Method for training a neural network

1 . A computer-implemented method for training a neural network, the neural network being configured to classify physical measuring variables, the method comprising the following steps:

training the neural network using a training data set, including:

drawing pairs, each including an input signal and an associated output signal, from the training data set for the training; and

adapting parameters of the neural network as a function of a respective output signal of the neural network, when the input signal of a drawn pair is supplied, and as a function of the associated output signal of the drawn pair;

wherein the drawing of the pairs always takes place from the entire training data set;

wherein the drawing of the pairs occurs regardless of which pair were previously drawn during the course of the training, wherein the input signal of the drawn pair is augmented using an augmentation function, wherein the augmentation function generates a variation of the input signal which leaves a classification of the input signal corresponding to an output signal of the neural network unchanged, wherein the augmentation function is selected from a set of provided augmentation functions which is dependent on the input signal of the drawn pair, wherein, during the drawing of the pairs from the training data set, a probability that a predefinable pair is drawn is dependent on a number of provided augmentation functions of the input signal of the predefinable pair, and wherein the variation is a rotation by a predefinable angle.

2 . The method as recited in claim 1 , wherein the adaptation of the parameters occurs as a function of an ascertained gradient and, for the ascertainment of the gradient, an estimated value of the gradient is refined, by taking a successively increasing number of pairs which are drawn from the training data set into consideration, until a predefinable termination condition which is dependent on the estimated value of the gradient is met.

3 . The method as recited in claim 2 , wherein the predefinable termination condition is also dependent on a covariance matrix of the estimated value of the gradient.

4 . The method as recited in claim 3 , wherein the predefinable termination condition encompasses a condition of whether the estimated value (m 1 ) and the covariance matrix (C) for a predefinable confidence value (λ) meet the condition (m 1 ,C −1 m 1 )≥λ 2 .

5 . A training system configured to train a neural network, the neural network being configured to classify physical measuring variables, the training system configured to:

train the neural network using a training data set, including:

draw pairs, each including an input signal and an associated output signal, from the training data set for the training; and

adapt parameters of the neural network as a function of a respective output signal of the neural network, when the input signal of a drawn pair is supplied, and as a function of the associated output signal of the drawn pair;

wherein the drawing of the pairs always takes place from the entire training data set;

wherein the drawing of the pairs occurs regardless of which pair were previously drawn during the course of the training, wherein the input signal of the drawn pair is augmented using an augmentation function, wherein the augmentation function generates a variation of the input signal which leaves a classification of the input signal corresponding to an output signal of the neural network unchanged, wherein the augmentation function is selected from a set of provided augmentation functions which is dependent on the input signal of the drawn pair, wherein, during the drawing of the pairs from the training data set, a probability that a predefinable pair is drawn is dependent on a number of provided augmentation functions of the input signal of the predefinable pair, and wherein the variation is a rotation by a predefinable angle.

6 . A method of using a neural network, the neural network being trained by drawing pairs, each including an input signal and an associated output signal, from the training data set for the training, and adapting parameters of the neural network as a function of a respective output signal of the neural network, when the input signal of a drawn pair is supplied, and as a function of the associated output signal of the drawn pair, wherein the drawing of the pairs always takes place from the entire training data set, and wherein the drawing of the pairs occurs regardless of which pair were previously drawn during the course of the training, the method comprising:

classifying first input signals which are present at an input of the neural network and were ascertained as a function of an output signal of a sensor, wherein the input signal of the drawn pair is augmented using an augmentation function, wherein the augmentation function generates a variation of the input signal which leaves a classification of the input signal corresponding to an output signal of the neural network unchanged, wherein the augmentation function is selected from a set of provided augmentation functions which is dependent on the input signal of the drawn pair, wherein, during the drawing of the pairs from the training data set, a probability that a predefinable pair is drawn is dependent on a number of provided augmentation functions of the input signal of the predefinable pair, and wherein the variation is a rotation by a predefinable angle; and

controlling an actuator based on the output of the neural network, the actuator actuating one of a robot or a motor vehicle.

7 . A method of using a neural network, the neural network being trained by drawing pairs, each including an input signal and an associated output signal, from the training data set for the training, and adapting parameters of the neural network as a function of a respective output signal of the neural network, when the input signal of a drawn pair is supplied, and as a function of the associated output signal of the drawn pair, wherein the drawing of the pairs always takes place from the entire training data set, and wherein the drawing of the pairs occurs regardless of which pair were previously drawn during the course of the training, the method comprising:

providing an activation signal for activating an actuator as a function of an first output signal of the neural network which is present at an output of the neural network, wherein the input signal of the drawn pair is augmented using an augmentation function, wherein the augmentation function generates a variation of the input signal which leaves a classification of the input signal corresponding to an output signal of the neural network unchanged, wherein the augmentation function is selected from a set of provided augmentation functions which is dependent on the input signal of the drawn pair, wherein, during the drawing of the pairs from the training data set, a probability that a predefinable pair is drawn is dependent on a number of provided augmentation functions of the input signal of the predefinable pair, wherein the variation is a rotation by a predefinable angle, and wherein in response to the activation signal the actuator actuates one of a robot or a motor vehicle.

8 . A non-transitory machine-readable memory medium on which is stored a computer program for training a neural network, the neural network being configured to classify physical measuring variables, the computer program, when executed by a computer, causing the computer to perform the following steps:

training the neural network using a training data set, including:

drawing pairs, each including an input signal and an associated output signal, from the training data set for the training; and

adapting parameters of the neural network as a function of a respective output signal of the neural network, when the input signal of a drawn pair is supplied, and as a function of the associated output signal of the drawn pair;

wherein the drawing of the pairs always takes place from the entire training data set;

wherein the drawing of the pairs occurs regardless of which pair were previously drawn during the course of the training, wherein the input signal of the drawn pair is augmented using an augmentation function, wherein the augmentation function generates a variation of the input signal which leaves a classification of the input signal corresponding to an output signal of the neural network unchanged, wherein the augmentation function is selected from a set of provided augmentation functions which is dependent on the input signal of the drawn pair, wherein, during the drawing of the pairs from the training data set, a probability that a predefinable pair is drawn is dependent on a number of provided augmentation functions of the input signal of the predefinable pair, and wherein the variation is a rotation by a predefinable angle.

9 . A method for using a neural network, the method comprising:

training the neural network using a training data set in a first phase, the neural network being trained by:

drawing pairs, each including an input signal and an associated output signal, from the training data set for the training, and

adapting parameters of the neural network as a function of a respective output signal of the neural network, when the input signal of a drawn pair is supplied, and as a function of the associated output signal of the drawn pair,

wherein the drawing of the pairs always takes place from the entire training data set,

wherein the drawing of the pairs occurs regardless of which pair were previously drawn during the course of the training; and

providing an activation signal for activating an actuator as a function of an first output signal of the neural network which is present at an output of the neural network, wherein the input signal of the drawn pair is augmented using an augmentation function, wherein the augmentation function generates a variation of the input signal which leaves a classification of the input signal corresponding to an output signal of the neural network unchanged, wherein the augmentation function is selected from a set of provided augmentation functions which is dependent on the input signal of the drawn pair, wherein, during the drawing of the pairs from the training data set, a probability that a predefinable pair is drawn is dependent on a number of provided augmentation functions of the input signal of the predefinable pair, and wherein the variation is a rotation by a predefinable angle, and wherein in response to the activation signal the actuator actuates one of a robot or a motor vehicle.