Distance Metric Relearning with Cluster Integration Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing distance metric learning models are adversely affected by labels set among data during relearning, leading to unintended integration of clusters and model degradation.
Innovation Solution
A learning support function that identifies pairs of clusters likely to be integrated without pre-set labels, based on embedded vectors, and adjusts relearning processes to suppress unintended integration, using a Siamese Network and embedded vector analysis.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Duration of action of stationary object
If relearning is performed with a greater number of epochs, then the model training is more complete, but unintended integration of clusters occurs adversely affecting the model
Solution Approach 1:
The system performs preliminary detection of cluster pairs likely to be integrated before completing the full relearning process. By identifying at-risk cluster pairs early using embedded vectors from intermediate training states, the system can take preventive actions (such as adjusting learning rates, adding regularization, or modifying loss functions) to prevent unintended integration while still allowing beneficial learning to occur.
Solution Approach 2:
The system implements feedback by continuously monitoring the distances between cluster representatives during relearning epochs. When the distance between certain cluster pairs falls below a threshold indicating potential unintended integration, the system provides feedback signals to adjust training parameters or intervene to prevent the harmful integration while maintaining progress in the learning process.
2Productivity
If labels are set among learning samples during relearning, then the training progresses, but unintended integration of clusters occurs
Solution Approach 1:
Before applying labels during relearning, the system performs preliminary analysis using embedded vectors to identify cluster pairs that are at risk of unintended integration. This preliminary action allows the system to prepare preventive measures (such as selective label application, modified loss weighting, or cluster-aware regularization) that enable training progress while protecting against harmful integration.
Solution Approach 2:
The system applies different quality control measures to different parts of the training process. Specifically, it identifies specific cluster pairs that require special attention and applies localized interventions (such as cluster-specific regularization terms or selective label masking) only to those at-risk pairs, while allowing other parts of the training to proceed normally with standard label-based learning.
3Reliability
If clusters are monitored during relearning, then unintended integration can be detected, but additional computational resources are required
Solution Approach 1:
The system extracts only the essential information needed for integration detection by computing and monitoring distances between cluster representatives rather than analyzing all individual sample pairs. This extraction approach significantly reduces computational complexity from O(n²) to O(k²) where k is the number of clusters, enabling effective monitoring with minimal additional computational overhead.
Solution Approach 2:
The system performs partial monitoring by focusing computational resources on detecting integration of specific high-risk cluster pairs identified through preliminary analysis, rather than uniformly monitoring all possible cluster pairs. This partial action approach achieves sufficient detection accuracy for the most critical cases while minimizing overall computational resource consumption.
Data Source
AI summary
A recording medium stores a program for causing a computer to execute a process including: calculating a first embedded vector for each cluster obtained by clustering samples in training data, by inputting the samples that represent the clusters to a first distance metric model; performing training of a second distance metric model, based on labels set in pairs of the samples; calculating a second embedded vector for each cluster, by inputting the samples that represent the clusters to the second distance metric model; detecting pairs of the clusters that are likely to be integrated when the training is performed with a greater number of epochs than a number of epochs at a time of the training of the second distance metric model, based on the first embedded vector and the second embedded vector; and outputting one of the pairs of the clusters in which a similarity label is not set.