EP1166227A1 - Procede d'initialisation d'un reseau de neurones - Google Patents
Procede d'initialisation d'un reseau de neuronesInfo
- Publication number
- EP1166227A1 EP1166227A1 EP00910976A EP00910976A EP1166227A1 EP 1166227 A1 EP1166227 A1 EP 1166227A1 EP 00910976 A EP00910976 A EP 00910976A EP 00910976 A EP00910976 A EP 00910976A EP 1166227 A1 EP1166227 A1 EP 1166227A1
- Authority
- EP
- European Patent Office
- Prior art keywords
- neural network
- module
- function
- linear
- parameters
- 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.)
- Withdrawn
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; 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
- G06N3/0499—Feedforward networks
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/082—Learning methods modifying the architecture, e.g. adding, deleting or silencing nodes or connections
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/09—Supervised learning
Definitions
- the invention relates to a method for initializing a neural network, more particularly called perceptron.
- a neural network makes it possible to perform many tasks, such as identification, control, classification or filtering.
- the object of the invention is therefore to construct a neural network which emulates a non-linear function known only by some of its operating points, then to initialize this neural network.
- the object of the invention is a method for initializing a neural network intended to emulate a non-linear function, characterized in that it is carried out by an initialization box receiving as input the data available on the function, i.e. a sample of inputs and a sample of outputs, and performing the following steps: automatic selection of the number of linear zones necessary, taking into account the information available on the non-linear function, to linearly approximate said function; the automatic calculation of the characteristics of the linear forms on each area of the function, ie their slope; - the automated design of the neural network which associates the corresponding linear form with each zone of the function; the calculation of the initial values of the parameters of the neural network, ie the weights and the biases of the layer of hidden neurons and of the layer of output neurons.
- FIG. 1 a functional diagram of a neural network
- Figures 2 a and 2 ⁇ two examples of activation functions of a neural network
- Figure 3 a symbolic representation of the transfer function of a neural network
- FIG. 4 a functional diagram of a neural network initialized according to the method of the invention
- - Figure 5 the initialization box of a neural network according to the invention
- FIG. 6 an example of a non-linear function to be emulated and of a corresponding linear initialization by pieces
- FIG. 7 the characteristics of a space band associated with a linear zone of a non-linear function
- FIGS. 8 a and 8 b the learning mode and the use mode of a neural network.
- a neural network 1 can be represented by a functional diagram like that of FIG. 1, which includes a module 2 constituting the transfer function of the neural network. For a given input E and a set of parameters ⁇ fixed in a module 3, it produces a given output S.
- x (Xi) i the input vector with d inputs
- W [W] i j : dxc real parameters
- b 1 [b 1 ] j : c real parameters
- w 2 [W21 J jk : cxa real parameters
- b [ b 1 k : a real parameters
- with c number of hidden neurons in the network.
- Sub-module 4 of module 2 of the transfer function produces the output vector z whose expression Ex) from the activation function ⁇ of the network is as follows:
- This activation function ⁇ can take several forms, two examples of which are shown in FIGS. 2 a and 2 ⁇ .
- the functional diagram of the neural network also includes a module 5 intended for learning the network and which receives as input:
- the learning module 5 delivers a correction C to be made to the parameters ⁇ of the network.
- Sub-module 6 contains the transfer function which corresponds to the following equations (E 2 ):
- each function R is an update method
- ⁇ are the network parameters defined previously, y and y * the state of the network
- ⁇ the error committed by the network
- T the parameters of the learning method.
- the error ⁇ committed by the network is treated according to the parameters T.
- the calculation of the partial derivatives of the error as a function of the weights W 2 and of the biases b is carried out, and a correction ⁇ W and ⁇ b is applied as a function of the derivatives and of the parameters T. Then, the same operation is carried out for the weights W and the biases b 1 .
- the invention consists in determining the size of the hidden layer of the neural network and the initial values of the parameters of the network. These two objectives are achieved by an initialization box I (FIG. 4) made up of several modules Ii to I 5 , the inputs E 'of which consist of a sample of inputs E, a sample of outputs S and a list of input ranges Z according to a variant.
- This box I is intended to deliver the size c of the hidden layer of neurons and the initial parameters ⁇ Q of the neural network, usually chosen at random by the user. As shown in Figure 5, it includes:
- a module I for implementing statistical methods making it possible to estimate the non-linearity of a database, for example by linear regressions by pieces, receiving as input the samples of inputs E and of outputs S; a module I for determining the number of hidden units by exploiting new theoretical results making it possible to relate the number c of hidden units to its non-linearity; - a module I 3 for calculating the minimum number of hidden units necessary for the emulation, by the neural network, of the Z zones supplied as an option;
- the aim of the invention being to construct a neural network which emulates a non-linear function, known by some of its operating points, and of which an example f is represented in FIG. 6, we proceed by linear approximations by pieces of so that the initial shape of the neural network is characterized by a set of slopes and corresponding zones as shown by the curve f 2 in FIG. 6.
- the method according to the invention is based on the following three points: the automatic choice of the number of zones required taking into account the information available on the non-linear function; the automatic calculation of the characteristics of the linear forms on each area of the function, ie their slope; the automated design of the neural network which associates the corresponding linear form with each area of the function.
- the automatic choice of the number of zones Z ⁇ necessary to linearly approximate the nonlinear function to be emulated is carried out by the two modules I ⁇ and I 2 , the first I lr receiving as input the data available on the function to be emulated, ie a sample of entries
- This module performs statistical tests to choose the necessary zones, that is to say, it performs; a random selection of zones Z ⁇ ; a statistical regression operation on each of. these zones selected to obtain a set of parameters Pi per zone Z; - Significance tests of the differences between said sets of parameters Pi to carry out, in the event of a non-significant difference between at least two zones, the merging of said corresponding zones.
- each strip Ba is characterized by a standard vector ni which is orthogonal to it, a width l and a fulcrum Xi which makes it possible to place the line of direction orthogonal to the vector nor in the affine plane.
- a first strip represents the speed N of a vehicle engine between 600 and 1200 revolutions / minute as a function of another band representing a fast CMI + couple between 20 and 140.
- the module I 2 is intended to link the non-linearity of the function to be emulated to the number c of hidden units of the neural network to be constructed.
- the module I 3 takes account of zones Z of the function which are already known to the user who supplies them as an option, in order to deliver the minimum number of hidden units necessary for the neural network. This number is zero when there is no previously known area.
- An element 9 calculates the maximum between the number of hidden units delivered by the module I 2 and that delivered by the module I 3 , and sends it to the module I, intended to perform two operations.
- the first operation is the automatic calculation of the characteristics of the linear forms on each zone Zi defined previously on the nonlinear function to be emulated, obtained by calculation of statistical regressions by the method of least squares.
- the second operation is the automated design of the neural network corresponding to the non-linear function to be emulated.
- a neural network is entirely characterized by the values of four sets of parameters which are on the one hand the weights W and the biases b 1 of its hidden layer of neurons, and on the other hand the weights W 2 and the biases b 2 of its output neuron layer.
- the method according to the invention calculates the parameters of the hidden layer by associating a hidden neuron with each band of the set constituted by the pooling of the bands which characterize all the zones, so that, to this hidden neuron, corresponds a line w of weight W 1 and an element ⁇ of bias b, which are automatically fixed from the width li, the vector ni and the support point Xi defining the associated space band, according to equations (E 3 ) following:
- the matrix and the vector b are then obtained by stacking the different values 2 i and bi calculated.
- the module I 4 of the initialization box I calculated the weights and biases completely characterizing the structure and the number of hidden units of the neural network corresponding to the non-linear function.
- This network is defined by its three vectors x, y and z:
- the method according to the invention then consists in determining the initial values of the parameters ⁇ of the neuron network, whereas they are usually chosen randomly by the user. This calculation is performed by the module I 5 , which can be emulated by software for example and embedded with the neural network in the vehicle, or which can be executed offline and used with a neural network which accepts any parameters as input. .
- These initial parameters are on the one hand the size c of the hidden layer of neurons and the initial values of the parameters:
- the neural network 10 has a set of data composed of inputs IN and desired outputs OUT *. Each IN input corresponds to an OUT * output.
- the network is initialized from the initial values [c, ⁇ Q ] defined according to the invention and it delivers OUT outputs from the IN inputs.
- the user calculates an error ⁇ representative of the difference 11 between the values of outputs OUT of the network and those of the outputs OUT * expected, and sends it to an input of the neural network, as well as the learning parameters T.
- the values of the network parameters are then automatically updated in order to minimize the error ⁇ .
- IN inputs are supplied to the network 10 which outputs an OUT value using the values of its parameters.
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Feedback Control In General (AREA)
Abstract
Description
Claims
Applications Claiming Priority (3)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| FR9903830 | 1999-03-26 | ||
| FR9903830A FR2791446B1 (fr) | 1999-03-26 | 1999-03-26 | Procede d'initialisation d'un reseau de neurones |
| PCT/FR2000/000653 WO2000058909A1 (fr) | 1999-03-26 | 2000-03-17 | Procede d'initialisation d'un reseau de neurones |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| EP1166227A1 true EP1166227A1 (fr) | 2002-01-02 |
Family
ID=9543703
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| EP00910976A Withdrawn EP1166227A1 (fr) | 1999-03-26 | 2000-03-17 | Procede d'initialisation d'un reseau de neurones |
Country Status (3)
| Country | Link |
|---|---|
| EP (1) | EP1166227A1 (fr) |
| FR (1) | FR2791446B1 (fr) |
| WO (1) | WO2000058909A1 (fr) |
Families Citing this family (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112446098B (zh) * | 2020-12-03 | 2023-08-25 | 武汉第二船舶设计研究所(中国船舶重工集团公司第七一九研究所) | 海洋装备中推进器的极限性能模拟方法 |
| CN117708592B (zh) * | 2023-12-12 | 2024-05-07 | 清新文化艺术有限公司 | 基于文化创意的艺术科技融合数字平台 |
Family Cites Families (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JPH06274660A (ja) * | 1993-03-18 | 1994-09-30 | Hitachi Ltd | 認識又は診断方法 |
| US5596681A (en) * | 1993-10-22 | 1997-01-21 | Nippondenso Co., Ltd. | Method of determining an optimal number of neurons contained in hidden layers of a neural network |
| DE19509186A1 (de) * | 1995-03-14 | 1996-09-19 | Siemens Ag | Einrichtung zum Entwurf eines neuronalen Netzes sowie neuronales Netz |
-
1999
- 1999-03-26 FR FR9903830A patent/FR2791446B1/fr not_active Expired - Fee Related
-
2000
- 2000-03-17 WO PCT/FR2000/000653 patent/WO2000058909A1/fr not_active Ceased
- 2000-03-17 EP EP00910976A patent/EP1166227A1/fr not_active Withdrawn
Non-Patent Citations (1)
| Title |
|---|
| See references of WO0058909A1 * |
Also Published As
| Publication number | Publication date |
|---|---|
| FR2791446B1 (fr) | 2001-09-21 |
| WO2000058909A1 (fr) | 2000-10-05 |
| FR2791446A1 (fr) | 2000-09-29 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| EP0546624B1 (fr) | Système de traitement de données opérant avec une fonction non linéaire formée de morceaux | |
| EP3877912B1 (fr) | Procédé de construction de réseau de neurones | |
| EP0521548B1 (fr) | Procédé et dispositif avec un réseau neuronal pour classer des données | |
| EP1470501B1 (fr) | Procedes et systemes de recherche et d'association de ressources d'information telles que des pages web | |
| Orwant | Doppelgänger goes to school: Machine learning for user modeling | |
| EP0322966A1 (fr) | Circuit et structure de réseau de neurones | |
| EP0372613B1 (fr) | Procédé de rétropropagation d'erreurs et structure de réseau de neurones | |
| EP3953662B1 (fr) | Procede de definition d'un chemin | |
| EP3736744A1 (fr) | Procédé de classification sécurisée d'une donnée d'entrée au moyen d'un réseau de neurones à convolution | |
| EP0514986B1 (fr) | Procédé d'apprentissage d'un réseau de neurones et dispositif de classification pour la mise en oeuvre de ce procédé | |
| CN113362852B (zh) | 一种用户属性识别方法和装置 | |
| EP0875032B1 (fr) | Procede d'apprentissage generant des reseaux de neurones de petites tailles pour la classification de donnees | |
| EP4162409A1 (fr) | Procédé de génération d'un système d'aide à la décision et systèmes associés | |
| EP3633548B1 (fr) | Procédé de classification sécurisée d'une donnée d'entrée au moyen d'un réseau de neurones à convolution | |
| FR3057378A1 (fr) | Systeme de detection de fraude dans un flux de donnees | |
| WO2000058909A1 (fr) | Procede d'initialisation d'un reseau de neurones | |
| EP0401927A1 (fr) | Méthode d'apprentissage, réseau de neurones et ordinateur pour simuler ledit réseau de neurones | |
| EP3805947A1 (fr) | Procédé de gestion d'une base de données partagée par un groupe d'applications, produit programme d'ordinateur et système embarqué associés | |
| US11675582B2 (en) | Neural networks to identify source code | |
| US11830064B2 (en) | Method and system for credit risk identification | |
| EP3955167A1 (fr) | Architecture de calcul reconfigurable pour la mise en oeuvre de réseaux de neurones artificiels | |
| EP0401926B1 (fr) | Procédé de traitement, stucture de réseau de neurones mettant en oeuvre le procédé et ordinateur pour simuler ladite structure de réseau de neurones | |
| US11971900B2 (en) | Rule-based data transformation using edge computing architecture | |
| EP0681246A1 (fr) | Procédé et dispositif d'extraction d'un sous-ensemble d'objets optimisant une mesure, en utilisant un réseau de neurones | |
| EP0432008A1 (fr) | Réseau neuronal électronique et procédé d'apprentissage |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PUAI | Public reference made under article 153(3) epc to a published international application that has entered the european phase |
Free format text: ORIGINAL CODE: 0009012 |
|
| 17P | Request for examination filed |
Effective date: 20010924 |
|
| AK | Designated contracting states |
Kind code of ref document: A1 Designated state(s): AT BE CH CY DE DK ES FI FR GB GR IE IT LI LU MC NL PT SE |
|
| GRAG | Despatch of communication of intention to grant |
Free format text: ORIGINAL CODE: EPIDOS AGRA |
|
| RTI1 | Title (correction) |
Free format text: METHOD AND APPARATUS FOR INITIALIZING A NEURAL NETWORK |
|
| 17Q | First examination report despatched |
Effective date: 20020419 |
|
| RAP1 | Party data changed (applicant data changed or rights of an application transferred) |
Owner name: RENAULT S.A.S. |
|
| STAA | Information on the status of an ep patent application or granted ep patent |
Free format text: STATUS: THE APPLICATION HAS BEEN WITHDRAWN |
|
| 18W | Application withdrawn |
Withdrawal date: 20021023 |