# US 2020/0265308 A1 · Model optimization method, data identification method and data identification device

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2020/0265308 A1
- Document ID: US20200265308A1
- Status: Published application
- Kind: A1
- Application number: 16/748,107
- Filed: 2020-01-21
- Published: 2020-08-20
- Earliest claimed: 2019-02-20
- Assignees: Fujitsu Limited
- Applicants: Fujitsu Limited
- Inventors: Mengjiao Wang; Rujie Liu
- Web page: https://patentgrep.com/patent/US20200265308A1

## CPC

- G06N 3/084: Backpropagation, e.g. using gradient descent
- G06N 3/04: Architecture, e.g. interconnection topology
- G06N 3/045: Combinations of networks
- G06N 3/0464: Convolutional networks [CNN, ConvNet]
- G06N 3/08: Learning methods
- G06N 3/088: Non-supervised learning, e.g. competitive learning
- G06N 3/09: Supervised learning
- G06V 10/764: using classification, e.g. of video objects
- G06V 10/82: using neural networks
- G06N 3/048: Activation functions
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06V: IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING

## Abstract

The present disclosure relates to a model optimization method, a data identification method and a data identification device. A method for optimizing a data identification model comprises: acquiring a loss function of a data identification model to be optimized; calculating weight vectors in the loss function which correspond to classes; performing normalization processing on the weight vectors; updating the loss function by increasing an included angle between any two of the weight vectors; optimizing the data identification model to be optimized based on the updated loss function.

## Claims

### Claim 1

A method for optimizing a data identification model, comprising:
acquiring a loss function of a data identification model to be optimized;
calculating weight vectors in the loss function which correspond to classes;
performing normalization processing on the weight vectors;
updating the loss function by increasing an included angle between any two of the weight vectors; and
optimizing the data identification model to be optimized based on the updated loss function.

### Claim 2 (depends on 1)

The method according to claim 1, wherein the data identification model to be optimized is obtained through deep neural network training.

### Claim 3 (depends on 1)

The method according to claim 1, wherein optimizing the data identification model to be optimized based on the updated loss function further comprises:
training a new data identification model based on the updated loss function, wherein the new data identification model is a data identification model after optimization is performed on the data identification model to be optimized.

### Claim 4 (depends on 2)

The method according to claim 2, wherein the data identification model is obtained through training by causing a function value of the loss function to be minimum.

### Claim 5 (depends on 1)

The method according to claim 1, wherein the loss function comprises Softmax loss function.

### Claim 6 (depends on 1)

The method according to claim 1, wherein one class corresponds to one weight vector.

### Claim 7 (depends on 6)

The method according to claim 6, wherein the weight vectors are M-dimensional vectors, where M is an integer greater than 1.

### Claim 8 (depends on 1)

The method according to claim 1, wherein the data comprises one of image data, voice data or text data.

### Claim 9 (depends on 1)

The method according to claim 1, wherein the loss function comprises Logit loss function.

### Claim 10 (depends on 1)

The method according to claim 1, wherein the loss function comprises feature loss function.

### Claim 11 (depends on 2)

The method according to claim 2, wherein the data identification model to be optimized is obtained through convolutional neural network training.

### Claim 12

A data identification method, comprising:
performing data identification using an optimized data identification model obtained by a method for optimizing a data identification model comprising:
acquiring a loss function of a data identification model to be optimized;
calculating weight vectors in the loss function which correspond to classes;
performing normalization processing on the weight vectors;
updating the loss function by increasing an included angle between any two of the weight vectors; and
optimizing the data identification model to be optimized based on the updated loss function.

### Claim 13

A computer readable recording medium having stored thereon program instructions that, when executed by a computer, are used for implementing a method for optimizing a data identification model comprising:
acquiring a loss function of a data identification model to be optimized;
calculating weight vectors in the loss function which correspond to classes;
performing normalization processing on the weight vectors;
updating the loss function by increasing an included angle between any two of the weight vectors; and
optimizing the data identification model to be optimized based on the updated loss function.
