Model learning device, model learning method, and program

Patent metadata

Field	Value
Publication number	US-11797845-B2
Application number	US-201917058099-A
Country	US
Kind code	B2
Filing date	May 28, 2019
Priority date	May 28, 2018
Publication date	Oct 24, 2023
Grant date	Oct 24, 2023

Abstract

Simultaneous learning of a plurality of different tasks and domains, with low costs and high precision, is enabled. A learning unit 160 , on the basis of learning data, uses a target encoder that takes data of a target domain as input and outputs a target feature expression, a source encoder that takes data of a source domain as input and outputs a source feature expression, a common encoder that takes data of the target domain or the source domain as input and outputs a common feature expression, a target decoder that takes output of the target encoder and the common encoder as input and outputs a result of executing a task with regard to data of the target domain, and a source decoder that takes output of the source encoder and the common encoder as input and outputs a result of executing a task with regard to data of the source domain, to learn so that the output of the target decoder matches training data, and the output of the source decoder matches training data.

First claim

The invention claimed is:

1. A computer-implemented method of training models for performing a task, the method comprising:
receiving learning data, wherein the learning data includes:
a first pair of data including:
target domain data from a target domain, and
first result data as first training data of performing a first task upon the target domain data, and

a second pair of data including:
source domain data from a source domain, and
second result as second training of performing a second task upon the source domain data;

generating, by a target encoder based on the learning data, a target feature expression using the target domain data;
generating, by a source encoder based on the learning data, a source feature expression using the source domain data, wherein the source encoder is distinct from the target encoder;
generating, by a common encoder based at least on one of the target domain data or the source domain data, a common feature expression using common parameters for encoding the at least one of the target domain data or the source domain data, wherein the common encoder is distinct from the source encoder and the target encoder;
generating, by a target decoder based at least on a first feature expression set including the generated target feature expression and the generated common feature expression, the first result data of performing the first task upon the target domain data;
generating, by a source decoder based at least on a second feature expression set including the generated source feature expression and the generated common feature expression, the second result data of performing the second task upon the source domain data; and
training a combination of the target encoder, the source encoder, the common encoder, the target decoder, and the source decoder,
wherein the first result data of performing the first task upon the target domain data is associated with the learning data, and
wherein the second result data of performing the second task upon the source domain data is associated with the learning data.

2. The computer-implemented method of claim 1 , the method further comprising:
training the combination of the target encoder, the source encoder, the common encoder, the target decoder, and the source decoder using at least a loss function indicating:
the generated first result data of performing the first task upon the target domain data matching the learning data,
the generated second result data of performing the second task upon the source domain data matching the learning data,
a first common feature expression based on the target domain data matching in expressions with a second common feature expression based on the source domain data,
the generated first result data of performing the first task upon the target domain data is distinct in expressions from the first common feature expression, and
the generated second result data of performing the second task upon the source domain data is distinct in expression from the second common feature expression.

3. The computer-implemented method of claim 1 , the method further comprising:
receiving, by the target decoder, either one of:
a first addition of the generated target feature expression and the generated common feature expression, or
a first combination of the generated target feature expression and the generated common feature expression; and

receiving, by the source decoder, either one of:
a second addition of the generated source feature expression and the generated common feature expression, or
a second combination of the generated source feature expression and the generated common feature expression.

4. The computer-implemented method of claim 1 , wherein one or more of the target encoder, the source encoder, the common encoder, the target decoder, and the source decoder is a neural network including a plurality of layers.

5. The computer-implemented method of claim 1 , the method further comprising:
receiving, by the target decoder, the generated target feature expression and the generated common feature expression;
generating, by the target decoder, the first result data of performing the first task upon the target domain data using a first intermediate expression from a first intermediate layer of the target encoder and a second intermediate expression from a second intermediate layer of the common encoder;
receiving, by the source decoder, the generated source feature expression and the generated common feature expression; and
generating, by the source decoder, the second result data of performing the second task upon the source domain data using a third intermediate expression from a third intermediate layer of the source encoder and a fourth intermediate expression from a fourth intermediate layer of the common encoder.

6. The computer-implemented method of claim 1 , wherein the target domain relates to a first view point of an object image for image recognition, wherein the source domain relates to a second view point of the object image for image recognition, and wherein the first view point and the second view point are distinct.

7. The computer-implemented method of claim 1 , wherein the first task is associated with one of detection or segmentation of an image recognition task, and wherein the second task is associated with the other of the detection or the segmentation of the image recognition task.

8. A system of training models for performing a task, the system comprises:
a processor; and
a memory storing computer-executable instructions that when executed by the processor cause the system to:
receive learning data, wherein the learning data includes:
a first pair of data including:
target domain data from a target domain, and
first result data as first training data of performing a first task upon the target domain data, and

a second pair of data including:
source domain data from a source domain, and
second result as second training of performing a second task upon the source domain data;

generate, by a target encoder based on the learning data, a target feature expression using the target domain data;
generate, by a source encoder based on the learning data, a source feature expression using the source domain data, wherein the source encoder is distinct from the target encoder;
generate, by a common encoder based at least on one of the target domain data or the source domain data, a common feature expression using common parameters for encoding the at least one of the target domain data or the source domain data, wherein the common encoder is distinct from the source encoder and the target encoder;
generate, by a target decoder based at least on a first feature expression set including the generated target feature expression and the generated common feature expression, the first result data of performing the first task upon the target domain data;
generate, by a source decoder based at least on a second feature expression set including the generated source feature expression and the generated common feature expression, the second result data of performing the second task upon the source domain data; and
train a combination of the target encoder, the source encoder, the common encoder, the target decoder, and the source decoder,
wherein the first result data of performing the first task upon the target domain data is associated with the learning data, and
wherein the second result data of performing the second task upon the source domain data is associated with the learning data.

9. The system of claim 8 , the computer-executable instructions when executed further causing the system to:
train the combination of the target encoder, the source encoder, the co

Assignees

• Nippon Telegraph & Telephone

Inventors

• Umeda Takayuki
• Murasaki Kazuhiko
• Ando Shingo
• Kinebuchi Tetsuya

Classifications

• G06N3/09

  Supervised learning · CPC title

• G06N3/0464

  Convolutional networks [CNN, ConvNet] · CPC title

• G06N3/0455

  Auto-encoder networks; Encoder-decoder networks · CPC title

• G06N3/08 Primary

  Learning methods · CPC title

• G06F18/214

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