Granted Patent B2

US 12,530,572 · App. 17/204,188 · Granted Jan 20, 2026

Inventors: Bence Tilk (Oslo, NO); Csaba Nemes (Dunakeszi, HU)

Assignee: CONTINENTAL AUTONOMOUS MOBILITY GERMANY GMBH

G06N3/08G06F18/2155G06F18/2163G06F18/241

Abstract

Claims (26)

1 . A computer-implemented method for configuring neural network model, wherein the method comprises:

providing the neural network model;

splitting the neural network model into a first portion and a second portion, the second portion comprising a first head, configured to classify a first type of classification data and a second head, configured to classify the second type of classification data, wherein the second type of classification data is a target data;

pre-processing, in a training phase, the second type of classification data in the first portion,

processing the pre-processed second type of classification data in the first and second heads and determining a first result of the processing of second type of classification data in the first head and a second result of the processing of second type of classification data in the second head;

calculating a consistency between the first result and the second result, both the first result and the second result relating to the target data;

calculating a first head accuracy for the first head;

only configuring the neural network model by updating a value of at least one parameter of the second head and at least one parameter of the first head based on the calculated consistency when the first head accuracy is greater than a pre-defined threshold, such that the first head has a substantiated basis in order to provide knowledge transfer to the second head; and

training the neural network model for detecting or classifying objects or features in images of an environment of a vehicle in an advanced driver assistance system (ADAS) or an automated driving (AD) camera system,

wherein at least portions of the first type of classification data and the second type of classification data are labeled, and wherein labels of the second type of classification data denominated as target labels are prioritized over labels of the first type of classification data denominated as source labels by mixing the updated values of the at least one parameter of the second head and the at least one parameter of the first head with a relation corresponding to the calculated consistency, to indicate that patterns of the target labels are more important than patterns of the source labels.

2 . The computer-implemented method according to claim 1 , wherein the artificial intelligence module is embodied as a neural network.

3 . The computer-implemented method according to claim 1 , wherein the first portion comprises encoder layers, and the second portion comprises classification layers.

4 . The computer-implemented method according to claim 1 , wherein the first type of classification data is correlated with the second type of classification data.

5 . The computer-implemented method according to claim 4 , wherein the images show the same scene, but a first image dataset is recorded at a day time and a second image dataset is recorded at a night time.

6 . The computer-implemented method according to claim 1 , wherein the neural network model provides domain adaptation, and the first type of classification data is source data.

7 . The computer-implemented method according to claim 1 , wherein, previous to the step of pre-processing the second type of classification data in the first portion, a training of the first portion and training of the first head is performed.

8 . The computer-implemented method according to claim 1 , wherein a first head standard loss is calculated for the first head, and the step configuring the neural network model by updating a value of at least one parameter of the second head and at least one parameter of the first head based on the calculated consistency is only performed if the first head standard loss is lower than a pre-defined threshold.

9 . The computer-implemented method according to claim 1 , wherein the parameter values to be updated in the first and/or second head are weights of the neural network.

10 . The computer-implemented method according to claim 1 , wherein the second type of classification data is partly labelled and partly unlabelled.

11 . The computer-implemented method according to claim 1 , wherein the output of the first and the second head is categorical, and the calculation of the consistency takes place on the basis of the Jensen-Shannon divergence.

12 . The computer-implemented method according to claim 2 , wherein the output of the first and the second head is continuous, and the calculation of the consistency is carried out on the basis of a mean square error.

13 . A neural network stored on a non-transitory computer readable medium, the neural network having a configuration achieved with the computer-implemented method according to claim 1 .

14 . Use of the neural network according to claim 13 for detecting a lane, an object, or a key point, for semantic segmentation, for classifying text, an acoustic pattern, or an object, or to estimate a pose.

15 . A computer program element stored on a non-transitory computer readable medium, the program element comprising instructions which, when the program element is executed by a computer, cause the computer to carry out the steps of the method according to claim 1 .

16 . A non-transitory computer-readable medium comprising instructions which, when executed by a computer, cause the computer to carry out the steps of the method according to claim 1 .

17 . The computer-implemented method according to claim 1 , wherein the artificial intelligence module is embodied as a convolutional neural network.

Assignments (2)

Priority Claims (1)

Continuity (1)

References Cited (25)