Meta-Learning Sample Weighting for Face Recognition Bias
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep face recognition models face challenges due to biased training data distributions, leading to lower accuracy in certain conditions like occlusions and unfairness towards minority ethnicities, as it is impractical to collect data covering a wide variety of variations.
Innovation Solution
A computer-implemented method using meta-learning to train deep face recognition models by calculating sample weights and updating the model based on gradients from both training and validation batches, incorporating instance-level variation-aware loss and meta-learning to rebalance data distributions across multiple factors of variation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If uniform sampling of training data is used, then training process is simple, but accuracy along certain factors (occlusions) decreases and model fairness deteriorates
Solution Approach 1:
The patent applies local quality by assigning different weights to different training samples based on their characteristics. Specifically, it calculates sample weights using a meta-learner that considers instance-level variations, giving higher weights to underrepresented or difficult samples (e.g., occluded faces, minority ethnicities) while maintaining uniform sampling for the overall training process. This resolves the contradiction by making the training process simple overall while achieving high accuracy and fairness through local sample-weighting.
2Reliability
If training data covers a wide variety of variations, then model accuracy and fairness improve, but data collection becomes impractical
Solution Approach 1:
The patent applies preliminary action by pre-calculating sample weights using a meta-learner trained on available data characteristics. Instead of collecting more diverse data, the method proactively identifies underrepresented variations in the existing training set and assigns higher weights to those samples. This preliminary weighting adjustment compensates for data distribution biases without requiring impractical data collection efforts, thereby improving accuracy and fairness while maintaining data collection feasibility.
3Reliability
If sample weights are dynamically updated using meta-learning, then dataset bias is mitigated, but computational complexity increases
Solution Approach 1:
The patent applies partial action by implementing meta-learning only for calculating sample weights rather than for the entire training process. The meta-learner is trained once to learn the mapping from sample characteristics to weights, and then these weights are applied during standard training. This partial application of meta-learning achieves effective bias mitigation while avoiding the excessive computational complexity of full meta-learning training, resolving the contradiction between effectiveness and complexity.
Data Source
AI summary
A method for training a model for face recognition is provided. The method forward trains a training batch of samples to form a face recognition model w(t), and calculates sample weights for the batch. The method obtains a training batch gradient with respect to model weights thereof and updates, using the gradient, the model w(t) to a face recognition model what(t). The method forwards a validation batch of samples to the face recognition model what(t). The method obtains a validation batch gradient, and updates, using the validation batch gradient and what(t), a sample-level importance weight of samples in the training batch to obtain an updated sample-level importance weight. The method obtains a training batch upgraded gradient based on the updated sample-level importance weight of the training batch samples, and updates, using the upgraded gradient, the model w(t) to a trained model w(t+1) corresponding to a next iteration.