DE202019103924U1 - Device for processing digital sensor data - Google Patents

Abstract

A device for processing digital sensor data, the device comprising a machine-readable storage medium having stored thereon instructions which, when executed by a computer, cause the computer to perform a method characterized by performing a plurality of training tasks from a distribution of training tasks (302), wherein the training tasks characterize the processing of digital sensor data, wherein a parameter set for an artificial neural network architecture and weights depends on at least a first training task from the distribution of the training tasks with a first gradient based learning algorithm and a second gradient-based learning algorithm is determined (304), wherein the artificial neural network is trained according to the parameter set and depending on a second training task with the first gradient-based learning algorithm ( 306), wherein digital sensor data is processed depending on the artificial neural network (200).

Description

State of the art

Artificial neural networks are suitable for processing digital sensor data. To train the artificial neural networks large amounts of this data and a high time and computational effort are required.

It is desirable to specify an improved procedure.

Disclosure of the invention

This is achieved by the subject matter of the independent claims.

A computer-implemented method for processing digital sensor data provides that a plurality of training tasks is provided from a distribution of training tasks, wherein the training tasks characterize the processing of digital sensor data, wherein a parameter set for an artificial neural network architecture and weights depends on at least one first training task is determined from the distribution of the training tasks with a first gradient-based learning algorithm and a second gradient-based learning algorithm, the artificial neural network depending on the parameter set and depending on a second training task with the first gradient-based learning Algorithm is trained, whereby digital sensor data is processed depending on the artificial neural network. The training tasks that characterize the digital sensor data may be prerecorded, simulated or calculated for offline training. Both the architecture and the weights of the artificial neural network can be trained with the at least one first training tasks in a first training phase for a specific application or independent of a specific application. For the specific application can thus be carried out in a second training phase training with only a second training task. This considerably reduces the training effort during an adaptation, in particular if the second training tasks correlate well with the first training tasks. For example, it is possible with little training effort to adapt the artificial neural network to a new sensor, which is used in a system for a previous sensor. This provides a machine learning model that has already been optimized for specific training tasks. In particular for deep neural networks, it is possible to easily adapt such an a priori optimized model for machine learning to a new training task. Fast means, for example, using very less newly identified training data, in a short time and / or with little computational effort compared to the training that was necessary for the a-priori optimization.

Preferably, the artificial neural network is defined by a plurality of layers, wherein elements of the plurality of layers share a common input and define a common output, wherein the architecture of the artificial neural network is defined by parameters in addition to the weights for neurons in the elements. wherein each of the parameters characterizes a contribution of one of the elements of the plurality of layers to the output. For example, the elements are arranged in parallel. For example, the parameters use their values to indicate the contribution of an element to which a parameter is assigned to the output. The outputs of individual elements are weighted by the values in addition to the weights provided by the artificial neural network for the neurons in the elements.

Preferably, the artificial neural network is trained in a first phase depending on a plurality of first training tasks with the first gradient-based learning algorithm and the second gradient-based learning algorithm, wherein the artificial neural network in a second phase depending on a second training task and dependent is trained by the first gradient-based learning algorithm and independent of the second gradient-based learning algorithm. The first phase takes place, for example, with first training tasks that originate from a generic application, in particular offline. The second phase takes place, for example, to adapt to a specific application with second training tasks, which originate from an operation of a specific application. The second training phase is executed, for example, during operation of the application.

Preferably, the artificial neural network is trained in a first phase depending on a plurality of first training tasks, wherein the artificial neural network is trained in a second phase depending on a fraction of the training data from the second training task. This will be a before Pre-trained artificial neural network in terms of architecture and weights adapted to a new application with little effort.

Preferably, the second gradient-based learning algorithm trains at least the parameters of the artificial neural network that define the architecture of the artificial neural network.

Furthermore, a device is proposed, which is set up to carry out one of the above-mentioned methods.

Preferably, a method for controlling a computer-controlled machine, in particular an at least semi-autonomous robot, a vehicle, a home application, a power tool, a personal support system, an access control system is provided, wherein training data for training tasks depending on digital sensor data generated, wherein an apparatus for machine learning, in particular for regression and / or for classification, and / or another application comprising an artificial neural network, is trained by means of training tasks according to the described method, wherein the computer-controlled machine is driven in response to an output signal of the thus-trained device. The training data are recorded for the specific application and used in particular in the second training phase for training. This facilitates adaptation of the artificial neural network and allows for immediate deployment.

The training data preferably comprise image data, video data and / or digital sensor data of a sensor, in particular from a camera, a thermal imaging camera, a LiDAR sensor, a radar sensor, an acoustic sensor, an ultrasonic sensor, a receiver for a satellite navigation system, a rotational speed sensor, a torque sensor, an acceleration sensor and / or a position sensor. These are particularly suitable for automation.

A computer-implemented method for training a machine learning, classification, or computer-controlled machine apparatus provides that a plurality of training tasks are provided from a distribution of training tasks, the training tasks characterizing the processing of digital sensor data using a parameter set for an architecture and for a computer Weights of an artificial neural network depending on at least a first training task from the distribution of training tasks with a first gradient-based learning algorithm and a second gradient-based learning algorithm is determined. Thus, this device is trained regardless of the specific application and before use then depending on the specific device and thus prepared for use in a specific application.

Preferably, it is provided that the artificial neural network is trained depending on the parameter set and dependent on a second training task with the first gradient-based learning algorithm. Thus, an adaptation to new training tasks can be efficiently implemented.

A device for processing digital sensor data, in particular for machine learning, classification or control of a computer-controlled machine, comprises a processor, and a memory for at least one artificial neural network, which are designed to execute the method. This device can be prepared independently of the specific application and then trained depending on the specific application.

Further advantageous embodiments will become apparent from the following description and the drawings.

• 1 eine schematische Darstellung von Teilen einer Vorrichtung für die Verarbeitung digitaler Sensordaten,
• 2 eine schematische Darstellung von Teilen eines künstlichen neuronalen Netzwerks,
• 3 Schritte in einem computerimplementierten Verfahren für die Verarbeitung digitaler Sensordaten,
• 4 Schritte in einem Verfahren zum Ansteuern einer computergesteuerten Maschine,
• 5 Schritte in einem computerimplementierten Verfahren zum Trainieren.

In the drawing shows

• 1 a schematic representation of parts of a device for processing digital sensor data,
• 2 a schematic representation of parts of an artificial neural network,
• 3 Steps in a computer-implemented method for the processing of digital sensor data,
• 4 Steps in a method for driving a computer-controlled machine,
• 5 Steps in a computer-implemented method for training.

In 1 is a device 100 for the processing of digital sensor data shown schematically. The device 100 includes a processor 102 , and a memory 104 , To capture digital sensor data in the example is a sensor 106 intended. The device 100 is in the example for controlling a computer-controlled machine 108 educated. The device 100 can also be designed for machine learning or a classification.

The sensor 106 is in the example via a signal line 110 with the processor 102 connectable. The processor 102 is formed in the example, digital signals of the sensor 106 to receive and into the store 104 save as training data. The training data include, for example, image data, video data, and / or other sensor digital sensor data 106 , The training data may be at least partially in an operation of the device 100 with the sensor 106 be recorded. Training data can also be independent of the sensor 106 detected or independent of the sensor 106 be provided digital signals.

The sensor 106 In particular, a camera, a thermal imaging camera, a LiDAR sensor, a radar sensor, an acoustic sensor, an ultrasonic sensor, a receiver for a satellite navigation system, a speed sensor, a torque sensor, an acceleration sensor and / or a position sensor. Several of these sensors can be provided.

The computer-controlled machine 108 is in the example via a signal line for an output signal 112 with the processor 102 connected. The processor 102 is in the example for driving the computer-controlled machine 108 formed depending on the digital signals.

The computer-controlled machine 108 In particular, it is an at least partially autonomous robot, a vehicle, a home application, a powered tool, a personal support system, or an access control system.

The memory 104 and the processor 102 are in the example with a signal line 114 connected. These components can be implemented distributed in a server infrastructure in particular. The device 100 may also be a controller that includes these components integrated into a microprocessor.

The device 100 is configured to perform the or one of the methods described below.

The device 100 includes at least one artificial neural network. An exemplary artificial neural network 200 is in 2 shown schematically.

The artificial neural network 200 is through a variety of layers 202 - 1 , ..., 202-m Are defined. In the example is an input 202 - 1 and an exit 202-m through one of the many layers 202 - 1 , ..., 202-m Are defined. The entrance 202 - 1 can be the input layer of the artificial neural network 200 or a hidden layer of the artificial neural network 200 his. The exit 202-m may be an output layer of the artificial neural network 200 or a hidden layer of the artificial neural network 200 his.

Certain elements 202-k , ..., 202-l the multitude of layers 202 - 1 , ..., 202-m have the entrance 202 - 1 as a common entrance. The Elements 202-k , ..., 202-l define the output in the example 202-m as a common output of the elements 202-k , ..., 202-l , That means the elements 202-k , ..., 202-l are in the artificial neural network 200 arranged in parallel with respect to their common input and with respect to their common output.

The artificial neural network 200 For example, it covers only a single hidden layer. This hidden layer contains several parallel elements. For example, a first element 202-k provided, which is designed as a 3x3 convolution. For example, an in 2 not shown second element provided, which is designed as a 5 × 5 convolution. For example, a third element 202-l provided, which is designed as MaxPooling. These three elements are arranged in parallel and form a search space consisting of the three elements {Conv3x3, Conv5x5, MaxPool}.

$$\text{output}=\text{Conv5x5}\left(\text{input}\right),$$

$$\text{output}=\text{MaxPool}\left(\text{input}\right).$$

A mathematical function describing for each of these three elements its output "output" dependent on a common input "input" is, for example, as follows: $$\text{output}=\text{Conv}3\text{x}3\left(\text{input}\right).$$

$$\text{output}=\text{Conv5x5}\left(\text{input}\right).$$

$$\text{output}=\text{maxpool}\left(\text{input}\right),$$

A mathematical function that describes a common output "output" of these three elements depending on the common input "input" is, for example, as follows: $$\text{output}={\alpha }_1*\text{Conv3x}3\left(\text{input}\right)+{\alpha }_2*\text{Conv}5\text{x}5\left(\text{input}\right)+{\alpha }_3*\text{maxpool}\left(\text{input}\right)$$

More general is the architecture of the artificial neural network 200 in addition to the weights w _a , ..., w _j for neurons 204-i , ..., 204-j in the elements 202-k , ..., 202-l through parameters α ₁ , ..., α _n Are defined. Each of the parameters α ₁ , ..., α _n characterizes a contribution of one of the elements 202-k , ..., 202 - 1 to the common exit. In the example, the parameters are for n = Ik elements α ₁ , ..., α _n Are defined. In the example, one of the parameters determines α ₁ , ..., α _n in a multiplication for all outputs of a single element its contribution to the output of the layer.

By appropriate determination of the parameters α ₁ , ..., α _n is possible that specifically one of the elements 202-k , ..., 202-l alone the result at the output of the layer determined. In the example, this would be just one of the parameters due to a non-zero value α ₁ , ..., α _n reachable. For example, of the three elements {Conv3x3, Conv5x5, MaxPool} described by way of example, α ₁ = 0, α ₂ = 1 and α ₃ = 0 mean that only the output of the Conv5x5 is considered, ie an architecture with the Conv5x5 layer. For the case α ₁ = 1, α ₂ = 0 and α ₃ = 0, this results in an architecture with the Conv3x3 layer. Generally, the parameter for each of the elements 202-k , ..., 202-l determined by a procedure described below by the artificial neural network 200 it is determined in which all elements 202-k , ..., 202-l are present parallel to each other. Every element 202-k , ..., 202-l is by a real-valued parameter α ₁ , ..., α _n weighted.

The parameters α ₁ , ..., α _n do not necessarily have to be 0 or 1, but can take on any real-valued numbers, for example α ₁ = 0.7, α ₂ = 0.2 and α ₃ = 0.1. This represents a relaxation of the search space. For example, for the parameters α ₁ , ..., α _n a boundary condition chosen so that a sum of the parameters α ₁ , ..., α _n gives the value one. This is for example by a determination of real valued values for the parameters α _1, ..., α _n and a normalization of the values for the parameters α ₁ , ..., α _n with the sum of all values possible. This relaxation represents a weighting of the individual elements 202-k , ..., 202-l in by all these elements 202-k , ..., 202-l defined architecture of the artificial neural network 200 represents.

With these, in particular real-valued parameters α ₁ , ..., α _n , a simple optimization of the architecture is possible. For example, optimization uses a gradient-based algorithm. Preferably, a stochastic gradient descent is used. Particularly preferred is the same kind of algorithms used for optimizing the weights w _a , ..., w _j for neurons 204-i , ..., 204-j in the elements 202-k , ..., 202-1 is used.

The artificial neural network 200 in 2 provides an example of such an arrangement of parallel elements 202-k , ..., 202-1 In general, an artificial neural network can have any number of such parallel elements, in particular in different, successive hidden layers. It can also be provided to arrange at least one of the elements parallel to another element or to a plurality of elements arranged in series.

Such, by the determination of the parameters α ₁ , ..., α _n Optimizable elements of the artificial neural network are parts that share a common input and define a common output. Several such layers may be provided which have respective inputs and outputs. In particular, each of the hidden layers may be so constructed. For each of these layers, a respective input and output can be provided.

A computer-implemented method for processing digital sensor data with such an artificial neural network will be described with reference to FIG 3 using the example of the artificial neural network 200 described.

In one step 302 will be a variety of p training tasks T ₁ . T ₂ , ..., T _p from a distribution p (T) of training tasks T provided.

In addition, the example for the three elements {Conv3x3, Conv5x5, MaxPool} becomes a meta-architecture α _meta provided. The meta-architecture α _meta is initially defined in this example as $$a_{meta}=\left(\mathrm{0.7,}\mathrm{0.2,}0.1\right)$$

These can be random, in particular real-valued quantities from zero to one. The example also includes meta weights w _meta initially defined.

The training tasks T characterize the processing of digital sensor data in the example. For example, these are data acquired by a sensor or determined or correlated with data acquired by a sensor. It may be image data, video data and / or digital sensor data of the sensor 106 underlie. The training tasks T For example, characterizing an association of the digital sensor data with a result of the processing. In particular, for digital sensor data from the at least one camera, the thermal imager, the LiDAR sensor, the radar sensor, the acoustic sensor, the ultrasonic sensor, the satellite navigation system receiver, the speed sensor, the torque sensor, the acceleration sensor and / or the position sensor Assignment to a classification of an event, in particular for at least partially autonomous control of the machine 108 be defined as a training task. For machine learning or regression corresponding training tasks can be defined.

In a subsequent step 304 becomes at least a first parameter set W ₁ . A ₁ for an artificial neural network architecture and weights depending on at least a first training task from the distribution of the training tasks T determined with a first gradient-based learning algorithm and a second gradient-based learning algorithm. The first parameter set W ₁ . A ₁ includes a first parameter value set A ₁ for the parameters α ₁ , ..., α _n and a first sentence W ₁ for the weights w _a , ..., w _j , The first sentence W ₁ For the weights also values for all other weights of all other neurons of the artificial neural network can be added 200 or part of the neurons of the artificial neural network 200 include. The final parameter value set resulting from the gradient descent procedures described below a _i defines the first parameter value set A ₁ , The last sentence resulting from the gradient descent procedures described below w _i with the weights defines the first sentence W ₁ for the weights.

Mit α_meta wird die Meta-Architektur bezeichnet. Mit w_meta werden die Meta-Gewichte bezeichnet.The first gradient-based learning algorithm involves for a specific training task T _i , a parameter value set α _i with the parameters α _{1, i} , ..., α _{n, i} , and a sentence w _i with the weights w _{a, i} , ..., _{wj, i} for example, an assignment $$\left(w_i.a_i\right)=\phi \left(w_{meta}.a_{meta}.T_i\right)$$

With α _meta is called the meta-architecture. With w _meta are called the meta weights.

Here, $ φ is an algorithm, in particular an optimization algorithm, training algorithm or learning algorithm, which optimizes both the weights and the architecture of a neural network for a specific training task for this training task. For example, with the execution of the algorithm φk steps gradient descent are performed to optimize the weights and the architecture. For the calculation, the algorithm Φ can be designed like the DARTS algorithm. DARTS refers to the algorithm "Differentiable Architecture Search", Hanxiao Liu, Karen Simonyan, Yiming Yang; ICLR; 2019; https://arxiv.org/abs/1806.09055.

The example will depend on this training task T _i an optimized architecture α _i depending on the initial meta-architecture α _meta and the initial weights w _meta determined as $$a_i=\left(0.80.00.2\right)=\left({\alpha }_1.{\alpha }_2.{\alpha }_3\right)$$

In addition, an optimized sentence w _i for the weights w _{a, i} , ..., _{wj, i} certainly.

The index i signals that α _i from the i-th training task T _i was determined. That means the parameters α _{1, i} , ..., α _{n, i} depend on the i-th training task T _i depending on the training task T _i vary.

In the example, the optimized architecture α _i depending on another training task T _i depending on the initial meta-architecture can also be determined as $$a_i=\left(0.0\text{1}\mathrm{.0,}0.0\right)=\left({\alpha }_1.{\alpha }_2.{\alpha }_3\right)$$

In addition, an optimized sentence w _i for the weights w _{a, i} , ..., _{wj, i} certainly.

Depending on the second gradient-based learning algorithm, at least one parameter is determined that defines the contribution of at least one of the elements to the output. In the example, the parameters become α ₁ , ..., α _n certainly.

The second gradient-based learning algorithm includes, for example, the plurality p training tasks T ₁ , ..., T _p an assignment $$\left(w_{meta}.a_{meta}\right)=\Psi \left(w_{meta}.w_1.....w_p.a_{meta}.a_1.....a_p.T_1.....T_p\right)$$

With Ψ is called a meta-learning algorithm. The meta-learning algorithm Ψ optimizes training tasks depending on a number of tasks T ₁ , ..., T _p with associated optimized architectures α ₁ ..., α _p and associated optimized weights w ₁ , ..., w _p the meta-architecture α _meta along with the meta weights w _meta , The optimized architectures are represented by the parameter value sets α ₁ , ..., α _p , The optimized weights are represented by the sentences w ₁ , ..., w _p for the weights.

The meta-learning algorithm Ψ is, for example, the MAML algorithm. The MAML refers to the Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks algorithm, Chelsea Finn, Pieter Abbeel, Sergey Levine; Proceedings of the 34 ^th International Conference on Machine Learning; 2017; https://arxiv.org/pdf/1703.03400.pdf. In contrast to meta-learning algorithms, which iteratively meta-learn the weights of a neural network, such as the original MAML algorithm, in which only the weights w of a fixed neural network are meta-learned, this also becomes the architecture of the neural network 200 meta-learned.

For a real-valued representation of the architecture of the artificial neural network, in the example with the MAML algorithm, gradients in the architecture space are also calculated for the architecture parameters. This gradient descent process optimizes both the weights and the architecture.

Anschließend wird in einem Schritt 306 geprüft, ob eine erste Phase abgeschlossen ist.For example, the gradient gradient method minimizes the following equation $$\underset{w_{meta}.a_{meta}}{min}{\displaystyle \underset{T}{\Sigma }LOss\left(T.\phi \right)}$$

Subsequently, in one step 306 Checked if a first phase has been completed.

The artificial neural network 200 In the example, in the first phase, it depends on the multitude of first training tasks T ₁ , ..., T _p trained with the first gradient-based learning algorithm and the second gradient-based learning algorithm.

The first parameter value set A ₁ for the parameters α ₁ , ..., α _n and the first sentence W ₁ for the weights w _a , ..., w _j defines in the example the artificial neural network 200 after training with the DARTS and the MAML algorithm.

The first phase is completed, for example, if a stop criterion applies. The stop criterion is, for example, the achievement of a time threshold or a resource budget. When the first phase is completed, one step becomes 308 executed. Otherwise, the step becomes 304 executed.

In step 308 becomes the artificial neural network 200 depending on the first parameter set W ₁ . A ₁ and trained according to a second training task with the first gradient-based learning algorithm. The final parameter value set resulting from training with the first gradient-based learning algorithm α _i defines a second parameter value set A ₂ , The final sentence resulting from training with the first gradient-based learning algorithm w _i with the weights defines the second sentence W ₂ for the weights.

That means the artificial neural network 200 is trained depending on a new training task and on the first gradient-based learning algorithm and independent of the second gradient-based learning algorithm. The second parameter value set A ₂ for the parameters α ₁ , ..., α _n and the second sentence W ₂ for the weights w _a , ..., w _j defines in the example the artificial neural network 200 after the final training only with the DARTS algorithm.

Subsequently, in one step 310 digital sensor data depending on the trained artificial neural network 200 processed.

Then the process ends.

In one aspect, the artificial neural network becomes 200 trained in the first phase depending on a plurality of first training tasks and in the second phase depending on a fraction of the training data, in particular from only a second training task.

Steps in a method for driving the computer-controlled machine 108 will be referred to below with reference 4 described.

The method of driving the computer controlled machine 108 starts, for example, when this is to be trained. In one aspect, the artificial neural network is 200 in the first phase as previously described, and in the device 100 for machine learning, for example for regression and / or for classification. The device 100 controls the computer-controlled machine 108 according to the method. The method starts, for example after switching on the computer-controlled machine 108 in the this artificial neural network 200 is implemented. It can also be an event, such as an exchange of the sensor 106 or a software update for the sensor 106 or for the computer-controlled machine 108 trigger the start.

After the start will be in one step 402 Training data for second training tasks depending on digital sensor data 110 generated. The training data may be image data, video data and / or sensor digital sensor data 106 his. For example, image data from the camera or the thermal imager is used. The image data can also come from the LiDAR sensor, the radar sensor, the acoustic sensor or the ultrasonic sensor. The training data may also include positions from the satellite navigation system receiver, rotational speed sensor readings, torque sensor torque, acceleration sensor acceleration and / or position sensor attitude information. The training data correlate in the example with the training data, in the first phase for the training of the artificial neural network 200 be used. The training tasks also correlate. For example, when replacing the sensor 106 or the first commissioning of the computer-controlled machine 108 with the sensor 106 first training tasks used in the first phase, in which generic sensor data used for the first phase by the actual sensor 106 certain sensor data will be replaced.

In a subsequent step 404 becomes the artificial neural network 200 trained by the second training tasks. In one aspect, the artificial neural network becomes 200 trained as described above for the second phase. This is how the device works 100 trained.

In a subsequent step 406 becomes the computer-controlled machine 108 depending on the output signal 112 the so-trained device 100 driven.

Subsequently, the method ends, for example when the computer-controlled machine 108 is switched off.

Steps in a computer-implemented method of training are described below 5 described.

After the start, a step 502 is executed.

In step 502 Training data are provided for the first training tasks according to the first phase. The training data is provided for example in a database.

In a subsequent step 504 The first training tasks for the first phase are determined. For example, for the distribution of training tasks p (T) determined for the first phase and the first training tasks from the distribution p (T) sampled. The second training tasks or the second training task need not be given or known at this time.

Subsequently, the artificial neural network 200 in one step 506 trained by means of the first training tasks according to the first phase.

• while (<some stopping criterion such as time or resource budget> ):
  • sample tasks T₁ , T₂ ,..., T_p from p(T)
  • for all T_i :
  • (w_i, α_i) = ϕ(W_meta, α_meta, T_i)
  • (w_meta, α_meta) = Ψ(w_meta, w_1, ..., w_p, α_meta, α₁, ..., α_p, T₁, ..., T_p)
  • return (w_meta, α_meta)

An exemplary implementation is given below for the distribution p (T) of the first training tasks:

• while (<some stopping criterion such as time or resource budget>):
  • sample tasks T ₁ . T ₂ , ..., T _p from p (T)
  • for all T _i :
  • (w _i , α _i ) = φ (W _meta , α _meta , T _i )
  • (w _meta , α _meta ) = Ψ (w _meta , w _1, ..., w _p , α _meta , α ₁ , ..., α _p , T ₁ , ..., T _p )
  • return (w _meta , α _meta )

Then the process ends.

Optionally, it can be provided that the artificial neural network 200 is trained by the second training tasks or only a second training task according to the second phase.

In one step 508 training data are provided for the second training tasks or only for the second training task according to the second phase.

Subsequently, in one step 510 determined at least a second training task for the second phase.

• (w_T, α_T) = ϕ(w_meta, α_meta, T)
• return (w_T, α_T)

Subsequently, the artificial neural network 200 trained according to the at least one second training tasks according to the second phase. An exemplary implementation of the step 512 is hereafter for a single second training task T reproduced:

• (w _T , α _T ) = φ (w _meta , α _meta , T)
• return (w _T , α _T )

The training tasks from the training assignment sentences are independently definable. A result of the training can be determined depending on the first phase of the method as well as on only one new training task. The step 510 can be applied to various new training tasks as needed, these are then independent of each other.

The described methods can be used to work with the artificial neural network 200 Predictions in particular depending on received sensor data. It can also be provided with the artificial neural network information about the sensor 106 extract from the received sensor data.

In the first phase, generic training data may be used for sensors of a particular sensor class, such as the sensor 106 counts. When replacing the sensor 106 Thus, the artificial neural network can be easily adapted by training in the second phase to a change of a hardware or software generation.

A concrete other application represents, for example, a traffic sign recognition. For example, country-specific traffic signs are used in the first phase, which are available only for a few countries, such as Germany or Austria. The first training data based on these country-specific traffic signs becomes the artificial neural network 200 trained in the first phase. If the traffic sign recognition is to be used in other countries, the artificial neural network becomes 200 trained in the second phase with a few second training data with traffic signs that are specific to these other countries.

Claims (9)

Apparatus for processing digital sensor data, the apparatus comprising a machine-readable storage medium having stored thereon instructions which cause, when executed by a computer, causes the computer to execute a method which is characterized in that a plurality of training tasks of a distribution of training tasks (302), wherein the training tasks characterize the processing of digital sensor data, wherein a parameter set for an artificial neural network architecture and weights depends on at least a first training task from the distribution of the training tasks with a first gradient based learning algorithm and a second gradient-based learning algorithm is determined (304), wherein the artificial neural network is trained according to the parameter set and depending on a second training task with the first gradient-based learning algorithm ( 306), wherein digital sensor data is processed depending on the artificial neural network (200). Device after Claim 1 characterized in that the stored instructions are arranged such that the method that the computer performs when executing these instructions on the computer is such that the artificial neural network (200) is defined by a plurality of layers (202-l, ..., 202-m), wherein elements (202-k, ..., 202-l) of the plurality of layers (202-l, ..., 202-m) have a common input (202-1 ) and define a common output (202-m), wherein the architecture of the artificial neural network (200) in addition to the weights (w _a , ..., w _j ) for neurons (204-i, ..., 204 -j) is defined in the elements (202-k, ..., 202-l) by parameters (α ₁ , ..., α _n ), each of the parameters (α ₁ , ..., an) Contribution of one of the elements (202-k, ..., 202-l) of the plurality of layers (202-1, ..., 202-m) to the output characterized. Device after Claim 1 or 2 characterized in that the stored instructions are arranged such that the method that the computer performs when executing these instructions on the computer is such that the artificial neural network (200) in a first phase depends on a plurality of first Training exercises with the first gradient-based learning algorithm and the second gradient-based learning algorithm is trained, wherein the artificial neural network (200) in a second phase depending on a second training task and depending on the first gradient-based learning algorithm and is trained independently of the second gradient-based learning algorithm. Device after Claim 3 characterized in that the stored instructions are arranged such that the method that the computer performs when executing these instructions on the computer is such that the artificial neural network (200) in the first phase depends on a plurality of first Training exercises is trained, wherein the artificial neural network (200) is trained in the second phase depending on a fraction of the training data from the second training task. Device after Claim 4 characterized in that the stored instructions are arranged such that the method that the computer carries out when these instructions are executed on the computer is such that at least the parameters (α ₁ , ..., α _n ) of the artificial neural network (200) that define the architecture of the artificial neural network (200). Device for driving a computer-controlled machine (108), in particular an at least semi-autonomous robot, a vehicle, a home appliance, a power tool, a personal access system, an access control system, characterized in that training data for training tasks depending on digital sensor data (110) are generated (402), wherein a device (100) comprising an artificial neural network (200), by means of training tasks according to the method of the Claims 1 to 5 is trained (404), wherein the computer-controlled machine (108) is driven (406) depending on an output signal (112) of the thus trained device (100). Device after Claim 6 , Characterized in that the training data is image data, video data and / or digital sensor data of a sensor (106), in particular of at least one camera, a thermal imaging camera, a LiDAR sensor, a radar sensor an acoustic sensor, an ultrasonic sensor, a receiver for a satellite navigation system, a speed sensor, a torque sensor, an acceleration sensor and / or a position sensor. Apparatus for training an apparatus (100) for machine learning, classification or control of a computer controlled machine (108), characterized in that a plurality of training tasks is provided from a distribution of training tasks (502, 504), wherein the training tasks are the processing of digital sensor data wherein a parameter set for an artificial neural network architecture and weights (200) is determined from the distribution of training tasks with a first gradient-based learning algorithm and a second gradient-based learning algorithm depending on at least one first training task ( 506). Device after Claim 8 , characterized in that the artificial neural network is trained depending on the parameter set and depending on a second training task with the first gradient-based learning algorithm (512).

Images