Equitable Loss Function for AI Model Sub-Group Bias
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current AI model training techniques fail to ensure equitable performance across different sub-groups, which is critical for regulatory approval in medical AI models, as they are often unaware of sub-groups during the training and validation process, leading to biased performance in demographic or medical condition meta-information-based sub-groups.
Innovation Solution
The implementation of an 'equitable loss function' (ELF) that identifies and penalizes performance variations across sub-groups during model training and validation, ensuring that the AI model performs uniformly across defined sub-groups by weighting losses inversely to sub-group sample sizes and minimizing standard deviation, thus driving equitable performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If conventional training techniques are used to train AI models, then overall model accuracy can be achieved, but equitable performance across different sub-groups cannot be ensured
Solution Approach 1:
The patent segments the training process by introducing sub-group labels that divide the data into distinct demographic or medical condition groups. The equitable loss function then operates on these segmented groups individually, calculating separate loss values for each sub-group and combining them with appropriate weighting. This segmentation allows the model to be evaluated and optimized for each sub-group's performance separately, ensuring equitable treatment across all groups while maintaining overall model accuracy.
2Reliability
If data augmentation or increased sampling is used to equalize sub-group sample sizes, then sub-group representation can be improved, but training time and computational resources increase
Solution Approach 1:
The patent changes the parameter being optimized from sample size equality to weighted loss equality. Instead of modifying the input data distribution through augmentation or oversampling, the equitable loss function modifies the loss calculation parameters by assigning different weights to different sub-groups based on their sample sizes. This approach achieves balanced sub-group representation in model performance without increasing training time or computational resources, as it operates on the existing data distribution.
3Measurement precision
If sub-group analysis is performed after model training, then performance can be evaluated, but biased performance in specific sub-groups cannot be prevented
Solution Approach 1:
The patent applies preliminary action by incorporating sub-group awareness directly into the training process through the equitable loss function. Rather than performing sub-group analysis after training to evaluate performance, the method proactively prevents biased performance by weighting the loss contributions from different sub-groups during training. This preliminary intervention ensures that the model learns equitable performance across all sub-groups from the beginning, rather than attempting to correct biases after they have been established.
Data Source
AI summary
Techniques are described that facilitate training an artificial intelligence (AI) model in a manner that ensures equitable model performance across different sub-groups. According to an embodiment, a system is provided that includes a memory that stores computer executable components, and a processor that executes the computer executable components stored in the memory. The computer executable components include a training component that trains a machine learning (ML) model on training data to perform an inferencing task using an equitable loss function that drives equitable performance of the ML model across different sub-groups represented by the training data, resulting in trained version of the ML model that provides a defined equitable performance level across the different sub-groups. The equitable loss function is “sub-group aware” and penalizes variation in model performance across the sub-groups during model training and validation.