Separating public and private knowledge in AI

Patent metadata

Field	Value
Publication number	US-11861035-B2
Application number	US-201916414087-A
Country	US
Kind code	B2
Filing date	May 16, 2019
Priority date	May 16, 2019
Publication date	Jan 2, 2024
Grant date	Jan 2, 2024

Abstract

A computer-implemented method comprises linking a private AI model to a public AI model to thereby form a combined AI model comprising the private AI model and the public AI model; and training the combined AI model with private samples while keeping the public AI model fixed so that only the private AI model is trained with the private samples.

First claim

What is claimed is:

1. A computer-implemented method comprising:
linking, by an AI model training application, a private AI model, to a public pre-trained AI model to thereby form a combined fusion AI model comprising the private AI model and the public pre-trained AI model; and
training, by the AI model training application, the combined fusion AI model with private samples while keeping the public pre-trained AI model fixed so that only the private AI model is trained with the private samples and feature tensor input from the public pre-trained AI model, wherein weights associated with the public AI model are not updated, allowing simultaneous training of a plurality of private AI models.

2. The computer-implemented method of claim 1 , wherein linking the private AI model to the public pre-trained AI model comprises using one or more features of the public pre-trained AI model with the private AI model.

3. The computer-implemented method of claim 1 , wherein the public pre-trained AI model is located on a public cloud.

4. The computer-implemented method of claim 1 , wherein the public pre-trained AI model is trained with public data comprising first image data and the private AI model is trained with private data comprising second image data.

5. The computer-implemented method of claim 1 , wherein the public pre-trained AI model is trained with public data comprising first speech data and the private AI model is trained with private data comprising second speech data.

6. The computer-implemented method of claim 1 , wherein the public pre-trained AI model is trained with public data comprising first computer vision data and the private AI model is trained with private data comprising second computer vision data.

7. The computer-implemented method of claim 1 , wherein the public pre-trained AI model is trained with public data comprising first language data and the private AI model is trained with private data comprising second language data.

8. The computer-implemented method of claim 1 , wherein the public pre-trained AI model is trained with public data comprising first text data and the private AI model is trained with private data comprising second text data.

9. The computer-implemented method of claim 1 , wherein the public pre-trained AI model is trained with public data comprising first audio data and the private AI model is trained with private data comprising second audio data.

10. The computer-implemented method of claim 1 , wherein private AI model is smaller than the public pre-trained AI model, based on number of parameters.

11. The computer-implemented method of claim 1 , wherein the computer implemented method further comprises:
linking, by an AI model training application, a private AI model to a first public pre-trained AI model to thereby form a first combined AI model comprising the private AI model and the first public pre-trained AI model;
training, by the AI model training application, the first combined AI model with first private samples while keeping the first public pre-trained AI model fixed so that only the private AI model is trained with the first private samples;
linking, by the AI model training application, the private AI model to a second public pre-trained AI model to thereby form a second combined AI model comprising the private AI model and the second public pre-trained AI model; and
training, by the AI model training application, the second combined AI model with second private samples while keeping the second public pre-trained AI model fixed so that only the private AI model is trained with the second private samples.

12. A computer program product comprising a computer readable storage medium having program instructions embodied therewith, the program instructions executable by one or more processors to cause the one or more processors to perform a method comprising:
linking, by an AI model training application, a private AI model, to a public pre-trained AI model to thereby form a combined fusion AI model comprising the private AI model and the public pre-trained AI model; and
training, by the AI model training application, the combined AI model with private samples while keeping the public pre-trained AI model fixed so that only the private AI model is trained with the private samples and feature tensor input from the public pre-trained AI model, wherein weights associated with the public AI model are not updated, allowing simultaneous training of a plurality of private AI models.

13. The computer program product of claim 12 , wherein linking the private AI model to the public pre-trained AI model comprises using one or more features of the public AI model with the private AI model.

14. The computer program product of claim 12 , wherein the public pre-trained AI model is located on a public cloud.

15. The computer program product of claim 12 , wherein the public pre-trained AI model is trained with public data comprising first image data and the private AI model is trained with private data comprising second image data.

16. The computer program product of claim 12 , wherein the public pre-trained AI model is trained with public data comprising first speech data and the private AI model is trained with private data comprising second speech data.

17. A system including one or more processors configured to implement a method comprising:
linking, by an AI model training application, a private AI model, to a public pre-trained AI model to thereby form a combined fusion AI model comprising the private AI model and the public pre-trained AI model; and
training, by the AI model training application, the combined AI model with private samples while keeping the public pre-trained AI model fixed so that only the private AI model is trained with the private samples and feature tensor input from the public pre-trained AI model, wherein weights associated with the public AI model are not updated, allowing simultaneous training of a plurality of private AI models.

18. The system of claim 17 , wherein linking the private AI model to the public pre-trained AI model comprises using one or more features of the public AI model with the private AI model.

19. The system of claim 17 , wherein the public pre-trained AI model is located on a public cloud.

20. The system of claim 17 , wherein the public pre-trained AI model is trained with public data comprising first image data and the private AI model is trained with private data comprising second image data.

Assignees

• IBM

Inventors

• Cho Minsik
• Kang Bumsoo
• Yoo Chungkuk

Classifications

• G06N3/09

  Supervised learning · CPC title

• G06N3/0464

  Convolutional networks [CNN, ConvNet] · CPC title

• G06N3/096

  Transfer learning · CPC title

• G06F21/6245 Primary

  Protecting personal data, e.g. for financial or medical purposes · CPC title

• G06N3/045

  Combinations of networks · CPC title