Classification Device Using Pseudo Data for Unrecognized Input Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In high-performance machine learning classification systems, such as deep learning, there is a risk that data difficult to classify may be incorrectly classified into classes with high probability or credibility, leading to low precision in presenting data that requires user review, particularly in medical image classification.
Innovation Solution
A classification device employing an Improved Generative Adversarial Network (IGAN) generates pseudo data similar to training data and learns a classification model to accurately classify input data into (K+1) classes, including a pseudo class, to identify and output data that does not belong to any classification class, thereby improving the precision of presenting difficult-to-classify data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a complicated classification model is created from training data using deep learning, then classification performance is improved, but the probability value or credibility of a class does not always represent the difficulty in classification
Solution Approach 1:
The patent introduces an intermediary mechanism (difficulty evaluation unit) that assesses classification difficulty based on multiple factors including probability values, credibility, and decision boundary distances, rather than relying solely on probability values. This intermediary evaluation layer resolves the contradiction by providing a more accurate representation of classification difficulty that complements the high-performance classification model.
Solution Approach 2:
The patent changes the parameters used to evaluate classification difficulty from simple probability values to a composite metric that includes probability values, credibility scores, and distances to decision boundaries. This parameter transformation allows the system to maintain high classification performance while accurately identifying difficult-to-classify instances.
2Productivity
If data with high probability value or high credibility is classified automatically, then classification efficiency is improved, but data difficult to be classified may be incorrectly classified
Solution Approach 1:
The patent implements a feedback mechanism where the difficulty evaluation unit continuously assesses classification difficulty and provides feedback to determine whether automatic classification or expert review is appropriate. This feedback loop ensures that high-probability classifications are efficient while difficult cases are identified and referred to experts, maintaining both productivity and reliability.
Solution Approach 2:
The patent introduces dynamic adjustment of classification strategy based on real-time difficulty assessment. The system dynamically determines whether to apply automatic classification or expert review for each input instance based on its classified difficulty level, allowing flexible adaptation between efficiency and accuracy requirements.
3Ease of operation
If a simple classification approach is used, then ease of operation is improved, but precision in presenting difficult-to-classify data is reduced
Solution Approach 1:
The patent segments the classification system into distinct functional modules: a high-performance classification model for accurate classification, a difficulty evaluation unit for assessing classification difficulty, and an output control unit for determining presentation strategy. This segmentation allows each module to specialize in its function while working together to achieve both operational simplicity and precise difficulty detection.
Data Source
AI summary
A classification device includes a generation unit, a learning unit, a classification unit, and an output control unit. The generation unit generates pseudo data having a feature similar to a feature of training data. The learning unit learns, by using the training data and the pseudo data, a classification model that classifies data into one of a pseudo class for classifying the pseudo data and a plurality of classification classes other than the pseudo class and that is constructed by a neural network. The classification unit classifies, by using the classification model, input data as a target for classification into one of the pseudo class and the plurality of classification classes. The output control unit outputs information indicating that the input data classified into the pseudo class is data not belonging to any of the plurality of classification classes.