Neural Network Classification of Rare Cases Using Dual Feature Spaces
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing deep neural networks struggle to accurately classify rare or unique data samples that significantly differ from previously learned data, leading to misclassification and limited improvement in decision spaces due to bounded feature spaces and weak feedback mechanisms.
Innovation Solution
A method involving a deep neural network that combines universal and class-specific feature spaces to classify both visually similar and visually non-similar data samples, utilizing unlabelled data to construct feature and decision spaces that can handle rare cases effectively, with a penalty matrix to focus training on high-risk misclassification areas.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a single deep neural network is trained on labelled data samples, then it can classify common cases effectively, but it cannot cope with previously unseen data samples efficiently at the classification stage
Solution Approach 1:
The system segments the classification task into two specialized networks: a first deep neural network trained on labelled data for common cases, and a second deep neural network trained on unlabelled data for unseen/rare cases. This segmentation allows each network to specialize in its respective domain, resolving the contradiction between reliability on common cases and adaptability to unseen data.
Solution Approach 2:
The system extends the feature space by introducing a second dimension of training data (unlabelled data) that is distinct from the original labelled data. This dimensional expansion allows the system to capture patterns in unseen data without compromising the classification performance on common cases, effectively adding a new capability dimension.
2Reliability
If an ensemble of models is used to reduce false alarms through agreement, then classification reliability improves, but device complexity increases
Solution Approach 1:
The system extracts the false alarm reduction capability from the ensemble approach and implements it within a single deep neural network through self-transfer learning. The network learns to identify and correct its own misclassifications by comparing predictions across different input perturbations, eliminating the need for multiple separate models while maintaining the reliability benefit.
3Quantity of substance
If data augmentation is applied to existing labelled data samples, then the training dataset size increases, but the feature space remains bounded by existing labelled data
Solution Approach 1:
The system creates synthetic training samples by copying and transforming unlabelled data through various augmentations (rotations, flips, color adjustments). These copied and transformed samples expand the training dataset and feature space without requiring additional labelled data, allowing the network to learn from a broader range of variations while maintaining training efficiency.
Data Source
Figure 1
Figure 2~4
Figure 5~7
AI summary
A method to reinforce deep neural network learning capacity to classify rare cases, which comprises the steps of training a first deep neural network (DNN-A) used to classify generic cases of original data (Data-A) into specified labels (Label-A); localizing discriminative class-specific features within the original data processed through DNN-A and mapping the discriminative class-specific features as spatial-probabilistic labels (Label-B); training a second deep neural network (DNN-C) used to classify rare cases of the original data into the spatial-probabilistic labels; and training a combined deep neural network (DNN-D) used to classify both generic and rare cases of the original data into primary combined specified and spatial-probabilistic labels (Label-A+B*).