Method and device for training a neural network

Patent metadata

Field	Value
Publication number	US-12548314-B2
Application number	US-202318330857-A
Country	US
Kind code	B2
Filing date	Jun 7, 2023
Priority date	Jun 8, 2022
Publication date	Feb 10, 2026
Grant date	Feb 10, 2026

Abstract

A method for training a first neural network is disclosed. The neural network is configured to ascertain, based on sensor signals of a technical system, an output signal characterizing a classification and/or a regression result regarding the sensor signal. The method includes (i) during operation of the technical system, receiving a sensor signal of the technical system, (ii) ascertaining a first output signal by way of the first neural network and based on the sensor signal, (iii) ascertaining a second output signal by way of a second neural network and based on the sensor signal, wherein the second neural network has a different architecture than the first neural network, and (iv) training the first neural network by adjusting parameters of the first neural network, wherein the first neural network is trained as a function of the second output signal.

First claim

What is claimed is:

1 . A method for training a first neural network to ascertain output signals characterizing classification and/or regression results based on sensor signals of a technical system, the method comprising:
(a) during operation of the technical system, receiving a first sensor signal of the technical system; (b) ascertaining a first output signal by providing the first sensor signal as an input to the first neural network and determining the first output signal based on the first sensor signal using the first neural network; (c) ascertaining a second output signal by providing the first sensor signal as an input to the second neural network and determining the second output signal based on the first sensor signal using the second neural network, the second neural network having been previously trained to ascertain output signals characterizing classification and/or regression results based on sensor signals of the technical system, the second neural network having a different architecture than the first neural network; and (d) training the first neural network by adjusting parameters of the first neural network based on a comparison between the first output signal and the second output signal, such that the first neural network is trained to mimic behavior of the second neural network.

2 . The method according to claim 1 , wherein the adjustment is ascertained based on a difference between the first output signal and the second output signal.

3 . The method according to claim 1 , wherein a difference in the architecture of the first neural network and the second neural network is such that the first neural network requires less computing capacity and/or less memory to ascertain the first output signal.

4 . The method according to claim 1 , wherein at least one layer of the first neural network corresponds in its parameterization to a layer of the second neural network and parameters of the layer of the first neural network are not adjusted during training.

5 . The method according to claim 1 , wherein step (b) and step (c) are carried out by an embedded computing unit of the technical system.

6 . The method according to claim 5 , wherein step (d) is carried out by a computer outside of the technical system.

7 . The method according to claim 1 , wherein control signals of the technical system are ascertained based on the second output signal.

8 . The method according to claim 1 , wherein the first sensor signal is provided by a camera sensor and characterizes an image.

9 . The method according to claim 1 , wherein the technical system is an at least partially automated vehicle or a robot.

10 . A training device which is configured to carry out the method according to claim 1 .

11 . The method according to claim 1 , wherein a processor executes a computer program to carry out the method.

12 . A machine-readable storage medium on which the computer program according to claim 11 is stored.

Assignees

• Bosch Gmbh Robert

Inventors

• Branz Thomas
• Hanselmann Markus
• Genssle Andreas

Classifications

• G06N3/045

  Combinations of networks · CPC title

• G06N3/008

  based on physical entities controlled by simulated intelligence so as to replicate intelligent life forms, e.g. based on robots replicating pets or humans in their appearance or behaviour · CPC title

• G06N3/0985

  Hyperparameter optimisation; Meta-learning; Learning-to-learn · CPC title

• G06N3/086

  using evolutionary algorithms, e.g. genetic algorithms or genetic programming · CPC title

• G06N3/082

  modifying the architecture, e.g. adding, deleting or silencing nodes or connections · CPC title