ML Model Bias Detection with Independent Test Data Retraining
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning models often exhibit bias, particularly in edge devices, leading to unfair and discriminatory outputs, and existing methods lack effective mechanisms for independent and secure bias detection and mitigation.
Innovation Solution
A system and method for periodically testing ML models for bias by injecting bias test data, generating alerts, and incrementally re-training models to maintain bias below a threshold, using separate ownership and control of bias test data from model code/weights, and employing a bias cost function to enforce fairness.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If ML models are deployed in edge devices with limited resources, then deployment efficiency is improved, but model bias detection and mitigation capability deteriorates due to lack of independent testing mechanisms
Solution Approach 1:
The patent segments the ML model deployment into two independent parts: the model code/weights and the bias test data. This segmentation allows the model to be deployed efficiently while separate bias testing mechanisms can independently evaluate and mitigate bias without interfering with deployment performance.
Solution Approach 2:
The patent introduces an intermediary bias testing system that acts as a mediator between model deployment and bias mitigation. This intermediary independently tests for bias using separate test data and provides feedback for retraining, enabling bias detection without compromising deployment efficiency.
2Device complexity
If bias test data is integrated with model code/weights, then system complexity is reduced, but independence and security of bias detection deteriorates
Solution Approach 1:
The patent explicitly segments bias test data from model code/weights, maintaining separate ownership and control. This segmentation ensures that bias detection remains independent and secure, preventing potential manipulation while managing system complexity through modular architecture.
Solution Approach 2:
The patent introduces an intermediary layer that manages the separation between model code and bias test data. This intermediary enables independent bias detection while maintaining system coherence, allowing the system to remain manageable despite the added separation.
3Reliability
If continuous bias testing is performed on deployed ML models, then fairness and non-discriminatory outputs are improved, but computational overhead and resource consumption increase
Solution Approach 1:
The patent implements periodic bias testing rather than continuous testing, where bias test data is injected at intervals to evaluate model fairness. This periodic approach maintains output fairness while significantly reducing computational overhead and resource consumption compared to continuous monitoring.
Solution Approach 2:
The patent performs preliminary bias testing during model deployment and retraining phases, identifying and mitigating bias before models are fully deployed. This preliminary action ensures fairness while minimizing ongoing computational overhead during production operations.
Data Source
AI summary
Testing for bias in a machine learning (ML) model in a manner that is independent of the code/weights deployment path is described. If bias is detected, an alert for bias is generated, and optionally, the ML model can be incrementally re-trained to mitigate the detected bias. Re-training the ML model to mitigate the bias may include enforcing a bias cost function to maintain a level of bias in the ML model below a threshold bias level. One or more statistical metrics representing the level of bias present in the ML model may be determined and compared against one or more threshold values. If one or more metrics exceed corresponding threshold value(s), the level of bias in the ML model may be deemed to exceed a threshold level of bias, and re-training of the ML model to mitigate the bias may be initiated.