Method and apparatus for automatically producing an artificial neural network

1. A computer-implemented method for automatically generating an artificial neural network that encompasses modules and connections that link those modules, successive modules and/or connections being added to a current starting network, the method comprising the following steps:

reading, by a computer processor, parameters representing the current starting network from a computer memory;

randomly selecting, using a random number generator, modules and/or connections that are to be added from a predefinable plurality of possible modules and connections that can be added;

respectively generating, by the computer processor, a plurality of possible refinements of a current starting network by adding to existing modules and connections of the current starting network the modules and/or connections that are to be added, wherein upon addition of the respective modules and/or connections that are to be added, predefinable parameters that characterize the respective modules and/or connections that are to be added are selected in such a way that the current starting network to which the modules and/or connections respectively to be added are added supplies, for each possible input variable, output variables unmodified respectively before and after addition of the respective modules and/or connections that are to be added;

selecting, by the computer processor, one of the refinements from the plurality of possible refinements to serve as a new starting network;

writing, by the computer processor, parameters representing the new starting network to the computer memory;

receiving, by an actuator control system including the artificial neural network, a sensor signal from a senor; and

determining, using the artificial neural network, a control signal for the actuator based on the sensor signal.

2. The computer-implemented method as recited in claim 1 , wherein the current starting network which serves as a starting point for a first-time addition of modules and/or connections is trained with an initial training data set prior to the first-time addition of modules and/or connections.

3. The computer-implemented method as recited in claim 1 , wherein the possible refinements of the current starting network are respectively trained with a training data set before one of the refinements from the plurality of possible refinements is then selected to serve as the current starting network in a next execution of the method.

4. The computer-implemented method as recited in claim 3 , wherein the refinements of the starting network are respectively trained using a gradient descent method, a parameter that characterizes a learning rate of the gradient descent method being selected to decrease with an increasing number of training epochs carried out.

5. The computer-implemented method as recited in claim 4 , wherein the decreasing profile of the parameter that characterizes the learning rate being characterized by way of a trigonometric function.

6. The computer-implemented method as recited in claim 1 , wherein the predefinable plurality of possible modules and/or connections include a normalization module and/or a skip connection and/or a module that contains a nonlinear function, the nonlinear function being idempotent for at least one value of a parameter characterizing its behavior.

7. The computer-implemented method as recited in claim 6 , wherein the predefinable plurality of possible modules and/or connections include the normalization module, and the skip connection, and the module containing the nonlinear module, and a module containing a nonlinear function, and a module that propagates an existing linear layer by adding units or channels, and a module that propagates a linear layer by duplicating that linear layer.

8. A non-transitory machine-readable memory medium on which is stored a computer program for automatically generating an artificial neural network that encompasses modules and connections that link those modules, successive modules and/or connections being added to a current starting network, the computer program, when executed by a computer, causing the computer to perform the following steps:

reading parameters representing the current starting network from a computer memory;

randomly selecting, using a random number generator, modules and/or connections that are to be added from a predefinable plurality of possible modules and connections that can be added;

respectively generating a plurality of possible refinements of a current starting network by adding to existing modules and connections of the current starting network the modules and/or connections that are to be added, wherein upon addition of the respective modules and/or connections that are to be added, predefinable parameters that characterize the respective modules and/or connections that are to be added are selected in such a way that the current starting network to which the modules and/or connections respectively to be added are added supplies, for each possible input variable, output variables unmodified respectively before and after addition of the respective modules and/or connections that are to be added;

selecting one of the refinements from the plurality of possible refinements to serve as the current starting network;

writing parameters representing the new starting network to the computer memory;

receiving, by an actuator control system including the artificial neural network, a sensor signal from a senor; and

determining, using the artificial neural network, a control signal for the actuator based on the sensor signal.

9. A computer system including a non-transitory machine-readable memory medium on which is stored a computer program for automatically generating an artificial neural network that encompasses modules and connections that link those modules, successive modules and/or connections being added to a current starting network, the computer program, when executed by the computer system, causing the computer system to perform the following steps:

reading parameters representing the current starting network from a computer memory;

randomly selecting, using a random number generator, modules and/or connections that are to be added from a predefinable plurality of possible modules and connections that can be added;

respectively generating a plurality of possible refinements of a current starting network by adding to existing modules and connections of the current starting network the modules and/or connections that are to be added, wherein upon addition of the respective modules and/or connections that are to be added, predefinable parameters that characterize the respective modules and/or connections that are to be added are selected in such a way that the current starting network to which the modules and/or connections respectively to be added are added supplies, for each possible input variable, output variables unmodified respectively before and after addition of the respective modules and/or connections that are to be added;

selecting one of the refinements from the plurality of possible refinements to serve as a new starting network;

writing parameters representing the new starting network to the computer memory;

receiving a sensor signal from a senor; and

determining, using the artificial neural network, a control signal for the actuator based on the sensor signal.

10. The computer-implemented method of claim 1 , wherein the random number generator is a pseudo random number generator.

11. The computer-implemented method of claim 1 , further comprising:

training, by the computer processor, the plurality of possible refinements of the current starting network with a training data set; and

after the training, determining, by the computer processor, a respective capability-characterizing characteristic value for each of the plurality of possible refinements;

wherein the selecting one of the refinements from the plurality of possible refinements to serve as the new starting network is based on the respective capability-characterizing characteristic value for each of the plurality of possible refinements.

12. The non-transitory machine-readable memory medium as recited in claim 8 , the steps further including:

training the plurality of possible refinements of the current starting network with a training data set; and

after the training, determining a respective capability-characterizing characteristic value for each of the plurality of possible refinements;

wherein the selecting one of the refinements from the plurality of possible refinements to serve as the new starting network is based on the respective capability-characterizing characteristic value for each of the plurality of possible refinements.

13. The computer system of claim 1 , the steps further including:

training the plurality of possible refinements of the current starting network with a training data set; and

after the training, determining a respective capability-characterizing characteristic value for each of the plurality of possible refinements;

wherein the selecting one of the refinements from the plurality of possible refinements to serve as the new starting network is based on the respective capability-characterizing characteristic value for each of the plurality of possible refinements.