Domain adaptation of deep neural networks

Patent metadata

Field	Value
Publication number	US-11580405-B2
Application number	US-201916727429-A
Country	US
Kind code	B2
Filing date	Dec 26, 2019
Priority date	Dec 26, 2019
Publication date	Feb 14, 2023
Grant date	Feb 14, 2023

Abstract

Disclosed herein are system, method, and computer program product embodiments for adapting machine learning models for use in additional applications. For example, feature extraction models are readily available for use in applications such as image detection. These feature extraction models can be used to label inputs (such as images) in conjunction with other deep neural network models. However, in adapting the feature extraction models to these uses, it becomes problematic to improve the quality of their results on target data sets, as these feature extraction models are large and resistant to retraining. Approaches disclosed herein include a transfer layer for providing fast retraining of machine learning models.

First claim

What is claimed is:

1. A computer implemented method, comprising:
extracting, by one or more computing devices, features from an input by a feature extractor;
processing, by the one or more computing devices, the features by transfer layers configured as a deep neural network to produce a transfer output;
classifying, by the one or more computing devices, the input with a label by a label predictor implemented as a deep neural network, based on the transfer output;
classifying, by the one or more computing devices, the input as either a labeled source input or an unlabeled target input by a domain classifier; and
back propagating, by the one or more computing devices, a loss at the domain classifier to the domain classifier and the transfer layers.

2. The computer implemented method of claim 1 , wherein back propagating the loss at the domain classifier to the transfer layers is configured to minimize reliance on labeled source inputs by the transfer layers over successive inputs.

3. The computer implemented method of claim 1 , wherein back propagating the loss at the domain classifier further comprises:
back propagating, by the one or more computing devices, the loss at the domain classifier to the domain classifier by adjusting a weight of the domain classifier to reduce the loss; and
back propagating, by the one or more computing devices, the loss at the domain classifier to the transfer layers by adjusting a weight of the transfer layers to increase the loss.

4. The computer implemented method of claim 1 , further comprising:
back propagating, by the one or more computing devices, a loss at the label predictor to the transfer layers.

5. The computer implemented method of claim 1 , wherein the feature extractor is configured as a deep neural network.

6. The computer implemented method of claim 5 , wherein neurons of the feature extractor are implemented in a fixed configuration.

7. The computer implemented method of claim 5 , wherein the deep neural network of the transfer layers comprises fewer neurons than the deep neural network of the feature extractor.

8. A system, comprising:
a memory configured to store operations; and
one or more processors configured to perform the operations, the operations comprising:
extracting features from an input by a feature extractor,
processing the features by transfer layers configured as a deep neural network to produce a transfer output,
classifying the input with a label by a label predictor implemented as a deep neural network, based on the transfer output,
classifying the input as either a labeled source input or an unlabeled target input by a domain classifier, and
back propagating a loss at the domain classifier to the domain classifier and the transfer layers.

9. The system of claim 8 , wherein back propagating the loss at the domain classifier to the transfer layers is configured to minimize reliance on labeled source inputs by the transfer layers over successive inputs.

10. The system of claim 8 , wherein back propagating the loss at the domain classifier further comprises:
back propagating the loss at the domain classifier to the domain classifier by adjusting a weight of the domain classifier to reduce the loss; and
back propagating the loss at the domain classifier to the transfer layers by adjusting a weight of the transfer layers to increase the loss.

11. The system of claim 8 , the operations further comprising:
back propagating a loss at the label predictor to the transfer layers.

12. The system of claim 8 , wherein the feature extractor is configured as a deep neural network.

13. The system of claim 12 , wherein neurons of the feature extractor are implemented in a fixed configuration.

14. The system of claim 12 , wherein the deep neural network of the transfer layers comprises fewer neurons than the deep neural network of the feature extractor.

15. A computer readable storage device having instructions stored thereon, execution of which, by one or more processing devices, causes the one or more processing devices to perform operations comprising:
extracting features from an input by a feature extractor;
processing the features by transfer layers configured as a deep neural network to produce a transfer output;
classifying the input with a label by a label predictor implemented as a deep neural network, based on the transfer output;
classifying the input as either a labeled source input or an unlabeled target input by a domain classifier; and
back propagating a loss at the domain classifier to the domain classifier and the transfer layers.

16. The computer readable storage device of claim 15 , wherein back propagating the loss at the domain classifier to the transfer layers is configured to minimize reliance on labeled source inputs by the transfer layers over successive inputs.

17. The computer readable storage device of claim 15 , wherein back propagating the loss at the domain classifier further comprises:
back propagating the loss at the domain classifier to the domain classifier by adjusting a weight of the domain classifier to reduce the loss; and
back propagating the loss at the domain classifier to the transfer layers by adjusting a weight of the transfer layers to increase the loss.

18. The computer readable storage device of claim 15 , further comprising:
back propagating a loss at the label predictor to the transfer layers.

19. The computer readable storage device of claim 15 , wherein the feature extractor is configured as a deep neural network.

20. The computer readable storage device of claim 19 , wherein neurons of the feature extractor are implemented in a fixed configuration.

Assignees

• Sap Se

Inventors

• Santillán Perez Erick David
• Kernert David

Classifications

• G06N3/0464

  Convolutional networks [CNN, ConvNet] · CPC title

• G06N3/0895

  Weakly supervised learning, e.g. semi-supervised or self-supervised learning · CPC title

• G06N3/09

  Supervised learning · CPC title

• G06N3/094

  Adversarial learning · CPC title

• G06N3/096

  Transfer learning · CPC title