Text Encoder Pruning for Fairer Generative AI Outputs
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Generative AI models often exhibit bias in their portrayal of sensitive groups, such as race, gender, and ethnicity, which is difficult to remove entirely from their training datasets.
Innovation Solution
A model pruning approach is employed to generate a debiased model by pruning neuron connections in the text encoder of a generative model associated with specified terms of interest and bias terms, using a binary mask based on a sparsity ratio.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If dataset curation is used to remove bias, then fairness is improved, but it is difficult or impossible to remove all bias instances and the process is impractical
Solution Approach 1:
The patent extracts and removes biased neuron connections from the text encoder by identifying connections associated with bias terms and pruning them using a binary mask. This extraction approach directly removes the source of bias without requiring complete dataset re-curation, making the fairness improvement practical while effectively reducing bias in model outputs.
2Reliability
If model pruning is applied to remove bias, then fairness is improved, but model complexity and performance may be affected
Solution Approach 1:
The patent applies local quality by selectively pruning only the neuron connections associated with bias terms in the text encoder, rather than uniformly pruning the entire model. This localized approach targets specific biased connections while preserving the overall model structure and functionality, thereby improving fairness without significantly impacting model complexity or performance.
3Reliability
If all biased connections are pruned, then fairness is improved, but model performance and functionality may deteriorate
Solution Approach 1:
The patent applies partial action by pruning only the necessary biased connections identified through the binary mask approach, rather than removing all connections. This selective pruning removes sufficient bias to improve fairness while maintaining the model's overall performance and functionality by preserving non-biased connections that are essential for model operation.
Data Source
AI summary
In one implementation, a device obtains one or more terms of interest. The device also obtains one or more bias terms. The device selects a generative model configured to generate an output given a textual prompt. The device generates a debiased model by pruning neuron connections in a text encoder of the generative model associated with the one or more terms of interest and the one or more bias terms.