Dual Sub-Network Architecture for Contextual Bias Detection in AI
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
AI models trained on large text corpora are susceptible to contextual and associative bias, with existing techniques failing to effectively account for context when generating perturbations to remove bias.
Innovation Solution
A dual parallel AI network with two sub-networks is employed, where the first sub-network determines various contexts of protected attributes across data samples, and the second sub-network adjusts its classification based on the impact of these contexts, allowing for targeted de-biasing.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If existing techniques generate perturbations to remove bias without considering context, then bias removal is simplified, but contextual bias detection accuracy deteriorates
Solution Approach 1:
The system segments the bias detection process into two distinct sub-networks: one for determining contextual information and another for classification. This segmentation allows each sub-network to specialize in its function, improving overall detection accuracy while maintaining computational efficiency.
Solution Approach 2:
The patent introduces an intermediary contextual information determination sub-network that bridges the input data and the classification sub-network. This intermediary processes and structures contextual information about protected attributes before it reaches the classifier, enabling more accurate bias detection without complicating the overall system architecture.
2Device complexity
If a single AI network is used for classification, then device complexity is reduced, but contextual bias mitigation capability deteriorates
Solution Approach 1:
The classification system is divided into two parallel sub-networks: a contextual information determination sub-network and a classification sub-network. This segmentation enables the system to simultaneously process contextual information and perform classification, improving bias mitigation capability while keeping each sub-network relatively simple in structure.
Solution Approach 2:
The patent merges the contextual analysis and classification functions into a unified dual-network architecture that operates in parallel. Both sub-networks process the same input data simultaneously, and their results are combined to produce the final classification output, achieving reliable bias mitigation without requiring an overly complex sequential processing structure.
3Measurement precision
If contextual information is fully analyzed before classification, then bias detection accuracy is improved, but processing time increases
Solution Approach 1:
The contextual information determination sub-network performs preliminary analysis of contextual information in parallel with the classification process. By preparing contextual data beforehand and in parallel, the system ensures that contextual analysis is complete before final classification decisions are made, maintaining high detection accuracy without significantly increasing sequential processing time.
Solution Approach 2:
The dual sub-network architecture enables dynamic parallel processing where contextual information determination and classification occur simultaneously rather than sequentially. This dynamic approach allows the system to analyze contextual information thoroughly while maintaining efficient processing speeds through concurrent operation of both sub-networks.
Data Source
AI summary
Methods, systems, and computer program products for detecting contextual bias in text are provided herein. A computer-implemented method includes identifying, by a machine learning network, a protected attribute in one or more data samples; processing the identified data samples using a first sub-network of the machine learning network, wherein the first sub-network is configured to determine a plurality of contexts of the protected attribute across the identified data samples; determining an impact of each of the plurality of contexts on a second sub-network of the machine learning network, wherein the second sub-network of the machine learning network is configured to classify a given data sample into one of a plurality of classes; and adjusting the second sub-network of the machine learning to account for the impact of at least one of the plurality of contexts on the second sub-network.