Multipath Inference Model Split Training for Latent Bias
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Inference models often exhibit latent bias due to training data that includes non-obvious correlations, leading to undesirable outcomes in computer-implemented services, such as discrimination against certain groups.
Innovation Solution
A modified split training method is employed to manage inference models by dividing them into multipath models with shared bodies and independent heads, where weights are frozen and unfrozen to fine-tune the model, reducing latent bias by controlling the training and untraining cycles for feature and bias predictions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If inference models are trained to improve predictive power for features, then the model's ability to make accurate predictions improves, but latent bias in the model increases
Solution Approach 1:
The inference model is divided into multiple independent inference paths, each headed by a different predictor. This segmentation allows the model to make multiple predictions for the same target variable using different feature sets, thereby reducing reliance on any single biased predictor while maintaining overall predictive power through aggregation of results.
2Productivity
If traditional training methods are used to maximize predictive accuracy, then model performance improves, but discrimination against certain groups occurs
Solution Approach 1:
Different predictors in the multipath model are assigned different local qualities based on their bias characteristics. Each predictor processes data through its own inference path with potentially different feature importance weights, allowing the system to leverage strengths of individual predictors while mitigating their respective biases through ensemble aggregation.
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 the likelihood of the inference models providing inferences indicative of bias features. The inference models may be managed using modified split training. The inferences provided by the inference models may be less likely to exhibit latent bias thereby reducing bias in computer-implemented services provided using the inferences. The latent bias may be managed granularly for different bias features to manage predictive power levels for features and bias features.