Looking for breakthrough ideas for innovation challenges? Try Patsnap Eureka!

Incremental learning method driven by task correlation

An incremental learning and correlation technology, applied in the field of image processing, can solve the problem of catastrophic forgetting of old tasks of the model

Inactive Publication Date: 2021-09-14
UNIV OF ELECTRONICS SCI & TECH OF CHINA
View PDF0 Cites 0 Cited by
  • Summary
  • Abstract
  • Description
  • Claims
  • Application Information

AI Technical Summary

Problems solved by technology

However, this can easily lead to catastrophic forgetting of old tasks by the model

Method used

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
View more

Image

Smart Image Click on the blue labels to locate them in the text.
Viewing Examples
Smart Image
  • Incremental learning method driven by task correlation
  • Incremental learning method driven by task correlation
  • Incremental learning method driven by task correlation

Examples

Experimental program
Comparison scheme
Effect test

Embodiment 1

[0024] Such as figure 1 As shown, the present embodiment provides a task-related incremental learning method, which includes the following steps:

[0025] image 3 Shows how a 4*4 filter in R&R-Net changes as the training progresses. Next, in a sequential fashion figure 2 Be explained:

[0026] Step 1, such as image 3 As shown in (a), before training task 1, the filter is initialized first, and all parameters in the initialized filter can be used for the training of task 1, and the labels of these parameters are marked as 1. After the initialized model learns for a certain period on the training set of task 1, it is pruned in order to cut out the redundant parameters in the model, and these redundant parameters can also be used for the learning of subsequent tasks. After pruning, it is determined which parameters in the filter are reserved for task 1 and which are available for subsequent tasks. The parameters reserved for task 1 will not be changed in the training of ...

Embodiment 2

[0036] In this embodiment, the method in the embodiment 1 is compared with the prior art and the experimental analysis is carried out.

[0037] First, we introduce the datasets used in the experiments. In total, three different datasets were used in the experiments:

[0038] (1) LIVE II: jointly established by the Department of Electrical and Computer Engineering and the Department of Psychology at the University of Texas at Austin. Contains 29 reference images and 779 distorted images, and the distorted images contain 5 types of distortion;

[0039] (2) LIVEMD: Established by the Image and Video Engineering Laboratory of the University of Texas at Austin, it contains two subsets. The first subset mixes blur distortion and JPEG distortion, the second subset mixes blur distortion and Gaussian noise, and each subset contains 225 distorted images generated from 15 original images;

[0040] (3) IVIPC_DQA: Established by the Intelligent Visual Information Processing and Communic...

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
Login to View More

PUM

No PUM Login to View More

Abstract

The invention relates to the technical field of image processing, and relates to a task correlation driven incremental learning method. The method comprises the following steps: 1, trimming a model after learning for a certain period on a training set of a task 1; performing fine adjustment on the model for a certain period after trimming to obtain a model 1; 2, calculating the correlation between the task 2 and the task 1; 3, initializing the parameters trimmed in the step 1 and learning the parameters by using the training data of the task 2; after learning is completed, trimming parameters of a fixed proportion, and performing fine tuning on the remaining parameters belonging to the task 2 by using the training data of the task 2 to obtain a model 2; 4, during subsequent task learning, calculating the correlation between the task and all the first tasks, and then learning by adopting a method which is the same as that in the step 3. According to the method, not only can disastrous forgetting be effectively avoided through the memory template, but also learning of the current task can be assisted by utilizing the correlation between the tasks, and the performance of the model on the current task is improved.

Description

technical field [0001] The present invention relates to the technical field of image processing, in particular to a task correlation-driven incremental learning method. Background technique [0002] When the original reference image is not available, blind image quality assessment (Blind Image Quality Assessment, BIQA) came into being. Many successful BIQA methods follow a one-to-many learning strategy, and this type of method uses the entire training data simultaneously to learn a static predictive model. These methods have achieved considerable success in many task-specific applications, including synthetic distortion, natural distortion, or augmented distortion task evaluation. [0003] In recent years, with the rapid development of image technology, multiple evaluation criteria and distortion types have emerged, which also bring new challenges to cross-task BIQA. The process of cross-task BIQA such as figure 2 shown. As new evaluation tasks are sequentially added, BI...

Claims

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
Login to View More

Application Information

Patent Timeline
no application Login to View More
IPC IPC(8): G06N20/00G06F9/48
CPCG06N20/00G06F9/4843
Inventor 吴庆波马瑞李宏亮孟凡满许林峰潘力立
Owner UNIV OF ELECTRONICS SCI & TECH OF CHINA
Who we serve
  • R&D Engineer
  • R&D Manager
  • IP Professional
Why Patsnap Eureka
  • Industry Leading Data Capabilities
  • Powerful AI technology
  • Patent DNA Extraction
Social media
Patsnap Eureka Blog
Learn More
PatSnap group products

Browse by: Latest US Patents, China's latest patents, Technical Efficacy Thesaurus, Application Domain, Technology Topic, Popular Technical Reports.

© 2024 PatSnap. All rights reserved.Legal|Privacy policy|Modern Slavery Act Transparency Statement|Sitemap|About US| Contact US: help@patsnap.com