Transfer Learning Similarity Screening Using Extremal Value Distributions
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for transfer learning in IoT equipment face high computation complexity due to the evaluation of similarity between training data and transfer sources, especially when using image data.
Innovation Solution
A similarity degree calculation device that evaluates transfer sources by calculating similarity between transfer source and destination data distributions based on extremal value groups, reducing computation complexity through feature extraction and extremal value analysis.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional similarity calculation methods (color histogram, gradient direction, local features) are used to evaluate transfer source candidates, then measurement precision of similarity is improved, but computation complexity increases
Solution Approach 1:
The patent extracts only the extremal values (maximum and minimum) from the complete feature amount distributions of transfer source and destination data. Instead of calculating full histograms or comparing all local features, the method identifies and compares only the extreme points of the distributions, significantly reducing computational complexity while maintaining the ability to assess similarity accuracy.
Solution Approach 2:
The patent uses a simplified, computationally inexpensive approach by relying solely on extremal value comparisons rather than expensive comprehensive similarity metrics. This disposable-like simplification sacrifices some detailed information but achieves sufficient similarity assessment with minimal computational resources.
2Reliability
If all transfer source candidates are evaluated one by one using comprehensive similarity metrics, then reliability of transfer source identification is improved, but productivity decreases
Solution Approach 1:
The patent extracts only the essential extremal values from each transfer source candidate's feature distribution, enabling rapid comparison with the destination data distribution. This extraction approach maintains reliability by focusing on the most discriminative features (extremes) while dramatically improving evaluation speed through reduced computational burden.
Solution Approach 2:
The patent performs partial action by comparing only extremal values rather than complete feature distributions. This partial comparison is sufficient for reliable transfer source identification in many cases, achieving a balance between reliability and productivity by doing just enough computation to make accurate decisions.
Data Source
AI summary
A similarity degree calculation device (1) includes a similarity degree calculation unit (19) and performs transfer learning with using a transfer source data candidate (2) and transfer destination data (3). The similarity degree calculation unit (19) finds a similarity degree between a transfer source data distribution and a transfer destination data distribution on a basis of a transfer source extremal value group and a transfer destination extremal value group. The transfer source extremal value group includes a transfer source extremal value indicating an extremal value of the transfer source data distribution indicating a distribution of a feature amount of the transfer source data candidate (2). The transfer destination extremal value group includes a transfer destination extremal value indicating an extremal value of the transfer destination data distribution indicating a distribution of a feature amount of the transfer destination data (3).