Semi-supervised Learning Using Distance-Based Score Assignment
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing semi-supervised learning techniques face challenges in accurately assigning labels to unlabeled data due to low separation properties or insufficient data, leading to reduced learning accuracy, especially when data distribution assumptions are not met.
Innovation Solution
A learning apparatus that sets scores for unlabeled data based on their distance to labeled data and uses a neural network to learn with both labeled and scored unlabeled data, employing label vectors to improve label assignment accuracy without relying on strong separation properties or sufficient data distribution knowledge.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If clustering is used to assign labels to unlabeled data, then the use of unlabeled data is enabled, but the accuracy of label assignment deteriorates when data has low separation property
Solution Approach 1:
The patent changes the parameter representation from discrete cluster assignments to continuous probability distributions. By representing label assignments as probability vectors that evolve through learning iterations, the system can capture nuanced relationships between data points and labels, overcoming the limitations of rigid clustering boundaries when separation property is low.
Solution Approach 2:
The patent introduces dynamic label assignment where probability distributions are iteratively updated during learning. Instead of static cluster assignments, the system continuously refines label probabilities based on learned representations, allowing adaptability to data structures with low separation property while maintaining accuracy through iterative optimization.
2Stability of the object's composition
If label assignment prioritizes cluster constraints, then clustering structure is maintained, but accuracy near cluster boundaries deteriorates
Solution Approach 1:
The patent transforms the label assignment from discrete cluster IDs to continuous probability values. This parameter change allows data points near boundaries to have probabilistic assignments reflecting their ambiguity, rather than being forced into rigid clusters. The probability vectors capture boundary uncertainty while maintaining overall clustering structure through the learned distribution shapes.
Solution Approach 2:
The patent applies partial label assignments through probability distributions rather than full commitment to single clusters. By allowing partial membership in multiple label classes simultaneously through probability vectors, the system maintains clustering structure while accurately representing boundary cases that don't fit cleanly into single clusters.
3Ease of manufacture
If distribution assumptions are made for label assignment, then theoretical framework is provided, but performance deteriorates when assumptions are not satisfied
Solution Approach 1:
The patent uses dynamic learning to adapt probability distributions directly from data rather than assuming fixed distributions a priori. The system iteratively learns label probability distributions through neural network training, allowing the distributions to emerge from the data itself rather than being constrained by theoretical assumptions, thus maintaining reliability when assumptions are violated.
Solution Approach 2:
The system performs self-service by automatically learning appropriate label probability distributions from the data without requiring external specification of distributional assumptions. The neural network learns the underlying data distribution and label relationships directly, making the method self-adapting to various data structures without relying on pre-established theoretical frameworks.
Data Source
AI summary
A non-transitory computer-readable recording medium stores therein a learning program that causes a computer to execute a process including: setting each of scores to each of a plurality of sets of unlabeled data with regard to each of labels used in a plurality of sets of labeled data based on a distance of each of the plurality of sets of unlabeled data with respect to each of the labels; and causing a learning model to learn using a neural network by using the plurality of sets of labeled data respectively corresponding to the labels of the plurality of sets of labeled data, and the plurality of sets of unlabeled data respectively corresponding to the scores of the plurality of sets of unlabeled data with regard to the labels.