Multiheaded Inference Model Latent Bias Management
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Inference models often exhibit latent bias due to training data limitations, leading to inaccurate and discriminatory predictions, which negatively impact computer-implemented services by introducing undesired correlations and discrimination.
Innovation Solution
A system and method for managing inference models through modified split training, where a multipath neural network is used with a shared body and separate heads for label and bias feature prediction, and features contributing to bias are identified and removed from the training data set to reduce latent bias.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional training methods are used to improve predictive power, then the model's accuracy improves, but latent bias increases
Solution Approach 1:
The inference model is segmented into multiple independent heads (feature head, bias feature head, label head) that process different aspects of the input data separately. This segmentation allows the model to learn predictive features while explicitly identifying and separating bias features, resolving the contradiction between accuracy and bias by handling them through distinct processing pathways.
Solution Approach 2:
The bias feature head is extracted as a separate component from the main inference model. This extracted head specifically identifies bias features in the input data, allowing the system to remove or mitigate these bias features before they affect the final predictions, thereby reducing latent bias while maintaining predictive accuracy through the remaining clean features.
2Productivity
If more features are added to improve model performance, then predictive power increases, but bias contribution increases
Solution Approach 1:
The bias feature head acts as an intermediary component between the input features and the label prediction. It identifies bias features and prevents them from contaminating the label head's predictions. This intermediary mechanism allows the model to utilize all input features for predictive power while filtering out bias contributions through the mediating bias detection layer.
Solution Approach 2:
Different heads of the model are assigned different local qualities or functions: the feature head focuses on predictive features, the bias feature head focuses on identifying bias, and the label head focuses on final prediction. This local specialization allows each component to optimize its specific function, enabling the overall system to achieve high predictive power while minimizing bias through localized bias detection and removal.
Data Source
AI summary
Methods, systems, and devices for providing computer-implemented services are disclosed. To provide the computer-implemented services, inference models used by data processing systems may be managed to reduce latent bias. The inference models may be managed by establishing supervised models based on the results of unsupervised learning. The supervised models may then be subjected to training to reduce the levels of latent bias, and analysis to identify features contributing to the latent bias. The supervised learning may then be performed without consideration for the identified features.