Dynamic Sampling Probability for Face Recognition Bias Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing face recognition training methods face challenges in reducing bias due to imbalances in training data, particularly in mini-batch training, where attributes like race and gender are affected by inconsistent sampling frequencies.
Innovation Solution
An image processing apparatus and method that acquires a degree of training for each training target, calculates a sampling probability based on the degrees of training of multiple targets, and performs training using sampled images to balance the sampling frequency, thereby reducing bias in face recognition models.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If mini-batch training is used with standard sampling, then training efficiency is improved, but sampling frequency inconsistency occurs causing attribute bias
Solution Approach 1:
The patent changes the sampling parameter by introducing a sampling probability distribution that varies based on training progress. Specifically, it uses a parameter λ (lambda) that changes over training epochs to dynamically adjust sampling probabilities for different attribute groups, thereby maintaining consistent sampling frequency across attributes while preserving mini-batch training efficiency
Solution Approach 2:
The patent implements dynamic sampling where the sampling probability for each attribute group is not fixed but changes during training. The sampling probability is dynamically adjusted based on the current training epoch and the degree of training for each attribute, allowing the system to adaptively balance sampling frequency across different attributes throughout the training process
2Reliability
If loss function adjustment is used to reduce bias, then attribute fairness is improved, but sampling frequency inconsistency remains unresolved
Solution Approach 1:
The patent segments the training data into different attribute groups (e.g., race, gender) and applies different sampling probabilities to each segment. This segmentation allows independent control of sampling frequency for each attribute group, enabling the system to address sampling frequency inconsistency while maintaining overall attribute fairness through targeted sampling adjustments
3Manufacturing precision
If sampling probability is increased for under-trained attributes, then training balance is improved, but data selection complexity increases
Solution Approach 1:
The patent implements a feedback mechanism where the sampling probability for each attribute group is determined by its current training status. The system continuously monitors the degree of training for each attribute and adjusts sampling probabilities accordingly, creating a closed-loop control system that automatically balances training across attributes without requiring complex manual intervention
Data Source
AI summary
An image processing apparatus comprises one or more processors, and one or more memories storing executable instructions which, when executed by the one or more processors, cause the image capturing control apparatus to function as: a first acquiring unit configured to acquire a degree of training of a training target in a training model, a second acquiring unit configured to acquire, based on degrees of training of a plurality of training targets, a sampling probability of an image of each of the plurality of training targets, and a training unit configured to perform training of the training model, based on a sampling image that is sampled from images of the plurality of training targets based on the sampling probability acquired by the second acquiring unit.