Incremental learning through reverse data distribution

Patent metadata

Field	Value
Publication number	US-12039418-B2
Application number	US-202016939621-A
Country	US
Kind code	B2
Filing date	Jul 27, 2020
Priority date	Jul 27, 2020
Publication date	Jul 16, 2024
Grant date	Jul 16, 2024

Abstract

A method, system, and computer program product for reconstructing training data and building a new incremental learning model with the reconstructed training data that can be further trained. The method may include receiving new data to be inputted into a previously trained machine learning model, where the previously trained machine learning model has inaccessible training data. The method may also include generating simulated training data using a reverse form of the previously trained machine learning model. The method may also include verifying the simulated training data. The method may also include creating a new machine learning model using the simulated training data, where the new machine learning model includes a same structure as the previously trained machine learning model. The method may also include inputting the new data into the new machine learning model, where the new machine learning model is further trained with the new data.

First claim

What is claimed is:

1. A computer-implemented method comprising:
receiving new data to be inputted into a previously trained machine learning model, wherein training data for the previously trained machine learning model is inaccessible;
creating a reverse form of the previously trained machine learning model, wherein the creating comprises:
removing parameters of the previously trained machine learning model, and
transposing operators of the previously trained machine learning model;

generating simulated training data using the reverse form of the previously trained machine learning model and a plurality of probability vectors;
creating a new machine learning model using the simulated training data to train the new machine learning model; and
inputting the new data as training data into the new machine learning model, wherein the new machine learning model is further trained with the new data.

2. The method of claim 1 , wherein the generating the simulated training data comprises:
inputting the plurality of the probability vectors into the reverse form of the previously trained machine learning model.

3. The method of claim 2 , wherein an output from the inputting the plurality of the probability vectors into the reverse form of the previously trained machine learning model comprises a vector equal to a size of an embedded layer of the previously trained machine learning model.

4. The method of claim 2 , wherein the simulated training data comprises a plurality of simulated data, wherein each simulated data output corresponds to a probability vector input of the plurality of probability vectors.

5. The method of claim 1 , further comprising verifying the simulated training data, wherein verifying the simulated training data comprises:
determining that the simulated training data can train a model similar to the previously trained machine learning model such that a vector distribution of the simulated training data is consistent with a vector distribution of the inaccessible training data.

6. The method of claim 5 , wherein creating the new machine learning model comprises:
selecting the model similar to the previously trained machine learning model as the new machine learning model.

7. The method of claim 5 , wherein inputting the new data into the new machine learning model is in response to the verifying the simulated training data, and wherein inputting the new data into the new machine learning model comprises:
mixing the new data with the simulated training data to form new training data.

8. The method of claim 7 , wherein the creating the new machine learning model and the inputting the new data into the new machine learning model are executed in one batch.

9. The method of claim 5 , wherein verifying the simulated training data comprises:
constructing a forward machine learning model from the simulated training data;
comparing the forward machine learning model to the previously trained machine learning model; and
determining, based on the comparing, whether the forward machine learning model meets a similarity threshold to the previously trained machine learning model.

10. The method of claim 9 , further comprising:
determining that the forward machine learning model does not meet the similarity threshold;
analyzing the simulated training data for outlier simulated training data; and
removing the outlier simulated training data.

11. The method of claim 1 , wherein the new data is inputted into the new machine learning model for incremental learning.

12. A system having one or more computer processors, the system configured to:
receive new data to be inputted into a previously trained machine learning model, wherein the previously trained machine learning model has inaccessible training data;
create a reverse form of the previously trained machine learning model, wherein the creating comprises:
removing parameters of the previously trained machine learning model, and
transposing operators of the previously trained machine learning model;

generate simulated training data using tie reverse form of the previously trained machine learning model and a plurality of probability vectors;
verify the simulated training data;
create a new machine learning model using the simulated training data, wherein the new machine learning model comprises a same structure as the previously trained machine learning model; and
input the new data as training data into the new machine learning model, wherein the new machine learning model is further trained with the new data.

13. The system of claim 12 , wherein generating the simulated training data using the reverse form of the previously trained machine learning model comprises: generating the plurality of probability vectors for the previously trained machine learning model; and inputting the plurality of the probability vectors into the reverse form of the previously trained machine learning model.

14. The system of claim 12 , wherein verifying the simulated training data comprises:
determining that the simulated training data can train a model similar to the previously trained machine learning model such that a vector distribution of the simulated training data is consistent with a vector distribution of the inaccessible training data.

15. The system of claim 14 , wherein inputting the new data into the new machine learning model is in response to the verifying the simulated training data, and wherein inputting the new data into the new machine learning model comprises:
mixing the new data with the simulated training data to form new training data.

16. The system of claim 12 , wherein verifying the simulated training data comprises:
constructing a forward machine learning model from the simulated training data;
comparing the forward machine learning model to the previously trained machine learning model; and
determining, based on the comparing, whether the forward machine learning model meets a similarity threshold to the previously trained machine learning model.

17. A computer program product comprising a computer readable storage medium having program instructions embodied therewith, the program instructions executable by a server to cause the server to perform a method, the method comprising:
receiving new data to be inputted into a previously trained machine learning model, wherein the previously trained machine learning model has inaccessible training data;
creating a reverse form of the previously trained machine learning model, wherein the creating comprises:
removing parameters of the previously trained machine learning model, and
transposing operators of the previously trained machine learning model;

generating simulated training data using the reverse form of the previously trained machine learning model and a plurality of probability vectors;
verifying the simulated training data;
creating a new machine learning model using the simulated training data, wherein the new machine learning model comprises a same structure as the previously trained machine learning model; and
inputting the new data as training data into the new machine learning model, wherein the new machine learning model is further trained with the new data.

18. The computer program product of claim 17 , wherein generating the simulated training data using the reverse form of the previously trained machine learning model comprises: generating the plurality of probability vectors for the previously trained machine learning model; and inputting the plurality of the probability vectors in

Assignees

• IBM

Inventors

• Yuan Zhong Fang
• Liu Tong
• Zhang li ni
• Shang Bin
• Liang yong fang
• Gao Chen

Classifications

• G06N3/09

  Supervised learning · CPC title

• G06N3/088

  Non-supervised learning, e.g. competitive learning · CPC title

• G06N3/08

  Learning methods · CPC title

• G06N3/02

  Neural networks · CPC title

• G06N20/00 Primary

  Machine learning · CPC title