Domain adaptation for machine learning models

Patent metadata

Field	Value
Publication number	US-11978272-B2
Application number	US-202217883811-A
Country	US
Kind code	B2
Filing date	Aug 9, 2022
Priority date	Apr 24, 2020
Publication date	May 7, 2024
Grant date	May 7, 2024

Abstract

Adapting a machine learning model to process data that differs from training data used to configure the model for a specified objective is described. A domain adaptation system trains the model to process new domain data that differs from a training data domain by using the model to generate a feature representation for the new domain data, which describes different content types included in the new domain data. The domain adaptation system then generates a probability distribution for each discrete region of the new domain data, which describes a likelihood of the region including different content described by the feature representation. The probability distribution is compared to ground truth information for the new domain data to determine a loss function, which is used to refine model parameters. After determining that model outputs achieve a threshold similarity to the ground truth information, the model is output as a domain-agnostic model.

First claim

What is claimed is:

1. A method comprising:
obtaining a machine learning model trained to process input data of a first domain and generate an output that classifies the input data of the first domain;
identifying one or more portions of the machine learning model trained to extract features from the input data in the first domain;
training the machine learning model to classify input data of a second domain that is different from the first domain by:
causing the identified one or more portions of the machine learning model to generate, for an instance of input data of the second domain, a feature representation that classifies the instance of input data of the second domain; and
refining at least one parameter of the machine learning model using the feature representation; and

outputting the trained machine learning model to be used for classifying input data of an image defined by the second domain.

2. The method of claim 1 , wherein the machine learning model is trained to generate an output that classifies image data that depicts an object and the input data of the first domain depicts a first type of object and the input data of the second domain depicts a second type of object that is different than the first type of object.

3. The method of claim 1 , wherein the machine learning model is trained to generate an output that classifies speech utterances and the input data of the first domain includes speech utterances in a first language and the input data of the second domain includes speech utterances in a second language that is different than the first language.

4. The method of claim 1 , wherein the machine learning model is trained to generate a graph classification describing user behavior and the input data of the first domain includes information describing user behavior on a first internet domain and the input data of the second domain includes information describing user behavior on a second internet domain that is different from the first internet domain.

5. The method of claim 1 , wherein training the machine learning model to classify the input data of the second domain further comprises generating at least one probability distribution by processing the feature representation using a convolutional neural network, the method further comprising computing a loss function by comparing a ground truth classification or the instance of input data of the second domain to the at least one probability distribution.

6. The method of claim 5 , wherein the convolutional neural network is trained upon a segmentation objective and the at least one probability distribution includes a probability distribution for each of a plurality of discrete regions of the input data of the second domain.

7. The method of claim 1 , wherein the machine learning model is trained to generate the output that classifies the input data of the first domain using a local feature network that extracts local features of input data and a global feature network that classifies global features of input data.

8. The method of claim 7 , causing the machine learning model to generate the feature representation that classifies the instance of input data of the second domain comprises processing the instance of input data of the second domain using the local feature network.

9. The method of claim 7 , wherein the local feature network is configured as a faster region convolutional neural network configured for object detection.

10. The method of claim 7 , wherein the global feature network is configured to classify the global features of input data by processing the local features extracted by the local feature network.

11. The method of claim 7 , wherein the machine learning model is configured to generate the output that classifies the input data of the first domain by assigning at least one label to the input data of the first domain using the local features and the global features.

12. The method of claim 1 , further comprising computing a loss function by using the feature representation and a ground truth classification for the instance of input data of the second domain, wherein refining the at least one parameter of the machine learning model is performed using the loss function.

13. A system comprising:
a memory component; and
a processing device coupled to the memory component, the processing device to perform operations comprising:
obtaining a machine learning model trained to generate an output that classifies input data of a first domain using a local feature network that classifies local features of input data and a global feature network that classifies global features of the input data;
training the machine learning model to classify input data of a second domain by:
causing the machine learning model to generate a classification of second domain input data using the local feature network and independent of use of the global feature network; and
refining at least one parameter of the machine learning model based on a difference between the classification of the second domain input data and ground truth information for the second domain input data; and

outputting the trained machine learning model as a trained machine learning model configured to classify input data of the first domain and input data of the second domain.

14. The system of claim 13 , wherein training the machine learning model to classify the input data of the second domain further comprises generating at least one probability distribution by processing the classification of the second domain input data using a convolutional neural network, the operations further comprising determining the difference between the classification of the second domain input data and the ground truth information by comparing the ground truth information to the at least one probability distribution.

15. The system of claim 13 , wherein the local feature network is configured as a faster region convolutional neural network configured for object detection.

16. The system of claim 13 , wherein the global feature network is configured to classify the global features of input data by processing the local features extracted by the local feature network.

17. The system of claim 13 , wherein the machine learning model is configured to generate the output that classifies the input data of the first domain by assigning at least one label to the input data of the first domain using the local features and the global features.

18. The system of claim 13 , wherein the machine learning model is configured with one of a document classification objective, a speech recognition objective, an image classification objective, or a graph classification objective.

19. A computer-readable storage medium that is non-transitory and storing executable instructions, which when executed by a processing device, cause the processing device to perform operations comprising:
obtaining a machine learning model trained to process input data of a first domain and generate an output that classifies the input data of the first domain using a first feature network and a second feature network;
identifying one or more portions of the machine learning model trained to extract features from the input data in the first domain;
training the machine learning model to classify input data of a second domain that is different from the first domain by:
causing the identified one or more portions of the machine learning model to generate, for an instance of input data of the second domain, a feature representation that classifies the instance of input

Assignees

• Adobe Inc

Inventors

• Li Kai
• Tensmeyer Christopher Alan
• Wigington Curtis Michael
• Zhao Handong
• Barmpalios Nikolaos
• Sun Tong
• Manjunatha Varun
• Morariu Vlad Ion

Classifications

• G06N3/096

  Transfer learning · CPC title

• G06N3/0464

  Convolutional networks [CNN, ConvNet] · CPC title

• G06N3/09

  Supervised learning · CPC title

• G06V30/413 Primary

  Classification of content, e.g. text, photographs or tables · CPC title

• G06F17/18

  for evaluating statistical data {, e.g. average values, frequency distributions, probability functions, regression analysis (forecasting specially adapted for a specific administrative, business or logistic context G06Q10/04)} · CPC title