Testing a neural network

1. A computer-implemented method for testing an object detection output of a trained neural network having a plurality of layers, which detects or classifies objects, the method comprising:

reading at least one result from at least one first layer and a confidence value thereof, which is generated in the first layer of the trained neural network during an inference process;

checking a plausibility of the result by taking into consideration the confidence value thereof so as to conclude whether the object detection by the trained neural network during the inference process is correct or false;

wherein the first checking step includes comparing the confidence value for the result with a predefined threshold value; and

wherein in the event that it is concluded in the first checking step that the object detection output is false, utilization of the object falsely detected by the trained neural network during the inference process is prevented.

2. The method according to claim 1 , wherein the results of at least two different layers of the neural network are checked for plausibility so as to conclude whether the object detection by the neural network is correct or false.

3. The method according to claim 2 , wherein the results of at least two checks of the at least two layers of the neural network are weighted to varying degrees during the concluding whether the object detection by the neural network is correct or false.

4. The method according to claim 1 , wherein it is concluded whether the object detection by the neural network is correct or false if the confidence value for the result exceeds or falls below the predefined threshold value.

5. The method according to claim 1 , wherein a second checking step is performed if the confidence value for the result exceeds or falls below the predefined threshold value, and output of the object falsely detected by the neural network is only prevented if the second checking step confirms that the object detection is false.

6. The method according to claim 5 , wherein the second checking step is implemented by means of i) a support vector machine, ii) a structured random forest, iii) a one-class classifier or iv) a small neural network, which is independent of the neural network having a plurality of layers.

7. The method according to claim 1 , wherein only the result of the last layer of the neural network or the result of each layer of the neural network is checked for plausibility.

8. The method according to claim 1 , wherein a partial result within at least one layer of the neural network is checked for plausibility.

9. The method according to claim 1 , wherein the method is deployed in a driver assistance system in a vehicle, wherein the driver assistance system includes detecting or classifying objects.

10. The method according to claim 9 , wherein the input data for the neural network for detecting or classifying objects include image data of at least one vehicle camera.

11. A system-for testing the output of a trained neural network having a plurality of layers, which detects or classifies objects, comprising:

an interface for obtaining at least one result and the confidence value thereof, which is generated in at least one first layer of the trained neural network during an inference process;

a plausibility tester for checking the result by taking into consideration the confidence value thereof so as to conclude whether the object detection by the trained neural network during the inference process is correct or false,

an output interface for outputting objects detected or classified by the trained neural network,

wherein the plausibility tester compares the confidence value for the result with a predefined threshold value, and

wherein the plausibility tester prevents output of an object falsely detected by the trained neural network inasmuch as the plausibility tester has concluded that the object detection is false during the inference process.

12. The system according to claim 11 wherein the plausibility tester and/or the neural network has/have been trained with a predefined reference data set.

13. A vehicle having a control system for influencing the driving or a condition of the vehicle, wherein the control system comprises a system according to claim 11 , wherein the control system performs a control action based on the output of the neural network Hollowing a check by the plausibility tester.

14. A non-transitory computer-readable medium having a program stored thereon, that when executed by a processor, instructs the processor to perform the computer-implemented method according to claim 1 .