WO2021089591A1 - Procédé d'entraînement d'un réseau de neurones artificiels, programme informatique, support d'enregistrement, dispositif, réseau de neurones artificiels et utilisation du réseau de neurones artificiels - Google Patents

Procédé d'entraînement d'un réseau de neurones artificiels, programme informatique, support d'enregistrement, dispositif, réseau de neurones artificiels et utilisation du réseau de neurones artificiels Download PDF

Info

Publication number
WO2021089591A1
WO2021089591A1 PCT/EP2020/080905 EP2020080905W WO2021089591A1 WO 2021089591 A1 WO2021089591 A1 WO 2021089591A1 EP 2020080905 W EP2020080905 W EP 2020080905W WO 2021089591 A1 WO2021089591 A1 WO 2021089591A1
Authority
WO
WIPO (PCT)
Prior art keywords
artificial neural
neural network
uncertainty
training
prediction
Prior art date
Application number
PCT/EP2020/080905
Other languages
German (de)
English (en)
Inventor
Di FENG
Lars Rosenbaum
Original Assignee
Robert Bosch Gmbh
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Robert Bosch Gmbh filed Critical Robert Bosch Gmbh
Publication of WO2021089591A1 publication Critical patent/WO2021089591A1/fr

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
    • G06N3/08—Learning methods

Definitions

  • the present invention relates to a method for training an artificial neural network, a corresponding computer program, a corresponding electronic storage medium, a corresponding device, an artificial neural network trained according to the method of the present invention and an application of the artificial neural network trained in this way.
  • the artificial neural networks used In order to, among other things, To select the correct reaction pattern, it is necessary that the artificial neural networks used not only fulfill their primary tasks of regression and classification efficiently and effectively, but also determine a measure of the uncertainty of their predictions.
  • DE 102018 220941.3 discloses a method for training a KI module as a function of a loss function, which takes into account an uncertainty of a prediction of the KI module determined by the KI module.
  • a KI module is designed to translate a set of input variables through an internal processing chain into at least a prediction of an output variable. This does not rule out that the KI module determines further output variables from the same input variables.
  • the KI module can be designed, for example, to determine the distance to an object as a continuous output variable and at the same time the type of this object from image data or other physical measurement data obtained by measuring, e.g. using at least one sensor, a spatial detection area Classify object. This does not rule out that, as an alternative or in combination with this, the continuous output variable is also used for classification tasks.
  • the behavior of the internal processing chain is determined by parameters. These parameters are learned during the training, and the Kl module works on the basis of the learned parameters in later operation.
  • One version of the KI module can include an artificial neural network.
  • an artificial neural network is to be understood as a network of artificial neurons for information processing.
  • Artificial neural networks essentially go through three phases. In an initial phase, a basic topology is specified, mostly depending on the primary task (e.g. classification, regression). This is followed by a training phase in which the basic topology for the efficient solution of the task is learned using training data. The topology of the network can also be adapted during the training phase. Training data is characterized by it assumes that the desired output data is typically available for the input data. Finally, there is an application phase in which the trained network is applied to input data for which there is no desired output data. The output data of the trained network then represent the output data searched for according to the task.
  • Artificial neural networks should determine an uncertainty in their predictions that corresponds to the natural frequency of the correct predictions. However, it has been shown that the previously known artificial neural networks are either too certain (eng .: over-confident) or too uncertain (eng .: under-confident) with regard to the question of the uncertainty of their predictions. This gives rise to the problem of mis-calibrated uncertainty, which, among other things, can lead to wrong decisions in the field of automated driving.
  • the present invention creates a method for training an artificial neural network as a function of a loss function, the artificial neural network being set up to determine a prediction and an uncertainty associated with the prediction as a function of an input, the loss function the ascertained uncertainty considered.
  • the method is characterized and differs in particular from the subject of DE 102018220941.3 in an unobvious way in that the uncertainty ascertained is a limited uncertainty.
  • a loss function can be understood to be a function that quantifies a measure of the distance between the predicted output variable and the expected output data of the training data set.
  • Typical loss functions in the field of artificial neural networks are the L t or Laplace loss and the L 2 or Gaussian loss
  • the determination of the uncertainties in comparison to the methods from the prior art does not require any additional computing resources (computational cost) during the application of the artificial neural network and the training time of the artificial neural network is only marginally extended.
  • the ascertained uncertainty is limited by means of a ground truth for the uncertainty.
  • ground truth can be understood as a value that is viewed as a correct value.
  • the predefined output data are regarded as ground truth.
  • the loss function of the primary task can be any standard loss function, such as the loss function of cross entropy (eng .: Cross Entropy Loess) for classification problems or the L t or Laplace loss function (eng .: L t Loess) for regression problems act. It is also conceivable that the primary task of the artificial neural network has both classification components and regression components and accordingly, the loss function includes terms corresponding to the primary task.
  • the calibration loss function can be described as follows:
  • the ground truth is formed as a function of the prediction determined and one of the output data assigned to the input data, in particular the training data set.
  • Another aspect of the present invention is a computer program which is set up to carry out all steps of the method according to the present invention.
  • Another aspect of the present invention is an electronic storage medium on which the computer program according to the present invention is stored.
  • Another aspect of the present invention is a device which is set up to carry out all steps of the method according to the present invention.
  • Another aspect of the present invention is an artificial neural network which is trained by means of the method according to the present invention.
  • a technical system can be a system for at least partially automated driving, a system for automated inspection, for example a production process, a system for automated access control, a system for monitoring buildings and public spaces with object tracking.
  • FIG. 1 shows a block diagram of an embodiment of the use of an artificial neural network
  • FIG. 2 shows a flowchart of an embodiment of a method for training an artificial neural network according to the present invention
  • FIGS. 3a to c each show a graph of a calibration plot
  • Figure 4 is a block diagram of areas of application of the present invention.
  • FIG. 1 shows a block diagram which shows the use of an artificial neural network in a technical system 1.
  • the technical system 1 shown is a vehicle.
  • the artificial neural network is used as part of a driver assistance function, for example an at least partially automated control of the vehicle.
  • the components required for this are at least one sensor 10 for measuring a detection area and for transmitting the detected data to a processing unit 11 in the vehicle.
  • the artificial neural network is used on the processing unit 11, for example for a classification task, such as object detection, or a regression task, such as distance estimation.
  • the processing unit 11 is also set up for this purpose to send corresponding signals to a control unit 12, for example for a braking system, in order to carry out a driver assistance function, such as automatic emergency braking.
  • the artificial neural network used for this purpose was trained on an external processing unit 13.
  • the trained network has been transferred to the processing unit 11 for execution.
  • the embodiment shown does not exclude the fact that the technical system 1 comprises more than one sensor 10 and uses it to measure the detection area.
  • the embodiment shown does not exclude that the processing unit 11 and the control unit 12 are implemented by means of a hardware unit or are implemented distributed over a plurality of hardware units.
  • the embodiment shown does not exclude that the data of the sensor 10 in an alternative or supplementary embodiment are alternatively or additionally transmitted to a further external processing unit, for example to a cloud.
  • the embodiment shown does not rule out that the external processing unit 13 is at least partially arranged in the technical system 1 in an alternative embodiment.
  • the illustrated embodiment does not exclude the fact that the artificial neural network used is (further) trained on the processing unit 11 in the technical system during the runtime.
  • FIG. 2 shows an embodiment of a method 200 for training an artificial neural network according to the present invention.
  • Training data sets are typically used for training an artificial neural network. Training datasets are characterized by the fact that they hold expected output data for a quantity of input data.
  • step 201 On the basis of such training data records, in step 201 a predetermined amount of input data of the training data records is given to the artificial neural network to be trained.
  • step 202 a prediction of the output data is determined for each input date using the artificial neural network to be trained.
  • an uncertainty of the prediction is additionally determined by means of the artificial neural network.
  • the artificial neural network to be trained is designed in such a way that, in particular in comparison to the prior art, the ascertained uncertainty is restricted. This can be achieved in that the ascertained uncertainty is taken into account when evaluating a loss function (see step 205) by means of the calibration loss function or loss term below.
  • Lcalib - II Px Px ⁇ with p x as the uncertainty determined by the artificial neural network when entering the input data x of the training data set, p x as the corresponding ground truth and INI as any loss function, such as the L ⁇ - or L 2 - Loss.
  • step 204 the prediction determined by the artificial neural network is compared with the expected output data of the training data records assigned to the respective input date.
  • a loss function is evaluated, which takes into account both the deviation of the prediction determined in the comparison from the expected output data and the determined uncertainty of the prediction.
  • the loss function L to be evaluated can be described as follows.
  • L std a corresponding loss function, for example a standard loss function
  • the primary task is a classification task, such as object detection or classification
  • the so-called "softmax function" i. H. the normalized exponential function can be provided as a loss term or loss function.
  • step 206 the parameters of the artificial neural network are optimized in such a way that when the input data is re-entered, the deviation determined by the loss function turns out to be smaller.
  • FIG. 3a shows a graph of a calibration plot. The uncertainty of the forecast is plotted on the abscissa. The probability of a correct prediction is plotted on the ordinate.
  • FIG. 3b shows a graph of a calibration plot of a specific artificial neural network trained by means of a training method from the prior art.
  • the solid diagonal curve shows an optimally calibrated artificial neural network.
  • the dashed curve stands for the calibration of the artificial neural network. This curve clearly shows that the artificial neural network has an incorrectly calibrated uncertainty, since the areas of the predictions O that are too reliable and the areas of predictions U that are too uncertain deviate significantly from the optimal calibration.
  • FIG. 3c shows a graph of a calibration plot of the artificial neural network according to FIG. 3b now trained with the method according to the present invention.
  • the dashed curve stands for the calibration of the artificial neural network. In comparison to the corresponding curve of the graph in FIG. 3b, this curve is now obviously closer to the curve of an optimally calibrated artificial neural network.
  • An improved calibration of artificial neural networks can therefore be effectively achieved by means of the method of the present invention.
  • FIG. 4A shows a block diagram of a system for automatic inspection 41, for example during the visual inspection of a production process 411.
  • an artificial neural network trained by means of the training method according to the present invention can be used to classify defective parts 412 as well as an indication of Determine the degree of uncertainty of the classification.
  • the present invention effectively helps to reduce the false positive rate, ie the proportion of parts 412 incorrectly classified as defective.
  • FIG. 4B shows a block diagram of an automated system 42, specifically an automated lawnmower.
  • Object recognition can be improved by using an artificial neural network trained by means of a method according to the present invention. This leads to improved results when using the automated system 42.
  • FIG. 4C shows a block diagram of a system for automated access control 43, for example by means of optical identification of people and the automated opening of doors 431.
  • an artificial neural network trained by means of a method of the present invention in addition to recognizing a person and thus comparing the access authorization also a measure for the uncertainty of the detection can be determined. This dimension can be taken into account when granting access.
  • FIG. 4D shows a block diagram of a system for monitoring 44 buildings or public places.
  • An artificial neural network trained by means of a method according to the present invention can be used to output an uncertainty of the object tracking in addition to the object tracking. In this way, for example, the quality or the reliability of the object tracking can be effectively taken into account.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Health & Medical Sciences (AREA)
  • Computing Systems (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Molecular Biology (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Image Analysis (AREA)
  • Feedback Control In General (AREA)

Abstract

L'invention concerne un procédé (200) d'entraînement d'un réseau de neurones artificiels en fonction d'une fonction de perte, le réseau de neurones artificiels étant conçu de manière à déterminer, en fonction d'une entrée, une prédiction et une incertitude associée à la prédiction, la fonction de perte tenant compte de l'incertitude déterminée, l'incertitude déterminée étant une incertitude limitée.
PCT/EP2020/080905 2019-11-08 2020-11-04 Procédé d'entraînement d'un réseau de neurones artificiels, programme informatique, support d'enregistrement, dispositif, réseau de neurones artificiels et utilisation du réseau de neurones artificiels WO2021089591A1 (fr)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
DE102019217300.4 2019-11-08
DE102019217300.4A DE102019217300A1 (de) 2019-11-08 2019-11-08 Verfahren zum Trainieren eines künstlichen neuronalen Netzes, Computerprogramm, Speichermedium, Vorrichtung, künstliches neuronales Netz und Anwendung des künstlichen neuronalen Netzes

Publications (1)

Publication Number Publication Date
WO2021089591A1 true WO2021089591A1 (fr) 2021-05-14

Family

ID=73059940

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/EP2020/080905 WO2021089591A1 (fr) 2019-11-08 2020-11-04 Procédé d'entraînement d'un réseau de neurones artificiels, programme informatique, support d'enregistrement, dispositif, réseau de neurones artificiels et utilisation du réseau de neurones artificiels

Country Status (2)

Country Link
DE (1) DE102019217300A1 (fr)
WO (1) WO2021089591A1 (fr)

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
DE102022207786A1 (de) 2022-07-28 2024-02-08 Robert Bosch Gesellschaft mit beschränkter Haftung Verfahren zum Trainieren eines künstlichen neuronalen Netzes

Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
DE102018220941A1 (de) 2018-12-04 2020-06-04 Robert Bosch Gmbh Auswertung von Messgrößen mit KI-Modulen unter Berücksichtigung von Messunsicherheiten

Patent Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
DE102018220941A1 (de) 2018-12-04 2020-06-04 Robert Bosch Gmbh Auswertung von Messgrößen mit KI-Modulen unter Berücksichtigung von Messunsicherheiten

Non-Patent Citations (5)

* Cited by examiner, † Cited by third party
Title
AUS FENG ET AL.: "Towards safe autonomous driving: Capture uncertainty in the deep neural network for lidar 3d vehicle detection", ITSC, 2018
DI FENG ET AL: "Towards Safe Autonomous Driving: Capture Uncertainty in the Deep Neural Network For Lidar 3D Vehicle Detection", ARXIV.ORG, CORNELL UNIVERSITY LIBRARY, 201 OLIN LIBRARY CORNELL UNIVERSITY ITHACA, NY 14853, 14 April 2018 (2018-04-14), XP081551113 *
FENG DI ET AL: "Leveraging Heteroscedastic Aleatoric Uncertainties for Robust Real-Time LiDAR 3D Object Detection", 2019 IEEE INTELLIGENT VEHICLES SYMPOSIUM (IV), IEEE, 9 June 2019 (2019-06-09), pages 1280 - 1287, XP033606084, DOI: 10.1109/IVS.2019.8814046 *
FENG ET AL., LEVERAGING HETEROSCEDASTIC ALEATORIC UNCERTAINTIES FOR ROBUST REAL-TIME LIDAR 3D OBJECT DETECTION, vol. IV, 2019
LE MICHAEL TRUONG ET AL: "Uncertainty Estimation for Deep Neural Object Detectors in Safety-Critical Applications", 2018 21ST INTERNATIONAL CONFERENCE ON INTELLIGENT TRANSPORTATION SYSTEMS (ITSC), IEEE, 4 November 2018 (2018-11-04), pages 3873 - 3878, XP033470240, ISBN: 978-1-7281-0321-1, [retrieved on 20181207], DOI: 10.1109/ITSC.2018.8569637 *

Also Published As

Publication number Publication date
DE102019217300A1 (de) 2021-05-12

Similar Documents

Publication Publication Date Title
DE102018128289A1 (de) Verfahren und vorrichtung für eine autonome systemleistung und zur einstufung
EP3646184A1 (fr) Procédé, dispositif et support d'enregistrement pouvant être lu par un ordinateur pourvu d'instructions pour la résolution d'une redondance de deux modules redondants ou plus
DE102018220941A1 (de) Auswertung von Messgrößen mit KI-Modulen unter Berücksichtigung von Messunsicherheiten
DE102017210787A1 (de) Verfahren und Vorrichtung zum Ermitteln von Anomalien in einem Kommunikationsnetzwerk
DE102019206720B4 (de) Überwachung eines KI-Moduls einer Fahrfunktion eines Fahrzeugs
WO2021089591A1 (fr) Procédé d'entraînement d'un réseau de neurones artificiels, programme informatique, support d'enregistrement, dispositif, réseau de neurones artificiels et utilisation du réseau de neurones artificiels
WO2019149664A1 (fr) Procédé de détermination d'une variation temporelle d'une grandeur de mesure, système de pronostic, système de commande d'actionneur, procédé d'apprentissage du système de commande d'actionneur, système d'apprentissage, programme d'ordinateur et support d'informations lisible par machine
DE102019209457A1 (de) Verfahren zum Trainieren eines künstlichen neuronalen Netzes, künstliches neuronales Netz, Verwendung eines künstlichen neuronalen Netzes sowie entsprechendes Computerprogramm, maschinenlesbares Speichermedium und entsprechende Vorrichtung
DE102017219269A1 (de) Klassifizierung mit automatischer Auswahl aussichtsreicher Lerndaten
EP3467722A1 (fr) Configuration d'un dispositif d'aide à la conduite d'un véhicule automobile doté d'un réseau neuronal en fonctionnement
DE102022201679A1 (de) Verfahren und Vorrichtung zum Trainieren eines neuronalen Netzes
DE102021200789A1 (de) Computer-implementiertes Verfahren und Vorrichtung zur Manipulationserkennung für Abgasnachbehandlungssysteme mithilfe Verfahren künstlicher Intelligenz
WO2020216621A1 (fr) Apprentissage de modules aptes à l'apprentissage avec des données d'apprentissage dont les étiquettes sont bruitées
DE102018216078A1 (de) Verfahren und Vorrichtung zum Betreiben eines Steuerungssystems
DE102022204863A1 (de) Detektion von Anomalien in einem technischen System durch Überwachung mit mehreren Sensoren
WO2018091160A1 (fr) Système de surveillance d'une fonction d'un dispositif de détection d'un véhicule à moteur
DE102019131639B4 (de) System zum Bereitstellen eines Erklärungsdatensatzes für ein KI-Modul
DE102022205084B3 (de) Verfahren, Computerprogramm und Vorrichtung zur Umfeldwahrnehmung im Fahrzeug sowie entsprechendes Fahrzeug
DE102016113310A1 (de) Verfahren zur Bewertung von Aussagen einer Mehrzahl von Quellen zu einer Mehrzahl von Fakten
DE102021207493A1 (de) Verfahren zur Unterstützung eines Betriebs eines Fahrzeuges mit einer Sensoreinheit, Computerprogrammprodukt sowie System
DE102022200729A1 (de) Verfahren zum Ermitteln eines aktuellen Steuermodus eines Lenkrades
DE102021205097A1 (de) Computerimplementiertes Verfahren und System zur Bestimmung einer Kostenfunktion
DE102022212455A1 (de) Verfahren zum Trainieren eines Algorithmus des maschinellen Lernens
DE102022208384A1 (de) Verfahren zum Ermitteln eines Qualitätszustands eines Prüfobjekts
DE102021208197A1 (de) Verfahren und Vorrichtung zum Erstellen eines maschinellen Lernsystems

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 20800919

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 20800919

Country of ref document: EP

Kind code of ref document: A1