Learning Device Discriminative Model Weighted Loss Fragmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods struggle with learning discriminative models for data with fragmented moving locus observations, leading to difficulties in class discrimination, particularly in identifying foreign objects in liquid containers due to factors like lens effects and illumination conditions.
Innovation Solution
A learning device and method that compute a discrimination score using a discriminative model and learn the model with a weighted loss function based on the relative height of the score, utilizing multiple pieces of data corresponding to the same object to mitigate fragmentation issues.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional learning methods are used for discriminative models, then the model can be trained with available data, but the model fails to accurately discriminate fragmented data leading to high error rates
Solution Approach 1:
The patent applies preliminary action by constructing synthetic training data that anticipates fragmentation issues before they occur in real applications. The training data is deliberately created with fragmented moving locus patterns that mimic real-world observation problems, allowing the model to learn robust discrimination features in advance. This preparatory synthetic data construction enables the model to handle fragmented real data reliably without requiring additional real fragmented samples.
Solution Approach 2:
The patent employs parameter changes by modifying the loss function to incorporate discrimination scores as weighting factors. The loss function dynamically adjusts weights based on the confidence levels (discrimination scores) of predictions, giving higher importance to harder discrimination cases. This parameter adjustment in the learning process improves both accuracy and reliability by focusing learning effort on challenging fragmented data patterns.
2Quantity of substance
If multiple pieces of fragmented data are used for training, then more training samples are available, but the fragmentation makes individual data points less informative
Solution Approach 1:
The patent applies merging by aggregating multiple fragmented data pieces belonging to the same object into unified training samples. The system identifies and groups fragmented observations that correspond to the same target object, then combines them to reconstruct more complete moving locus information. This merging process recovers lost information while maintaining the benefits of having multiple training samples, effectively converting quantity of fragments into quality of complete trajectories.
Solution Approach 2:
The patent uses copying by creating synthetic replicas of fragmented data patterns through simulation. Instead of relying solely on limited real fragmented data, the system generates multiple copies of fragmented trajectories with varied characteristics through computational simulation. These synthetic copies expand the training dataset while preserving the essential fragmentation patterns, allowing the model to learn from abundant replicated examples without requiring proportionally more real fragmented observations.
Data Source
AI summary
A learning device includes a learning means for learning a discriminative model that discriminates a class to which second data belongs, the second data being data corresponding to an unknown object, by using first training data that includes a group including a plurality of pieces of first data corresponding to the same object, and a first data label with respect to the group. The learning means computes a discrimination score with respect to the first data by using the discriminative model, and learns the discriminative model by using a loss weighted by a weight that depends on a relative height of the discrimination score in the group.