WO1993000652A1 - Verfahren zur verarbeitung von unsicherheiten von eingangsdaten in neuronalen netzwerken - Google Patents
Verfahren zur verarbeitung von unsicherheiten von eingangsdaten in neuronalen netzwerken Download PDFInfo
- Publication number
- WO1993000652A1 WO1993000652A1 PCT/DE1992/000494 DE9200494W WO9300652A1 WO 1993000652 A1 WO1993000652 A1 WO 1993000652A1 DE 9200494 W DE9200494 W DE 9200494W WO 9300652 A1 WO9300652 A1 WO 9300652A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- input signals
- vector
- neutral
- neuron
- training data
- Prior art date
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/043—Architecture, e.g. interconnection topology based on fuzzy logic, fuzzy membership or fuzzy inference, e.g. adaptive neuro-fuzzy inference systems [ANFIS]
Definitions
- the invention is based on the object of specifying a method with which known uncertainties of input data of neural networks can be taken into account in order to improve the behavior of a neural network in the test phase. This object is achieved by a method for processing uncertainties of input data in neural networks with features according to claim 1.
- 1 shows the space of the possible input signals of a neuron with two inputs, in which two classes of training data, which are separated by a one-dimensional hyperplane (straight line), are shown.
- FIG. 2 shows the same situation as FIG. 1 with test data, the variance of which exceeds that of the training data.
- FIG. 3 shows the same situation as FIG. 2 with a second hyperplane (straight line) which appropriately separates the two classes of test data.
- Neural networks are generally constructed from a multiplicity of neurons which calculate a weighted sum from a multiplicity of inputs using weight factors and subject this weighted sum to a threshold value decision using a preferably sigmoid transfer function.
- connection weights weight factors
- the connection weights (weight factors) between the inputs and the neuron are set or learned in a training phase by adapting the weight factors to a set of training data.
- a possible training method (Rumelhart, 1986) realizes a gradient descent over an error surface. The best possible separability of the two classes is achieved by using a sigmoidal transfer function.
- This neuron can differentiate between two classes, the initial activity of the neuron being "0" for one class and "1" for the other. It is further assumed that in this example the classes can be separated by a linear plane and that only one input line contributes the necessary information for this distinction. The second input line should only carry redundant information here.
- Fig. 1 shows two classes K1 and K2 of training data, which can be separated by a linear hyperplane H1 (straight line).
- FIG. 2 shows the same situation as in FIG. 1, but using a larger amount of data - the test data, which also include the training data.
- the two classes K11 and K21 of test data can no longer be separated by a horizontally running hyperplane, but only by a vertically running hyperplane.
- the diagonally running hyperplane H1 visible in FIGS. 1 and 2 is capable of correctly dividing the training data into two classes, but not the test data.
- test data have a greater variance - mainly in the input signal I2 - than the training data. It is therefore input I1 that carries the information for the appropriate separation of test data classes K11 and K21. If, as in many applications, it is known from the outset that one of the inputs or, in general, several inputs is subject to an uncertainty and therefore does not allow reliable separation of the test data, it is desirable to use a method which, when using the training data from FIG 1 leads to the hyperplane H2, which is shown in FIG. The hyperplane H2 separates the training data and the test data into two classes K11 and K21. Such a method must make use of the knowledge about the uncertainty of the information of the input 12. The situation shown in FIGS.
- 1, 2 and 3 is a great simplification of the in many applications, such as. B. in the image and speech processing situation in which high-dimensional input vectors occur, the individual components of which have considerable redundancy.
- the general solution according to the invention to the described problem consists in not multiplying and adding up the input values of one or more neurons of the neural network directly by the weighting factors, but instead first calculating modified input signals from the input signals of the neurons, which appropriately take into account the different uncertainties of different input signals .
- a neuron calculates an output value a according to the formula from its input values i k
- n k the kth value of a neutral vector N.
- each neuron in addition to its output value (neuron activity), has an average security of its input signals according to the following relationship
- neural network can thus be used in the most general meaning of this word, with particular reference to neural networks which are implemented by appropriate computer programs on suitable computer systems.
- the method can also be used in the context of neural networks implemented in terms of circuitry.
- the neutral vector N can be calculated as the intersection of the hyperplane set by the training method with those hyperplanes which result when an input signal of a neuron is not taken into account in each case.
- K different hyperplanes for One neuron each with K inputs a hyperplane that divides the K-dimensional space, and K - 1 other hyperplanes that subdivide the K - 1-dimensional subspaces.
- Such a multiple training process can be carried out simultaneously, whereby approximately K times the amount of computing time and storage space is necessary.
- K linear equations can be solved to determine the intersection of all hyperplanes.
- Step 1 ensures that all modifications of N occur in the parting plane H.
- Step 2 iteratively calculates the focus of all projections.
- N new N old + ⁇ ⁇ (I '- N old ), where N old is the previous value of the neutral vector, N new is the corrected value of the neutral vector, and ⁇ is a factor between 0 and 1, which is the speed of the Correction of the neutral vector is determined, ⁇ can be chosen to be constant if there are approximately the same number of examples for each class. If this is not the case, ⁇ must be selected in proportion to the number of examples in this class in accordance with the class belonging to the current training pattern.
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Software Systems (AREA)
- Artificial Intelligence (AREA)
- Data Mining & Analysis (AREA)
- Mathematical Optimization (AREA)
- Pure & Applied Mathematics (AREA)
- Fuzzy Systems (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Computational Mathematics (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Mathematical Analysis (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Automation & Control Theory (AREA)
- Image Analysis (AREA)
- Complex Calculations (AREA)
Abstract
Description
Claims
Priority Applications (3)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
JP4510691A JPH06508948A (ja) | 1991-06-24 | 1992-06-16 | ニューラルネットワークにおける入力データの不確定性状態の処理方法 |
EP92911631A EP0591259B1 (de) | 1991-06-24 | 1992-06-16 | Verfahren zur verarbeitung von unsicherheiten von eingangsdaten in neuronalen netzwerken |
DE59203485T DE59203485D1 (de) | 1991-06-24 | 1992-06-16 | Verfahren zur verarbeitung von unsicherheiten von eingangsdaten in neuronalen netzwerken. |
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
DEP4120828.5 | 1991-06-24 | ||
DE4120828A DE4120828A1 (de) | 1991-06-24 | 1991-06-24 | Verfahren zur verarbeitung von unsicherheiten von eingangsdaten in neuronalen netzwerken |
Publications (1)
Publication Number | Publication Date |
---|---|
WO1993000652A1 true WO1993000652A1 (de) | 1993-01-07 |
Family
ID=6434630
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/DE1992/000494 WO1993000652A1 (de) | 1991-06-24 | 1992-06-16 | Verfahren zur verarbeitung von unsicherheiten von eingangsdaten in neuronalen netzwerken |
Country Status (4)
Country | Link |
---|---|
EP (1) | EP0591259B1 (de) |
JP (1) | JPH06508948A (de) |
DE (2) | DE4120828A1 (de) |
WO (1) | WO1993000652A1 (de) |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
EP0892956A1 (de) * | 1996-02-09 | 1999-01-27 | Sarnoff Corporation | Verfahren und apparat zum trainieren eines neuralen netzwerks um objekte mit unsicheren trainingsdaten zu erkennen und zu klassifizieren |
WO2004090807A2 (de) * | 2003-04-10 | 2004-10-21 | Bayer Technology Services Gmbh | Verfahren zum trainieren von neuronalen netzen |
Families Citing this family (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
DE4340977C1 (de) * | 1993-12-01 | 1995-03-16 | Deutsche Aerospace | Adaptives Signalauswerte-System |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US4876731A (en) * | 1988-02-19 | 1989-10-24 | Nynex Corporation | Neural network model in pattern recognition using probabilistic contextual information |
EP0378689A1 (de) * | 1988-05-20 | 1990-07-25 | Matsushita Electric Industrial Co., Ltd. | Verfahren zur bestimmung der inferenzregel sowie inferenzmotor |
Family Cites Families (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US5077677A (en) * | 1989-06-12 | 1991-12-31 | Westinghouse Electric Corp. | Probabilistic inference gate |
DE4018779A1 (de) * | 1989-10-03 | 1991-04-11 | Westinghouse Electric Corp | Neurale netzwerke und lernfaehiges schlussfolgerungssystem |
-
1991
- 1991-06-24 DE DE4120828A patent/DE4120828A1/de not_active Withdrawn
-
1992
- 1992-06-16 JP JP4510691A patent/JPH06508948A/ja active Pending
- 1992-06-16 DE DE59203485T patent/DE59203485D1/de not_active Expired - Fee Related
- 1992-06-16 EP EP92911631A patent/EP0591259B1/de not_active Expired - Lifetime
- 1992-06-16 WO PCT/DE1992/000494 patent/WO1993000652A1/de active IP Right Grant
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US4876731A (en) * | 1988-02-19 | 1989-10-24 | Nynex Corporation | Neural network model in pattern recognition using probabilistic contextual information |
EP0378689A1 (de) * | 1988-05-20 | 1990-07-25 | Matsushita Electric Industrial Co., Ltd. | Verfahren zur bestimmung der inferenzregel sowie inferenzmotor |
Non-Patent Citations (3)
Title |
---|
IEEE FIRST INTERNATIONAL CONFERENCE ON NEURAL NETWORKS Bd. 3, 21. Juni 1987, SAN DIEGO,USA Seiten 51 - 58 GAINES 'Uncertainty as a foundation of computational power in neural networks' * |
INTERNATIONAL NEURAL NETWORK CONFERENCE INNC 90 PARIS Bd. 2, 9. Juli 1990, PARIS, FRANCE Seiten 809 - 812 MARSHALL 'Representation of uncertainty in self-organizing neural networks' * |
INTERNATIONAL NEURAL NETWORK CONFERENCE INNC 90 PARIS Bd. 2, 9. Juli 1990, PARIS,FRANCE Seiten 902 - 907 SUN 'The discrete neuronal model and the probabilistic discrete neuronal model' * |
Cited By (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
EP0892956A1 (de) * | 1996-02-09 | 1999-01-27 | Sarnoff Corporation | Verfahren und apparat zum trainieren eines neuralen netzwerks um objekte mit unsicheren trainingsdaten zu erkennen und zu klassifizieren |
EP0892956A4 (de) * | 1996-02-09 | 2002-07-24 | Sarnoff Corp | Verfahren und apparat zum trainieren eines neuralen netzwerks um objekte mit unsicheren trainingsdaten zu erkennen und zu klassifizieren |
WO2004090807A2 (de) * | 2003-04-10 | 2004-10-21 | Bayer Technology Services Gmbh | Verfahren zum trainieren von neuronalen netzen |
WO2004090807A3 (de) * | 2003-04-10 | 2005-12-22 | Bayer Technology Services Gmbh | Verfahren zum trainieren von neuronalen netzen |
US7406451B2 (en) | 2003-04-10 | 2008-07-29 | Bayer Technology Services Gmbh | Method for training neural networks |
Also Published As
Publication number | Publication date |
---|---|
DE4120828A1 (de) | 1993-01-07 |
EP0591259A1 (de) | 1994-04-13 |
JPH06508948A (ja) | 1994-10-06 |
DE59203485D1 (de) | 1995-10-05 |
EP0591259B1 (de) | 1995-08-30 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
DE69333664T2 (de) | Verfahren und Gerät zur Einstellung eines Neurons | |
DE68928612T2 (de) | Verfahren zur bestimmung der inferenzregel sowie inferenzmotor | |
DE212020000731U1 (de) | Kontrastives Vortraining für Sprachaufgaben | |
DE69414752T2 (de) | Sprecherunabhängiges Erkennungssystem für isolierte Wörter unter Verwendung eines neuronalen Netzes | |
EP2165296B1 (de) | Verfahren und vorrichtung zur rechnergestützten vorhersage von intendierten bewegungen | |
DE2524804A1 (de) | Verfahren und vorrichtung zur automatischen spracherkennung | |
DE69719058T2 (de) | Verbessertes neuronales netzwerk | |
DE3938645C1 (de) | ||
DE69223447T2 (de) | Lernverfahren für neuronales Netzwerk und Klassifizieranlage zum Anwenden dieses Verfahrens | |
EP0901658B1 (de) | Verfahren zur optimierung eines fuzzy-regelsatzes durch einen rechner | |
DE69314293T2 (de) | Neuronalanlage und -Bauverfahren | |
WO2019081241A1 (de) | Verfahren, vorrichtung und computerprogramm zur erstellung eines tiefen neuronalen netzes | |
EP0925541B1 (de) | Verfahren und vorrichtung zur rechnergestützten generierung mindestens eines künstlichen trainingsdatenvektors für ein neuronales netz | |
DE69333247T2 (de) | Trainingsverfahren und Gerät zur Erzeugung eines neues Neurons | |
DE19504664C2 (de) | Chaotisches rekurrentes neuronales Netz und Lernverfahren dafür | |
EP0750764B1 (de) | Verfahren und anordnung zur fuzzy-regelung | |
DE10047718A1 (de) | Verfahren zur Spracherkennung | |
EP0591259B1 (de) | Verfahren zur verarbeitung von unsicherheiten von eingangsdaten in neuronalen netzwerken | |
DE4121453C2 (de) | Näherungsschlußfolgerungsvorrichtung | |
DE69802372T4 (de) | Klassifizierungssystem und -verfahren mit N-Tuple- oder RAM-basiertem neuronalem Netzwerk | |
DE69030812T2 (de) | Verfahren und gerät um einen prozess zu steuern mit hilfe eines trainierten parallel verteilten verarbeitungsnetzwerkes | |
DE19703964C1 (de) | Verfahren zur Transformation einer zur Nachbildung eines technischen Prozesses dienenden Fuzzy-Logik in ein neuronales Netz | |
DE69809402T2 (de) | Assoziativneuron in einem künstlichen neuralen netzwerk | |
DE4209746A1 (de) | Verfahren zur Optimierung eines technischen Neuro-Fuzzy-Systems | |
WO1998007100A1 (de) | Rechnergestütztes verfahren zur auswahl von trainingsdaten für ein neuronales netz |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
AK | Designated states |
Kind code of ref document: A1 Designated state(s): JP US |
|
AL | Designated countries for regional patents |
Kind code of ref document: A1 Designated state(s): AT BE CH DE DK ES FR GB GR IT LU MC NL SE |
|
DFPE | Request for preliminary examination filed prior to expiration of 19th month from priority date (pct application filed before 20040101) | ||
WWE | Wipo information: entry into national phase |
Ref document number: 1992911631 Country of ref document: EP |
|
ENP | Entry into the national phase |
Ref country code: US Ref document number: 1993 162192 Date of ref document: 19931215 Kind code of ref document: A Format of ref document f/p: F |
|
WWP | Wipo information: published in national office |
Ref document number: 1992911631 Country of ref document: EP |
|
WWG | Wipo information: grant in national office |
Ref document number: 1992911631 Country of ref document: EP |
|
ENP | Entry into the national phase |
Ref country code: US Ref document number: 1995 565047 Date of ref document: 19951130 Kind code of ref document: A Format of ref document f/p: F |