Bias Detection via Text Embedding Clustering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current bias detection algorithms in machine-learning models are limited by their reliance on image-based embeddings, which fail to identify subtle biases and overlook non-dominant features, leading to incomplete bias mitigation.
Innovation Solution
The proposed system uses text descriptions to distill image datasets into a hierarchy of categories and features via text embedding clustering, identifying pairwise feature correlations, and allowing human experts to select biased correlations for mitigation through data re-weighting.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If image-based embeddings are used for bias detection, then the detection process is computationally efficient, but subtle biases and non-dominant features are overlooked
Solution Approach 1:
The patent introduces text descriptions as an intermediary modality between images and bias detection. Text embeddings serve as a mediator that captures subtle semantic information and auxiliary features more effectively than direct image embeddings, enabling more accurate bias detection without requiring complex modifications to the core detection algorithm
Solution Approach 2:
The patent transitions from a single-modality approach (image-only) to a multi-modality approach by adding text descriptions. This dimensional expansion allows the system to detect biases in both visual and linguistic representations, capturing subtle biases that exist in text but are invisible in image space alone
2Measurement precision
If text descriptions are added to improve bias detection, then subtle biases are identified, but the processing complexity and time increase
Solution Approach 1:
The patent performs text embedding and feature extraction as preliminary steps before the actual bias detection process. By pre-processing the text data and extracting relevant features in advance, the system reduces the computational burden during the main detection phase, thereby minimizing additional processing time while maintaining high detection accuracy
3Measurement precision
If comprehensive feature extraction is performed, then all biases are detected, but false positives increase requiring human review
Solution Approach 1:
The patent applies different analysis strategies to different types of features. Rather than uniformly reviewing all detected correlations, the system identifies and flags only those correlations that exhibit characteristics of potential bias (such as correlations between auxiliary features and target labels), allowing human reviewers to focus on specific high-risk cases rather than comprehensively reviewing all features
Data Source
AI summary
An exemplary method for reducing bias in a training image dataset for training a machine-learning model comprises: receiving a plurality of text strings comprising at least one text string describing each image in the training image dataset; generating a plurality of embeddings based on the plurality of text strings; identifying, based on the plurality of embeddings, a plurality of visual features in the training image dataset; identifying one or more correlations between the plurality of visual features in the training image dataset; receiving a user input identifying at least one biased correlation from the one or more correlations; and training the machine-learning model at least partially by adjusting one or more data sampling weights associated with one or more training images in the training image dataset based on the user input.