Training neural networks with a lesser requirement for labelled training data
A method for training a neural network for determining a task output with respect to a given task. The method includes: providing unlabeled and/or labelled encoder training records of measurement data; training the encoder network to map encoder training records to representations towards the goal that these representations, and/or or one or more work products derived from the representations, fulfil a self-consistency condition or correspond to ground truth; providing task training records that are labelled with ground truth; and training the association network and the task head networks towards the goal that, when a task training record is mapped to a representation using the encoder network, and the representation is mapped to a task output by the combination of the association network and the task head networks, the so-obtained task output corresponds to the ground truth with which the training record is labelled, as measured by a task loss function.
1 . A method for training a neural network for determining, from an input record of measurement data, a task output with respect to a given task, the neural network including:
an encoder network that is configured to map the input record to a representation, wherein the representation includes multiple independent components;
one or more task head networks that are configured to map representation components of the input record to the task output; and
an association network configured to provide, to each task head network of the one of more task head networks, a linear combination of those of the representation components of the input record x that are relevant for the task of the respective task head network,
the method comprising the following steps:
providing unlabeled and/or labelled encoder training records of measurement data;
training the encoder network to map encoder training records to representations towards a goal that the representations, and/or or one or more work products derived from the representations:
based on the representation being derived from an unlabeled encoder training record, fulfil a self-consistency condition that does not rely on ground truth, and
based on the representation being derived from a labelled encoder training record, correspond to the ground truth with which the encoder training record is labelled;
providing task training records that are labelled with ground truth; and
training the association network and the one or more task head networks towards a goal that, when the task training record is mapped to a representation using the encoder network, and the representation is mapped to a task output by a combination of the association network and the task head networks, a so-obtained task output corresponds to the ground truth with which the training record is labelled, as measured by a task loss function.
2 . The method of claim 1 , wherein the self-consistency condition includes that a reconstruction of the encoder training record obtained from the representation by a decoder network corresponds to an original encoder training record.
3 . The method of claim 1 , wherein the neural network is configured to determine a classification of the input record of measurement data as task output.
4 . The method of claim 1 , wherein the neural network is configured to process images as the input records of measurement data.
5 . The method of claim 3 , wherein:
the task output is a combination of one or more object values and one or more attribute values; and
the ground truth includes ground truth for the one or more object values and ground truth for the one or more attribute values;
an object task head network of the task head networks is configured to map the representation components of the input record to the one or more object values; and
an attribute task head network of the task head networks is configured to map the representation components of the input record to the one or more attribute values.
6 . The method of claim 5 , wherein an attribute value of the one or more attribute values is a color or a texture of an object.
7 . The method of claim 4 , wherein the task training records include images of road traffic situations.
8 . The method of claim 1 , further comprising:
providing, for each respective component of the representation, a factor head network that is configured to map the respective component to a predetermined basic factor of the input record;
providing factor training records that are labelled with ground truth values with respect to the basic factors represented by the components as encoder training records;
mapping, by the encoder network and the factor head networks, the factor training records to values of the basic factors;
rating deviations of the values of the basic factors from the ground truth values using an encoder loss function; and
optimizing parameters that characterize a behavior of the encoder network and parameters that characterize behavior of the factor head networks towards a goal that, when further factor training records are processed, the rating by the encoder loss function is likely to improve.
9 . The method of claim 7 , wherein the basic factors that correspond to the components of the representation include one or more of:
a time of day;
lighting conditions;
a season of the year; and
weather conditions;
in which the image is acquired.
10 . The method of claim 8 , wherein the neural network is configured to process images as the input records of measurement data, and wherein the basic factors that correspond to the components of the representation comprise one or more of:
a shape of at least one object in the image;
a color or at least one object in the image and/or area of the image;
a lighting condition in which the image was acquired;
a texture pattern of at least one object in the image.
11 . The method of claim 8 , wherein the providing of factor training records includes:
applying, to at least one given starting record, processing that impacts at least one of the basic factors thereby producing a factor training record; and
determining the ground truth values with respect to the basic factors based on the applied processing.
12 . The method of claim 8 , wherein, in each factor training record, each basic factor takes a particular value, and a set of the factor training records includes at least one factor training record for each combination of values of the basic factors.
13 . The method of claim 8 , wherein:
a combined loss function is formed as a weighted sum of the encoder loss function and the task loss function; and
parameters that characterize the behaviors of all networks are optimized with a goal of improving a value of the combined loss function.
14 . The method of claim 1 , wherein combinations of one encoder network on the one hand and multiple different combinations of an association network and task head networks on the other hand are trained based on one and the same training of the encoder network with encoder training records.
15 . The method of claim 1 , wherein the task loss function includes an entropy loss term, and the lower a number of representation components that the association network provides to each task head network, the better a value of the entropy loss term.
16 . The method of claim 1 , wherein the task loss function includes a suppression loss term, and the lower a number of task networks that each representation component is provided to be the association network, the better a value of the suppression loss term.
17 . A non-transitory storage medium on which is stored a computer program for training a neural network for determining, from an input record of measurement data, a task output with respect to a given task, the neural network including:
an encoder network that is configured to map the input record to a representation, wherein the representation includes multiple independent components;
one or more task head networks that are configured to map representation components of the input record to the task output; and
an association network configured to provide, to each task head network of the one of more task head networks, a linear combination of those of the representation components of the input record x that are relevant for the task of the respective task head network,
the computer program, when executed by one or more computers, causing the one or more computers to perform the following steps:
providing unlabeled and/or labelled encoder training records of measurement data;
training the encoder network to map encoder training records to representations towards a goal that the representations, and/or or one or more work products derived from the representations:
based on the representation being derived from an unlabeled encoder training record, fulfil a self-consistency condition that does not rely on ground truth, and
based on the representation being derived from a labelled encoder training record, correspond to the ground truth with which the encoder training record is labelled;
providing task training records that are labelled with ground truth; and
training the association network and the one or more task head networks towards a goal that, when the task training record is mapped to a representation using the encoder network, and the representation is mapped to a task output by a combination of the association network and the task head networks, a so-obtained task output corresponds to the ground truth with which the training record is labelled, as measured by a task loss function.
18 . One or more computers configured to train a neural network for determining, from an input record of measurement data, a task output with respect to a given task, the neural network including:
an encoder network that is configured to map the input record to a representation, wherein the representation includes multiple independent components;
one or more task head networks that are configured to map representation components of the input record to the task output; and
an association network configured to provide, to each task head network of the one of more task head networks, a linear combination of those of the representation components of the input record x that are relevant for the task of the respective task head network,
the one or more computers configured to:
provide unlabeled and/or labelled encoder training records of measurement data;
train the encoder network to map encoder training records to representations towards a goal that the representations, and/or or one or more work products derived from the representations:
based on the representation being derived from an unlabeled encoder training record, fulfil a self-consistency condition that does not rely on ground truth, and
based on the representation being derived from a labelled encoder training record, correspond to the ground truth with which the encoder training record is labelled;
provide task training records that are labelled with ground truth; and
train the association network and the one or more task head networks towards a goal that, when the task training record is mapped to a representation using the encoder network, and the representation is mapped to a task output by a combination of the association network and the task head networks, a so-obtained task output corresponds to the ground truth with which the training record is labelled, as measured by a task loss function.