Fairness Deduplication for Digital Image Databases
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional systems for training vision-language models on large datasets suffer from computational inefficiencies and biases due to redundant images and lack of consideration for semantic concepts, leading to biased neural network outputs.
Innovation Solution
A fairness deduplication algorithm that generates preservation prototypes representing semantic concepts, compares these with semantic embeddings from digital images, and prunes images based on similarity to these prototypes, ensuring fair representation of underrepresented groups.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If conventional deduplication techniques are applied to reduce computational cost, then training efficiency is improved, but model fairness deteriorates due to removal of images from underrepresented groups
Solution Approach 1:
The patent applies local quality by differentiating the treatment of images based on their semantic concept representation. Instead of uniform deduplication, the system preserves images that represent underrepresented semantic concepts while removing redundant images from overrepresented concepts. This selective approach maintains fairness by ensuring adequate representation of minority groups while achieving computational efficiency through removal of redundant majority group images.
Solution Approach 2:
The patent changes the parameter of image selection criteria from conventional similarity-based deduplication to fairness-aware selection. The system modifies the deduplication process by introducing semantic concept analysis and underrepresentation detection, changing how images are selected for retention or removal. This parameter change enables the system to balance computational efficiency with model fairness by considering semantic representation distribution rather than merely image similarity.
2Reliability
If images are removed from training database to reduce bias, then model fairness is improved, but computational resources are wasted on redundant images
Solution Approach 1:
The patent applies the taking out principle by extracting and removing only the harmful redundant images that cause bias, while preserving images that are necessary for maintaining fairness. The system identifies images that can be safely removed without compromising the representation of underrepresented groups, and extracts only those specific images for removal rather than removing all images from overrepresented concepts. This selective extraction reduces computational resource waste by removing only the minimal necessary images to achieve fairness.
3Measurement precision
If semantic concept preservation is implemented, then model accuracy is improved, but algorithm complexity increases
Solution Approach 1:
The patent applies universality by using a multi-functional approach where the same semantic embedding framework serves multiple purposes: it represents image content, identifies semantic concepts, detects underrepresentation, and guides image selection. Rather than implementing separate systems for each function, the system uses a unified semantic embedding model that performs all these tasks, reducing overall algorithmic complexity while maintaining high model accuracy through comprehensive semantic analysis.
Data Source
AI summary
The present disclosure relates to systems, non-transitory computer-readable media, and methods for generating and modifying databases using a fairness deduplication algorithm. In particular, in one or more embodiments, the disclosed systems generate, within an embedding space, semantic embeddings from a plurality of digital images stored in a database. In some embodiments, the disclosed systems identify, from among the semantic embeddings in the embedding space, a preservable embedding according to a preservation prototype indicating a semantic concept to preserve within the database. In one or more embodiments, the disclosed systems generate a modified database by pruning one or more digital images corresponding to semantic embeddings other than the preservable embedding from the database.