Granted Patent B2

US 12,455,945 · App. 17/654,417 · Granted Oct 28, 2025

Inventors: Lukas Lange (Pforzheim, DE); Jannik Stroetgen (Karlsruhe, DE)

Assignee: ROBERT BOSCH GMBH

G06F18/2413G06F18/214G06F18/2431

Abstract

Claims (61)

1. A computer-implemented method for classifying data sets, the method comprising:

predefining a model for solving a task;

predefining a plurality of training data sets;

defining, for each training data set from the plurality of training data sets, a respective trained model for solving the task by pretraining the model on the training data set and training the model on a reference training data set;

determining a trained reference model for solving the task by training the model on the reference training data set without pretraining with the plurality of training data sets;

determining, for each respective trained model, a respective quality of solving the task;

determining, for the trained reference model, a reference quality of solving the task;

classifying each respective trained model as suitable or unsuitable for the pretraining as a function of a deviation of the respective quality from the reference quality;

determining, in the plurality of training data sets, nearest neighbors of a data set of the plurality of training data sets; and

either classifying the data set as suitable or unsuitable for the pretraining as a function of how the trained models, which have been trained with the nearest neighbors, are classified, or classifying the nearest neighbors of the data set as suitable for the pretraining, wherein for training of the model, the model is pretrained with at least one of the training data sets, which is classified as suitable for the pretraining, and the model for solving the task is subsequently trained with the data set.

2. A computer-implemented method for classifying data sets, the method comprising:

predefining a model for solving a task;

predefining a plurality of training data sets;

defining, for each training data set from the plurality of training data sets, a respective trained model for solving the task by pretraining the model on the training data set and training the model on a reference training data set;

determining a trained reference model for solving the task by training the model on the reference training data set without pretraining with the plurality of training data sets;

determining, for each respective trained model, a respective quality of solving the task;

determining, for the trained reference model, a reference quality of solving the task;

classifying each respective trained model as suitable or unsuitable for the pretraining as a function of a deviation of the respective quality from the reference quality;

determining, in the plurality of training data sets, nearest neighbors of a data set of the plurality of training data sets; and

either classifying the data set as suitable or unsuitable for the pretraining as a function of how the trained models, which have been trained with the nearest neighbors, are classified, or classifying the nearest neighbors of the data set as suitable for the pretraining, wherein for training of the model, the model is pretrained for solving the task with the data set when the data set is classified, by a classifier, as suitable for the pretraining, and otherwise the model for solving the task is not pretrained with the data set.

3. A computer-implemented method for classifying data sets, the method comprising:

predefining a model for solving a task;

predefining a plurality of training data sets;

defining, for each training data set from the plurality of training data sets, a respective trained model for solving the task by pretraining the model on the training data set and training the model on a reference training data set;

determining a trained reference model for solving the task by training the model on the reference training data set without pretraining with the plurality of training data sets;

determining, for each respective trained model, a respective quality of solving the task;

determining, for the trained reference model, a reference quality of solving the task;

classifying each respective trained model as suitable or unsuitable for the pretraining as a function of a deviation of the respective quality from the reference quality;

determining, in the plurality of training data sets, nearest neighbors of a data set of the plurality of training data sets; and

either classifying the data set as suitable or unsuitable for the pretraining as a function of how the trained models, which have been trained with the nearest neighbors, are classified, or classifying the nearest neighbors of the data set as suitable for the pretraining, wherein for each training data set of the plurality of training data sets, at least one distance from the data set is determined, either a predefined number of training data sets from the plurality of training data sets being determined as nearest neighbors, whose distance is less than the other of the training data sets from the plurality of training data sets, or the training data sets from the plurality of training data sets being determined as nearest neighbors, whose distance is less than a predefined distance.

4. The method as recited in claim 3 , wherein for each training data set of the plurality of training data sets, a plurality of distances from the data set is determined using various distance measures, the training data sets from the plurality of training data sets being determined as nearest neighbors, for which at least one distance from the plurality of distances is less than the predefined distance.

5. A computer-implemented device for classifying data sets, the computer-implemented device having a processor that executes computer-readable instructions configured to:

predefine a model for solving a task;

predefine a plurality of training data sets;

define, for each training data set from the plurality of training data sets, a respective trained model for solving the task by pretraining the model on the training data set and training the model on a reference training data set;

determine a trained reference model for solving the task by training the model on the reference training data set without pretraining with the plurality of training data sets;

determine, for each respective trained model, a respective quality of solving the task;

determine, for the trained reference model, a reference quality of solving the task;

classify each respective trained model as suitable or unsuitable for the pretraining as a function of a deviation of the respective quality from the reference quality;

determine, in the plurality of training data sets, nearest neighbors of a data set of the plurality of training data sets; and

either classify the data set as suitable or unsuitable for the pretraining as a function of how the trained models, which have been trained with the nearest neighbors, are classified, or classify the nearest neighbors of the data set as suitable for the pretraining, wherein for training of the model, the model is pretrained with at least one of the training data sets, which is classified as suitable for the pretraining, and the model for solving the task is subsequently trained with the data set.

6. A computer-implemented device for classifying data sets, the computer-implemented device having a processor that executes computer-readable instructions configured to:

predefine a model for solving a task;

predefine a plurality of training data sets;

define, for each training data set from the plurality of training data sets, a respective trained model for solving the task by pretraining the model on the training data set and training the model on a reference training data set;

determine a trained reference model for solving the task by training the model on the reference training data set without pretraining with the plurality of training data sets;

determine, for each respective trained model, a respective quality of solving the task;

determine, for the trained reference model, a reference quality of solving the task;

classify each respective trained model as suitable or unsuitable for the pretraining as a function of a deviation of the respective quality from the reference quality;

determine, in the plurality of training data sets, nearest neighbors of a data set of the plurality of training data sets; and

either classify the data set as suitable or unsuitable for the pretraining as a function of how the trained models, which have been trained with the nearest neighbors, are classified, or classify the nearest neighbors of the data set as suitable for the pretraining, wherein the computer-readable instructions executed by the processor of the computer-implemented device are configured to classify the data set, and configured to determine the model for solving a task, which is trained or pretrained with the data set, when the data set is classified as suitable for the pretraining, and otherwise to determine the model without training or pretraining with the data set.

7. A non-transitory computer-readable medium on which is stored a computer program including computer-readable instructions for classifying data sets, the instructions, when executed by a computer, causing the computer to perform the following steps:

predefining a model for solving a task;

predefining a plurality of training data sets;

defining, for each training data set from the plurality of training data sets, a respective trained model for solving the task by pretraining the model on the training data set and training the model on a reference training data set;

determining a trained reference model for solving the task by training the model on the reference training data set without pretraining with the plurality of training data sets;

determining, for each respective trained model, a respective quality of solving the task;

determining, for the trained reference model, a reference quality of solving the task;

classifying each respective trained model as suitable or unsuitable for the pretraining as a function of a deviation of the respective quality from the reference quality;

determining, in the plurality of training data sets, nearest neighbors of a data set of the plurality of training data sets; and

either classifying the data set as suitable or unsuitable for the pretraining as a function of how the trained models, which have been trained with the nearest neighbors, are classified, or classifying the nearest neighbors of the data set as suitable for the pretraining, wherein for training of the model, the model is pretrained with at least one of the training data sets, which is classified as suitable for the pretraining, and the model for solving the task is subsequently trained with the data set.

Assignments (1)

Priority Claims (1)

Continuity (1)

References Cited (12)