Learning device, program, and learning method

The learning device and method address domain adaptation challenges by determining and retraining specific layers of deep learning models based on activation level comparisons, improving accuracy in diverse domains.

JP7740560B2Active Publication Date: 2025-09-17KONICA MINOLTA INC
View PDF 5 Cites 0 Cited by

Patent Information

Application Number
JP2024540063
Authority / Receiving Office
JP · JP
Patent Type
Patents
Current Assignee / Owner
Priority Date
2023-01-13
Filing Date
2023-10-12
Publication Date
2025-09-17
Estimated Expiration
2043-10-12

AI Technical Summary

Technical Problem

Existing deep learning models struggle with domain adaptation when the difference between the source and target domains is significant, leading to insufficient accuracy, and the determination of learning conditions relies heavily on expert intuition.

Method used

A learning device and method that determine the layers of a deep learning model to be re-learned based on comparisons of activation levels and distributions between source and target domain data, adjusting the model's parameters to improve accuracy.

Benefits of technology

Enables effective relearning of deep learning models by targeting specific layers for retraining, enhancing classification accuracy even with limited target domain data, especially when domains differ significantly.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 0007740560000001
    Figure 0007740560000001
  • Figure 0007740560000002
    Figure 0007740560000002
  • Figure 0007740560000003
    Figure 0007740560000003
Patent Text Reader

Abstract

A learning device (100) comprises: a re-learning unit (113) that causes a deep learning model having been trained by using first domain data (see source domain data (140)) to undergo re-learning by using second domain data (see target domain data (150)); and a learning target layer determination unit (112) that determines a layer of a deep learning model to be a target of re-learning on the basis of either a node which is activated by the first domain data and which is included in the layer of the deep learning model or a node which is activated by the second domain data and which is included in the layer of the deep learning model.
Need to check novelty before this filing date? Find Prior Art

Citation Information

Patent Citations

  • Machine learning device, method for mechanical learning, and learned model

    JP2022085356A

  • Machine learning device, machine learning method, and machine learning program

    JP2022122326A

  • Semantic coherence analysis of deep neural networks

    US20210117778A1

  • Machine learning device, machine learning method, and trained model

    WO2022113534A1

  • Machine learning device, machine learning method, and machine learning program

    WO2022172569A1