Method and apparatus with deep neural network model fusing

1. A processor-implemented method comprising:

implementing a deep neural network (DNN) model using input data;

generating, by implementing the DNN model, first output data from the DNN model;

changing the DNN model;

generating, by implementing the changed DNN model using the input data, second output data of the changed DNN model; and

determining result data by combining the first output data and the second output data.

2. The method of claim 1 , wherein the input data is inference input data.

3. The method of claim 2 , wherein the changing of the DNN model comprises changing at least one connection weight of the DNN model.

4. The method of claim 2 , wherein the changing of the DNN model comprises:

determining a noise value based on a noise model; and

applying the determined noise value to at least one connection weight of the DNN model.

5. The method of claim 4 , wherein the determining of the noise value comprises determining the noise value a Gaussian model.

6. The method of claim 4 , wherein the noise value is a quantized noise value.

7. The method of claim 2 , wherein the DNN model is a single DNN model.

8. The method of claim 2 , wherein the DNN model is one of a face authentication model, a liveness model, a translation model, and an objection recognition model, and an object detection model.

9. A non-transitory computer-readable storage medium storing instructions that, when executed by a processor, cause the processor to perform the method of claim 1 .

10. A processor-implemented data processing method comprising:

generating a first changed deep neural network (DNN) model by changing a trained initial DNN model;

generating first output data from the first changed DNN model by implementing the first changed DNN model using input data;

generating a second changed DNN model by changing the initial DNN model;

generating second output data from the second changed DNN model by implementing the second changes DNN model using the input data; and

determining result data by combining the first output data and the second output data.

11. The method of claim 10 , wherein

the generating of the first changed DNN model comprises changing at least one connection weight of the initial DNN model,

the generating of the second changed DNN model comprises changing at least one connection weight of the initial DNN model, and

the first changed DNN model is different from the second changed DNN model.

12. The method of claim 10 , further comprising determining the result data based on an output of the initial DNN model.

13. A processor-implemented training method comprising:

changing a deep neural network (DNN) model;

implementing the changed DNN model using trained input data;

generating, by implementing the changed DNN model, output data from the changed DNN model;

updating parameters of the DNN model based on the generated output data; and

training the DNN model by repeatedly performing the changing, the generating, and the updating.

14. The method of claim 13 , wherein the changing of the DNN model comprises changing at least one connection weight of the DNN model.

15. The method of claim 13 , wherein the changing of the DNN model comprises:

determining a noise value based on a noise model; and

applying the determined noise value to at least one connection weight of the DNN model.

16. The method of claim 15 , wherein the noise value is a quantized noise value.

17. The method of claim 13 , wherein the generating of the output data comprises generating the output data by inputting training data containing noise to the changed DNN model.

18. An apparatus comprising

one or more processors configured to:

implement a deep neural network (DNN) model using input data;

generate, by implementing the DNN model, first output data of the DNN model;

change the DNN model;

generate, by implementing the changed DNN model using the input data, second output data from the changed DNN model; and

determine result data by combining the first output data and the second output data.

19. The apparatus of claim 18 , wherein the processor is further configured to change at least one connection weight of the DNN model.

20. The apparatus of claim 19 , wherein the processor is further configured to determine a noise value based on a noise model and apply the determined noise value to the at least one connection weight of the DNN model, to change the at least one connection weight of the DNN model.

21. The apparatus of claim 18 , further comprising a memory storing instructions that, when executed by the one or more processors, configure the one or more processors to perform the generating of the first output data, the changing of the DNN model, the generating of the second output data, and the determining of the result data.

22. The apparatus of claim 18 , further comprising an image capture device configured to capture the input data.

23. An apparatus comprising

one or more processors configured to:

generate a first changed deep neural network (DNN) model by changing a trained initial DNN model;

generate first output data of the first changed DNN model by implementing the first changed DNN model using input data;

generate a second changed DNN model by changing the initial DNN model;

generate second output data from the second changed DNN model by implementing the second changes DNN model using the input data; and

determine result data by combining the first output data and the second output data.

24. The apparatus of claim 23 , wherein

the processor is further configured to:

generate the first changed DNN model by changing at least one connection weight of the DNN model; and

generate the second changed DNN model by changing at least one connection weight of the DNN model, and

wherein the first changed DNN model is different from the second changed DNN model.