Class-Adaptive Percentile Thresholds for Multi-Label Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing semi-supervised learning methods struggle to effectively leverage unlabeled data for multi-label classification tasks due to challenges such as class independence, class imbalance, and varying learning difficulties across classes, leading to noisy pseudo-labels and suboptimal performance.
Innovation Solution
A percentile-based threshold adjusting scheme is introduced for positive and negative pseudo-labels, combined with dynamic unlabeled loss weights, to adaptively select and weight pseudo-labels based on class-specific learning status, reducing noise and improving model performance in multi-label classification.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If semi-supervised learning is applied to multi-label classification, then the ability to leverage unlabeled data is improved, but the quality of pseudo-labels deteriorates due to class independence and class imbalance
Solution Approach 1:
The patent applies local quality by introducing class-specific thresholding mechanisms that adapt to the unique characteristics of each class. Different classes have different difficulty levels and distribution patterns, so a unified threshold approach fails. The method adjusts thresholds and loss weights individually for each class based on its learning status, ensuring that pseudo-labels are generated with appropriate quality control for each specific class context.
Solution Approach 2:
The patent implements dynamics through dynamic threshold adjustment and adaptive loss weighting that evolve during training. The thresholds for positive and negative pseudo-labels are not fixed but adapt based on class performance and distribution. Loss weights are dynamically adjusted to down-weight difficult classes and up-weight easier ones, allowing the system to maintain reliable pseudo-labels across heterogeneous multi-label datasets.
2Device complexity
If a fixed threshold is used for pseudo-label selection, then the simplicity of the method is maintained, but the performance deteriorates due to varying learning difficulties across classes
Solution Approach 1:
The patent transitions from static to dynamic thresholding by adjusting thresholds based on class-specific learning status. The method calculates adaptive thresholds using percentile-based approaches that respond to the distribution of predictions for each class. This dynamic adjustment maintains simplicity in the overall framework while significantly improving accuracy by tailoring thresholds to each class's specific challenges.
Solution Approach 2:
The patent applies parameter changes by modifying threshold values and loss weights based on class characteristics. Instead of using fixed thresholds, the method computes adaptive parameters (thresholds, weights) that change according to class distribution, prediction confidence, and learning progress. This parameter adaptation allows the same simple pseudo-labeling framework to achieve high performance across diverse classes with varying difficulties.
3Ease of operation
If all classes are treated equally in multi-label classification, then the fairness of treatment is maintained, but the effectiveness of training deteriorates due to class imbalance and varying difficulty
Solution Approach 1:
The patent applies local quality by transitioning from uniform to differentiated treatment of classes. Each class receives customized threshold settings and loss weights based on its specific characteristics such as frequency, difficulty, and distribution. This localized adaptation ensures that training effectiveness is optimized for each class while maintaining a coherent overall framework that treats all classes within the multi-label context.
Solution Approach 2:
The patent implements parameter changes through adaptive loss weighting and dynamic threshold adjustment that respond to class-specific metrics. The method modifies training parameters (loss weights, thresholds) based on class frequency, prediction confidence distributions, and learning status. This parameter adaptation enables the system to address class imbalance and varying difficulty while preserving operational simplicity through automated parameter adjustment.
Data Source
Figure 1A~1C
Figure 2A~2C
Figure 3A~3C
AI summary
A method and system provide for augmenting a photograph. An unlabeled photograph is obtained. A weakly augmented photograph and a strongly augmented photograph are obtained from the unlabeled photograph based on different types of data augmentation methods. The weakly augmented photograph is processed through a model to generate multiple weakly augmented photograph class predictions (with assigned probabilities). The multiple weakly augmented photograph class predictions are converted into positive pseudo-labels (indicating a presence of a class) or negative pseudo-labels (indicating absence of a class) using different fixed percentile thresholds. The strongly augmented photograph is processed through the model to generate a strongly augmented photograph class prediction. The model is trained to make the strongly augmented photograph label prediction match the positive pseudo-label via a cross-entropy loss. The trained model is then utilized to label the unlabeled photograph with multiple labels.