System improvement for deep neural networks
The technology relates to tuning a data translation block (DTB) including a generator model and a discriminator model. One or more processors may be configured to receive training data including an image in a second domain. The image in the second domain may be transformed into a first domain with a generator model. The transformed image may be processed to determine one or more outputs with one or more deep neural networks (DNNs) trained to process data in the first domain. An original objective function for the DTB may be updated based on the one or more outputs. The generator and discriminator models may be trained to satisfy the updated objective function.
1. A method for tuning a data translation block (DTB) including a generator model and a discriminator model, the method including:
receiving, by one or more processors, training data including an image in a second domain;
transforming, by a generator model executing on the one or more processors, the image from the second domain to a transformed image in a first domain;
processing, by one or more processors executing one or more deep neural networks (DNNs) trained to process data in the first domain, the transformed image to determine one or more outputs;
updating, by the one or more processors, an original objective function for the DTB based on the one or more outputs; and
training, by the one or more processors, the generator model and discriminator model to satisfy the updated objective function.
2. The method of claim 1 , wherein each of the one or more DNNs outputs a respective output of the one or more outputs, and
each respective output includes one or more predictions or estimated values.
3. The method of claim 2 , wherein the training data further includes markings associated with the image in the second domain, the markings associated with the image in the second domain being separated into subsets, with each subset of markings being assigned to the one or more DNNs.
4. The method of claim 3 , further including:
calculating a loss value for each of the one or more outputs.
5. The method of claim 4 , wherein the loss value (Loss i ) is calculated for each DNN, using the formula:
Loss i =(M D2-i −M ˜D2-i ) 2 , where i is the DNN index, where M D2-i is the subset of markings assigned to DNNi and M ˜D2-i is the one or more predictions or estimated values output by DNNi.
6. The method of claim 5 , further including:
weighting the loss value for each of the one or more DNNs;
aggregating the weighted loss value for each of the one or more DNNs; and
determining an average loss value (DNN_AVG_LOSS).
7. The method of claim 6 , wherein the updated objective function is:
O CGAN +α DNN_AVG_LOSS *DNN_AVG_LOSS, where O CGAN is the original objective function of the DTB and α DNN_AVG_LOSS is a weight factor.
8. The method of claim 1 , wherein the training data further includes a training image in the first domain, and
the discriminator is trained based on the transformed image and the training image.
9. A non-transitory computer-readable medium storing instructions, which when executed by one or more processors, cause the one or more processors to:
receive training data including an image in a second domain;
transform the image from the second domain to a transformed image in a first domain with a generator model;
process the transformed image to determine one or more outputs with one or more deep neural networks (DNNs) trained to process data in the first domain;
update an original objective function for a DTB based on the one or more outputs; and
train the generator model and discriminator model to satisfy the updated objective function.
10. The non-transitory computer-readable medium of claim 9 , wherein each of the one or more DNNs outputs a respective output of the one or more outputs, and
each respective output includes one or more predictions or estimated values.
11. The non-transitory computer-readable medium of claim 9 , wherein the training data further includes a training image in the first domain, and
the discriminator is trained based on the transformed image and the training image.
12. The non-transitory computer-readable medium of claim 10 , wherein the training data further includes markings associated with the image in the second domain, the markings associated with the image in the second domain being separated into subsets, with each subset of markings being assigned to the one or more DNNs.
13. The non-transitory computer-readable medium of claim 12 , wherein the instructions further cause the one or more processors to calculate a loss value for each of the one or more outputs.
14. The non-transitory computer-readable medium of claim 13 , wherein the loss value (Loss i ) is calculated for each DNN, using the formula:
Loss i =(M D2-i −M ˜D2-i ) 2 , where i is the DNN index, where M D2-i is the subset of markings assigned to DNNi and M ˜D2-i is the one or more predictions or estimated values output by DNNi.
15. The non-transitory computer-readable medium of claim 14 , wherein the instructions further cause the one or more processors to:
weight the loss value for each of the one or more DNNs;
aggregate the weighted loss value for each of the one or more DNNs; and
determine an average loss value (DNN_AVG_LOSS).
16. The non-transitory computer-readable medium of claim 15 , wherein the updated objective function is:
O CGAN +α DNN_AVG_LOSS *DNN_AVG_LOSS, where O CGAN is the original objective function of the DTB and α DNN_AVG_LOSS is a weight factor.
17. A system for tuning a data translation block (DTB) including a generator model and a discriminator model, the system including:
one or more computing devices having one or more processors; and
memory storing instructions, the instructions, when executed by the one or more processors, cause the one or more processors to:
receive training data including an image in a second domain;
transform the image from the second domain to a transformed image in a first domain with a generator model;
process the transformed image to determine one or more outputs with one or more deep neural networks (DNNs) trained to process data in the first domain;
update an original objective function for the DTB based on the one or more outputs; and
train the generator model and discriminator model to satisfy the updated objective function.
18. The system of claim 17 , wherein each of the one or more DNNs outputs a respective output of the one or more outputs, and
each respective output includes one or more predictions or estimated values.
19. The system of claim 17 , wherein the training data further includes a training image in the first domain, and
the discriminator is trained based on the transformed image and the training image.
20. The system of claim 18 , wherein the training data further includes markings associated with the image in the second domain, the markings associated with the image in the second domain being separated into subsets, with each subset of markings being assigned to the one or more DNNs.
21. The system of claim 20 , wherein the instructions further cause the one or more processors to calculate a loss value for each of the one or more outputs.
22. The system of claim 21 , wherein the loss value (Loss i ) is calculated for each DNN, using the formula:
Loss i =(M D2-i −M ˜D2-i ) 2 , where i is the DNN index, where M D2-i is the subset of markings assigned to DNNi and M ˜D2-i is the one or more predictions or estimated values output by DNNi.
23. The system of claim 22 , wherein the instructions further cause the one or more processors to:
weight the loss value for each of the one or more DNNs;
aggregate the weighted loss value for each of the one or more DNNs; and
determine an average loss value (DNN_AVG_LOSS).
24. The system of claim 23 , wherein the updated objective function is:
O CGAN +α DNN_AVG_LOSS *DNN_AVG_LOSS, where O CGAN is the original objective function of the DTB and α DNN_AVG_LOSS is a weight factor.