Augmented Training Data for Rare-Case Feature Differentiation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing AI models are biased towards non-distinguishing characteristics, leading to errors in differentiating parts-of-interest, particularly in rare cases, due to training on naturally biased data, resulting in failures that can be dangerous, especially in medical applications.
Innovation Solution
The method involves data augmentation and corresponding labeling to emphasize distinguishing characteristics in the training dataset, creating an augmented sample set that teaches the model to focus on these features, reducing the likelihood of overfitting and improving differentiation performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If naturally biased training data is used, then the model can process common cases efficiently, but the model becomes biased and fails on rare conditions
Solution Approach 1:
The patent applies preliminary action by proactively generating augmented training data that includes rare conditions before the model is trained. The system identifies rare conditions in the training data and generates augmented samples by combining rare and common samples, ensuring the model learns to handle rare conditions correctly from the outset rather than attempting to fix bias after training.
Solution Approach 2:
The patent changes the parameter composition of the training data by adjusting the proportion of rare to common samples through augmentation. Instead of using natural data distributions, the system modifies data parameters by generating synthetic samples with controlled characteristics, transforming the training data to balance rare and common condition representation.
2Speed
If the model focuses on non-distinguishing characteristics like backgrounds, then it can quickly process images, but it misclassifies rare conditions such as melanoma as acne
Solution Approach 1:
The patent applies local quality by making the training data itself have different characteristics for different regions. Through augmentation, the system creates samples where distinguishing characteristics (local features) are enhanced and emphasized, while non-distinguishing characteristics (backgrounds) are either reduced or made varied. This ensures the model learns to focus on locally important features for accurate classification.
Solution Approach 2:
The system performs preliminary analysis of training data to identify samples where non-distinguishing characteristics dominate, then proactively generates augmented samples that correct this imbalance by emphasizing distinguishing characteristics before the model training begins, preventing the model from learning incorrect patterns.
3Reliability
If high sensitivity AI is designed to detect all rare conditions, then it can reduce false negatives, but it increases false positives and is easier to approve while being more dangerous
Solution Approach 1:
The patent changes the training data parameters to include augmented samples that balance rare and common conditions. By modifying the data composition through augmentation techniques, the system enables the model to achieve high sensitivity for rare conditions without the extreme false positive rates associated with traditional high-sensitivity approaches, as the augmented data teaches the model to distinguish subtle patterns accurately.
Solution Approach 2:
The system performs preliminary data augmentation to pre-teach the model accurate differentiation between rare and common conditions before deployment. This preliminary training with balanced augmented data prevents the model from adopting overly sensitive thresholds that cause excessive false positives, while still maintaining high detection capability for rare conditions.
Data Source
AI summary
A method and apparatus is provided for generating an augmented sample set for enriching a first training dataset for training a model. The method comprises: using data augmentation and corresponding labeling or using label augmentation to add a first augmented sample set to the first training dataset, wherein the data augmentation and corresponding labeling, or the label augmentation purposely puts a first distinguishing characteristic of a first part-of-interest or an associated label into the first training dataset to cause the first distinguishing characteristic of the first part-of-interest to be emphasized to enable the model to learn a generalizable principle of the first distinguishing characteristic, wherein the first distinguishing characteristic is for differentiating the first part-of-interest from a second part-of-interest. Methods for training a model, using a model to differentiate part-of-interests and using a model to infer a dataset are also provided.