Data-based updating of the training of classifier networks

A method for training a neural network that maps input images onto an association with one or multiple classes of a predefined classification, onto a semantic segmentation, and/or onto a recognition of one or multiple objects, as output data, the method comprising the following steps:

• providing learning input images and associated learning output data onto which the neural network ideally is to map the learning input images;
• providing auxiliary input images;
• generating modifications of the auxiliary input images by introducing at least one predefined change into each auxiliary input image of the auxiliary input images;
• supplying the modifications to the neural network;
• ascertaining predictions for the predefined change, using output data onto which the neural network maps each of the modifications;
• assessing deviations of the predictions from the predefined change, using a predefined auxiliary cost function;
• optimizing parameters that characterize a behavior of the neural network, with an objective of improving the assessment by the auxiliary cost function during further processing of the auxiliary input images;
• supplying the learning input images to the neural network;
• assessing deviations of output data, onto which the neural network maps the supplied learning input images, from the learning output data, using a main cost function;
• optimizing parameters which characterize the behavior of the neural network with an objective of improving the assessment by the main cost function during further processing of the learning input images.

A method for operating a neural network that maps input images onto an association with one or multiple classes of a predefined classification, onto a semantic segmentation, and/or onto a recognition of one or multiple objects, as output data, the method comprising the following steps:

• training the neural network including:
• providing learning input images and associated learning output data onto which the neural network ideally is to map the learning input images,
• providing auxiliary input images,
• generating modifications of the auxiliary input images by introducing at least one predefined change into each auxiliary input image of the auxiliary input images,
• supplying the modifications to the neural network,
• ascertaining predictions for the predefined change, using output data onto which the neural network maps each of the modifications,
• assessing deviations of the predictions from the predefined change, using a predefined auxiliary cost function,
• optimizing parameters that characterize a behavior of the neural network, with an objective of improving the assessment by the auxiliary cost function during further processing of the auxiliary input images,
• supplying the learning input images to the neural network,
• assessing deviations of output data, onto which the neural network maps the supplied learning input images, from the learning output data, using a main cost function, and
• optimizing parameters which characterize the behavior of the neural network with an objective of improving the assessment by the main cost function during further processing of the learning input images;
• providing input images that have been recorded using at least one sensor;
• generating modifications of the input images by introducing at least one predefined change into each of the input images;
• supplying the modifications of the input images to the neural network;
• ascertaining predictions for the predefined change. using output data onto which the neural network maps each of the modifications of the input images;
• assessing deviations of the predictions from the predefined change, using the predefined auxiliary cost function;
• and
• optimizing parameters that characterize the behavior of the neural network, with the objective of improving the assessment by the auxiliary cost function during further processing of modifications of the input images.

A non-transitory machine-readable data medium on which is stored a computer program for training a neural network that maps input images onto an association with one or multiple classes of a predefined classification, onto a semantic segmentation, and/or onto a recognition of one or multiple objects, as output data, the computer program, when executed by a computer, causing the computer to perform the following steps:

• providing learning input images and associated learning output data onto which the neural network ideally is to map the learning input images;
• providing auxiliary input images;
• generating modifications of the auxiliary input images by introducing at least one predefined change into each auxiliary input image of the auxiliary input images;
• supplying the modifications to the neural network;
• ascertaining predictions for the predefined change, using output data onto which the neural network maps each of the modifications;
• assessing deviations of the predictions from the predefined change, using a predefined auxiliary cost function;
• optimizing parameters that characterize a behavior of the neural network, with an objective of improving the assessment by the auxiliary cost function during further processing of the auxiliary input images;
• supplying the learning input images to the neural network;
• assessing deviations of output data, onto which the neural network maps the supplied learning input images, from the learning output data, using a main cost function;
• optimizing parameters which characterize the behavior of the neural network with an objective of improving the assessment by the main cost function during further processing of the learning input images.

A computer configured to train a neural network that maps input images onto an association with one or multiple classes of a predefined classification, onto a semantic segmentation, and/or onto a recognition of one or multiple objects, as output data, the computer configured to:

• provide learning input images and associated learning output data onto which the neural network ideally is to map the learning input images;
• provide auxiliary input images;
• generate modifications of the auxiliary input images by introducing at least one predefined change into each auxiliary input image of the auxiliary input images;
• supply the modifications to the neural network;
• ascertain predictions for the predefined change, using output data onto which the neural network maps each of the modifications;
• assess deviations of the predictions from the predefined change, using a predefined auxiliary cost function;
• optimize parameters that characterize a behavior of the neural network, with an objective of improving the assessment by the auxiliary cost function during further processing of the auxiliary input images;
• supply the learning input images to the neural network;
• assess deviations of output data, onto which the neural network maps the supplied learning input images, from the learning output data, using a main cost function;
• optimize parameters which characterize the behavior of the neural network with an objective of improving the assessment by the main cost function during further processing of the learning input images.