Neural network training

Patent metadata

Field	Value
Publication number	US-12020475-B2
Application number	US-202217676469-A
Country	US
Kind code	B2
Filing date	Feb 21, 2022
Priority date	Feb 21, 2022
Publication date	Jun 25, 2024
Grant date	Jun 25, 2024

Abstract

A deep neural network (DNN) can be trained based on a first training dataset that includes first images including annotated first objects. The DNN can be tested based on the first training dataset to determine first object predictions including first uncertainties. The DNN can be tested by inputting a second training dataset and outputting first object predictions including second uncertainties, wherein the second training dataset includes second images including unannotated second objects. A subset of images included in the second training dataset can be selected based on the second uncertainties, The second objects in the selected subset of images included in the second training dataset can be annotated. The DNN can be trained based on the selected subset of images included in the second training dataset including the annotated second objects.

First claim

The invention claimed is:

1. A computer, comprising:
a processor; and
a memory, the memory including instructions executable by the processor to:
train a deep neural network (DNN) based on receiving as input a first training dataset that includes first images including annotated first objects;
test the DNN based on the first training dataset to determine first object predictions including first uncertainties;
test the DNN by inputting a second training dataset and outputting first object predictions including second uncertainties, wherein the second training dataset includes second images including unannotated second objects;
select a subset of images included in the second training dataset based on the second uncertainties;
annotate the second objects in the selected subset of images included in the second training dataset; and
train the DNN based on the selected subset of images included in the second training dataset including the annotated second objects.

2. The computer of claim 1 , wherein a second computer includes instructions to operate a vehicle based on a third object prediction including a third uncertainty determined by the DNN.

3. The computer of claim 1 , wherein the annotated first objects and the annotated second objects include ground truth.

4. The computer of claim 3 , the instructions including further instructions to train the DNN by inputting images included in the first training dataset to the DNN a plurality of times to determine object predictions which are compared to the ground truth included in the first training dataset to determine a loss function.

5. The computer of claim 4 , the instructions including further instructions to backpropagate the loss function through layers of the DNN from the layers closest to the output to the layers closest to the input to select DNN processing weights.

6. The computer of claim 1 , wherein the DNN is a convolutional neural network that includes a plurality of convolutional layers and fully connected layers.

7. The computer of claim 1 , wherein the first uncertainties and the second uncertainties are probabilities that the object predictions are correct.

8. The computer of claim 1 , the instructions including further instructions to select the subset of images included in the second training dataset based on comparing the second uncertainties with first uncertainties.

9. The computer of claim 8 , the instructions including further instructions to compare the first uncertainties with the second uncertainties includes determining a mean and standard deviation for the first uncertainties based on Gaussian statistics.

10. The computer of claim 1 , wherein the first images and the second images include traffic scenes.

11. The computer of claim 1 , wherein the annotated first objects and the unannotated second objects include one or more of vehicles and pedestrians.

12. The computer of claim 1 , wherein training the DNN includes a plurality of datasets.

13. A method, comprising:
training a deep neural network (DNN) based on a first training dataset that includes first images including annotated first objects;
testing the DNN based on the first training dataset to determine first object predictions including first uncertainties;
testing the DNN by inputting a second training dataset and outputting first object predictions including second uncertainties, wherein the second training dataset includes second images including unannotated second objects;
selecting a subset of images included in the second training dataset based on the second uncertainties;
annotating the second objects in the selected subset of images included in the second training dataset; and
training the DNN based on the selected subset of images included in the second training dataset including the annotated second objects.

14. The method of claim 13 , wherein a second computer includes instructions to operate a vehicle based on a third object prediction including a third uncertainty determined by the DNN.

15. The method of claim 13 , wherein the annotated first objects and the annotated second objects include ground truth.

16. The method of claim 15 , further comprising training the DNN by inputting images included in the first training dataset to the DNN a plurality of times to determine object predictions which are compared to the ground truth included in the first training dataset to determine a loss function.

17. The method of claim 16 , further comprising backpropagating the loss function through layers of the DNN from the layers closest to the output to the layers closest to the input to select DNN processing weights.

18. The method of claim 17 , wherein the DNN is a convolutional neural network that includes a plurality of convolutional layers and fully connected layers.

19. The method of claim 13 , wherein the first uncertainties and the second uncertainties are probabilities that the object predictions are correct.

20. The method of claim 13 , further comprising selecting the subset of images included in the second training dataset based on comparing the second uncertainties with first uncertainties.

Assignees

• Ford Global Tech Llc
• Georgia Tech Res Inst

Inventors

• Parchami Mostafa
• Corona Enrique
• Alregib Ghassan
• Prabhushankar Mohit
• Benkert Ryan

Classifications

• G06V10/225

  based on a marking or identifier characterising the area · CPC title

• G06V10/765

  using rules for classification or partitioning the feature space · CPC title

• G06V10/774

  Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting · CPC title

• G06V10/82 Primary

  using neural networks · CPC title

• G06V10/454

  Integrating the filters into a hierarchical structure, e.g. convolutional neural networks [CNN] · CPC title