Further training of neural networks for the evaluation of measurement data
A method for further training of a neural network for processing measurement data, which neural network has been pre-trained with training examples from a set M. In the method: a batch B of new training examples is provided; a subset D⊆M of the previous training examples is provided; the new training examples from batch B and the previous training examples from subset D are processed by the neural network into outputs respectively; the deviations of the outputs from the respective target outputs are evaluated using a predefined cost function; parameters characterizing the behavior of the neural network are optimized with the aim that, during further processing of previous and new training examples, the evaluation with the cost function is improved in regard to new training examples from batch B and is not made worse in regard to previous training examples from subset D.
1 . A method for further training of a neural network for processing measurement data, the neural network having been pre-trained with training examples from a set M, each of the training examples being labeled with a target output of the neural network, the method comprising the following steps:
providing a batch B of new training examples, which are labeled with target outputs of the neural network;
providing a subset D⊆M of the previous training examples;
processing, by the neural network both the new training examples from the batch B and the previous training examples from the subset D, into outputs respectively;
evaluating deviations of the outputs from the respective target outputs using a predefined cost function; and
optimizing parameters characterizing the behavior of the neural network with a goal that, during further processing of previous and new training examples, the evaluation with the cost function: i) is improved in regard to new training examples from the batch B, and ii) is not made worse in regard to previous training examples from the subset D.
2 . The method as recited in claim 1 , wherein:
a first gradient G 0 of an expected value of the cost function for the previous training examples from the subset D is ascertained according to the parameters,
a second gradient G 1 of an expected value of the cost function for the new training examples from the batch B is ascertained according to the parameters, and
from the gradients G 0 and G 1 , a gradient g is ascertained along which the parameters are modified.
3 . The method as recited in claim 2 , wherein the gradient g is ascertained as a projection of the first gradient G 1 onto an orthogonal complement G 0 * of the first gradient G 0 .
4 . The method as recited in claim 1 , wherein, in response to a situation where a deterioration of the evaluation by the cost function in regard to previous training examples from the subset D exceeds a predefined threshold, the further training is terminated.
5 . The method as recited in claim 1 , wherein the cost function that was used for an original training of the neural network on the training examples from the set M is also used for the further training.
6 . The method as recited in claim 1 , wherein the neural network is a classifier configured to classify measurement data including images, into classes of a predefined classification.
7 . The method as recited in claim 6 , wherein the neural network includes:
a feature extractor, which converts an input of the neural network into a feature map, which may be represented as a vector with K elements, and
a classification head which, by applying a k×K matrix A, and a k-dimensional bias vector b, maps the feature map onto a vector having k classification scores relating to k available classes as output.
8 . The method as recited in claim 6 , wherein the neural network includes:
a feature extractor, which converts an input of the neural network into a feature map, which may be represented as a vector with K elements, and
a classification head which, by applying a k×K matrix A, maps the feature map onto a vector having k classification scores relating to k available classes as output.
9 . The method as recited in claim 7 , wherein, in response to a situation where at least one new training example in the batch B is labeled with a target output that does not correspond to any of the k available classes, wherein:
matrix A is extended with an additional k+1th row, which is filled with a mean of existing k rows, and
the bias vector b is extended with a k+1th component, which is filled with the mean of existing k components.
10 . The method as recited in claim 8 , wherein, in response to a situation where at least one new training example in the batch B is labeled with a target output that does not correspond to any of the k available classes, wherein:
matrix A is extended with an additional k+1th row, which is filled with a mean of existing k rows.
11 . The method as recited in claim 6 , wherein the subset D⊆M is chosen such that it contains substantially equal numbers of previous training examples for the classes of the predefined classification.
12 . The method as recited in claim 6 , wherein the classes represent road users, and/or roadways, and/or road boundaries, and/or traffic signs, and/or obstacles, and/or other objects of relevance when driving a vehicle.
13 . The method as recited in claim 1 , wherein the measurement data include images, and/or audio signals, and/or time series of measurements, and/or radar data and/or lidar data.
14 . The method as recited in claim 1 , further comprising:
supplying measurement data to the further-trained neural network in order for the further-trained neural network to generate outputs;
forming a control signal from the outputs of the neural network; and
using the control signal control: a vehicle, and/or a driver assistance system, and/or a system for quality control, and/or a system for monitoring areas, and/or a system for medical imaging.
15 . A non-transitory machine-readable data carrier on which is stored a computer program including machine-readable instructions for further training of a neural network for processing measurement data, the neural network having been pre-trained with training examples from a set M, each of the training examples being labeled with a target output of the neural network, the machine-readable instructions, when executed on one or more computers and/or compute instances, cause the computers and/or compute instances to perform the following steps:
providing a batch B of new training examples, which are labeled with target outputs of the neural network;
providing a subset D⊆M of the previous training examples;
processing, by the neural network both the new training examples from the batch B and the previous training examples from the subset D, into outputs respectively;
evaluating deviations of the outputs from the respective target outputs using a predefined cost function; and
optimizing parameters characterizing the behavior of the neural network with a goal that, during further processing of previous and new training examples, the evaluation with the cost function: i) is improved in regard to new training examples from the batch B, and ii) is not made worse in regard to previous training examples from the subset D.
16 . One or more computers and/or compute instances configured to further train a neural network for processing measurement data, the neural network having been pre-trained with training examples from a set M, each of the training examples being labeled with a target output of the neural network, the one or more computers and/or compute instances configured to:
provide a batch B of new training examples, which are labeled with target outputs of the neural network;
provide a subset D⊆M of the previous training examples;
process, by the neural network both the new training examples from the batch B and the previous training examples from the subset D, into outputs respectively;
evaluate deviations of the outputs from the respective target outputs using a predefined cost function; and
optimize parameters characterizing the behavior of the neural network with a goal that, during further processing of previous and new training examples, the evaluation with the cost function: i) is improved in regard to new training examples from the batch B, and ii) is not made worse in regard to previous training examples from the subset D.