Inclusive ML Model Training via Auxiliary Data Feature Space
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current machine learning models often exhibit bias due to under-representation of demographic groups in training and testing datasets, leading to poor performance for under-represented populations.
Innovation Solution
The method involves generating a common feature space from training data and auxiliary data, using data augmentation and pseudo-labeling to create a combined learned feature representation, which is then used to train an inclusive machine learning model that improves performance for under-represented groups without compromising overall model accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If training data is used to train machine learning models, then model performance is improved, but bias against under-represented demographic groups is introduced
Solution Approach 1:
The patent introduces auxiliary data as an intermediary component that mediates between the biased training data and the final model. The auxiliary data, which contains demographic information about under-represented groups, serves as a bridge to transfer knowledge and reduce bias without directly modifying the original training data. This intermediary data source enables the model to learn fairer representations while maintaining performance on the primary task.
2Adaptability or versatility
If additional labeled data for under-represented groups is collected, then inclusivity is improved, but data collection cost and time increase
Solution Approach 1:
The patent creates pseudo-labeled copies of auxiliary data by using the trained model to generate labels for unlabeled auxiliary data. Instead of manually collecting and labeling additional data for under-represented groups, the system copies the structure and patterns from the auxiliary data and applies synthetic labels generated by the model itself. This copying approach enables scalable data augmentation without the time-consuming manual labeling process.
Solution Approach 2:
The system performs self-service by using its own trained model to generate pseudo-labels for the auxiliary data. The model serves itself by creating the labeling mechanism it needs, eliminating the requirement for external human annotators. This self-labeling capability allows the system to continuously improve inclusivity without incurring additional data collection costs or time delays.
3Reliability
If data augmentation and pseudo-labeling are applied, then performance for under-represented groups is improved, but model complexity increases
Solution Approach 1:
The patent applies preliminary action by first training an initial model on the biased training data before using it to generate pseudo-labels for the auxiliary data. This preliminary model training establishes a baseline that can then be used to create synthetic labels, simplifying the overall process compared to attempting to train a fair model directly from scratch. The preliminary action creates a foundation that makes subsequent debiasing operations more manageable and less complex.
Data Source
AI summary
Transfer knowledge from auxiliary data for more inclusive machine learning models is provided. A method can include generating a common feature space comprising first data features, wherein the first data features are present in training data used to train a first machine learning model, and wherein the first data features are present in auxiliary data that are independent of the training data; generating a combined learned feature representation, the combined learned feature representation being representative of the first data features of the common feature space and second data features that are unique to the training data; and training a second machine learning model based on the combined learned feature representation.