special purpose neural net computer system for pattern recognition and application method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN1700250A CN1700250A (en) | 2005-11-23 |
| CN1331092C true CN1331092C (en) | 2007-08-08 |
Family
ID=35476308
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CNB2004100379654A Expired - Lifetime CN1331092C (en) | 2004-05-17 | 2004-05-17 | Special purpose neural net computer system for pattern recognition and application method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN1331092C (en) |
Families Citing this family (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN1838159B (en) * | 2006-02-14 | 2010-08-11 | 北京未名博思生物智能科技开发有限公司 | Cognition logic machine and its information processing method |
| JP5171118B2 (en) * | 2007-06-13 | 2013-03-27 | キヤノン株式会社 | Arithmetic processing apparatus and control method thereof |
| CN101527010B (en) * | 2008-03-06 | 2011-12-07 | 上海理工大学 | Hardware realization method and system for artificial neural network algorithm |
| CN101383023B (en) * | 2008-10-22 | 2011-04-06 | 西安交通大学 | Neural network short-term electric load prediction based on sample dynamic organization and temperature compensation |
| CN108427990B (en) | 2016-01-20 | 2020-05-22 | 中科寒武纪科技股份有限公司 | Neural network computing system and method |
| CN107329936A (en) * | 2016-04-29 | 2017-11-07 | 北京中科寒武纪科技有限公司 | A kind of apparatus and method for performing neural network computing and matrix/vector computing |
| CN109086877B (en) * | 2016-04-29 | 2020-05-08 | 中科寒武纪科技股份有限公司 | Apparatus and method for performing convolutional neural network forward operation |
| TWI653584B (en) | 2017-09-15 | 2019-03-11 | 中原大學 | Method of judging neural network with non-volatile memory cells |
| CN111582464B (en) * | 2017-12-29 | 2023-09-29 | 中科寒武纪科技股份有限公司 | Neural network processing method, computer system and storage medium |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN1282069A (en) * | 1999-07-27 | 2001-01-31 | 中国科学院自动化研究所 | On-palm computer speech identification core software package |
| US6377941B1 (en) * | 1998-11-26 | 2002-04-23 | International Business Machines Corporation | Implementing automatic learning according to the K nearest neighbor mode in artificial neural networks |
| CN1445715A (en) * | 2002-03-15 | 2003-10-01 | 微软公司 | System and method for mode recognising |
| CN1464477A (en) * | 2002-06-18 | 2003-12-31 | 中国科学院半导体研究所 | A process for constructing multiple weighing value synapse nerve cell |
| CN1470022A (en) * | 2000-11-30 | 2004-01-21 | Neural cortex |
-
2004
- 2004-05-17 CN CNB2004100379654A patent/CN1331092C/en not_active Expired - Lifetime
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US6377941B1 (en) * | 1998-11-26 | 2002-04-23 | International Business Machines Corporation | Implementing automatic learning according to the K nearest neighbor mode in artificial neural networks |
| CN1282069A (en) * | 1999-07-27 | 2001-01-31 | 中国科学院自动化研究所 | On-palm computer speech identification core software package |
| CN1470022A (en) * | 2000-11-30 | 2004-01-21 | Neural cortex | |
| CN1445715A (en) * | 2002-03-15 | 2003-10-01 | 微软公司 | System and method for mode recognising |
| CN1464477A (en) * | 2002-06-18 | 2003-12-31 | 中国科学院半导体研究所 | A process for constructing multiple weighing value synapse nerve cell |
Also Published As
| Publication number | Publication date |
|---|---|
| CN1700250A (en) | 2005-11-23 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| JP7289918B2 (en) | Object recognition method and device | |
| WO2019228358A1 (en) | Deep neural network training method and apparatus | |
| Xiang et al. | ThunderNet: A turbo unified network for real-time semantic segmentation | |
| KR102224253B1 (en) | Teacher-student framework for light weighted ensemble classifier combined with deep network and random forest and the classification method based on thereof | |
| CN111507378A (en) | Method and apparatus for training image processing model | |
| Peemen et al. | Speed sign detection and recognition by convolutional neural networks | |
| CN109800628A (en) | A kind of network structure and detection method for reinforcing SSD Small object pedestrian detection performance | |
| CN108062569B (en) | Unmanned vehicle driving decision method based on infrared and radar | |
| CN110059741A (en) | Image-recognizing method based on semantic capsule converged network | |
| WO2021203865A9 (en) | Molecular binding site detection method and apparatus, electronic device and storage medium | |
| CN112232490A (en) | Deep simulation reinforcement learning driving strategy training method based on vision | |
| CN101944174A (en) | Identification method of characters of licence plate | |
| CN109325430A (en) | Real-time Activity recognition method and system | |
| CN1331092C (en) | Special purpose neural net computer system for pattern recognition and application method | |
| CN109948707A (en) | Model training method, device, terminal and storage medium | |
| CN114120432A (en) | Online learning attention tracking method based on sight estimation and application thereof | |
| CN109784159A (en) | The processing method of scene image, apparatus and system | |
| Yao et al. | Unsupervised transfer learning for spatiotemporal predictive networks | |
| CN109101108A (en) | Method and system based on three decision optimization intelligence cockpit human-computer interaction interfaces | |
| CN111352419B (en) | Path planning method and system for updating experience playback cache based on time sequence difference | |
| CN113238970B (en) | Training method, evaluation method, control method and device of automatic driving model | |
| CN115457395A (en) | Lightweight remote sensing target detection method based on channel attention and multi-scale feature fusion | |
| CN111339935A (en) | Optical remote sensing picture classification method based on interpretable CNN image classification model | |
| CN113011304A (en) | Human body posture estimation method and system based on attention multi-resolution network | |
| Oeljeklaus | An integrated approach for traffic scene understanding from monocular cameras |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| C14 | Grant of patent or utility model | ||
| GR01 | Patent grant |