New Regularization Improves Fairness in Self-Supervised Learning

This paper introduces Unbiased Open World Regularization (UOWReg), a novel framework designed to enhance fairness in self-supervised learning (SSL) models, particularly Joint-Embedding Predictive Architectures (JEPAs). Existing SSL methods often learn spurious biases from datasets, as their global regularization constraints are insufficient to prevent task-irrelevant features from segregating the latent space. UOWReg shifts the objective from global to conditional distribution matching, explicitly guaranteeing statistical independence between the learned representations and specific sensitive attributes. This ensures that the model's internal understanding of data is not unduly influenced by biases present in the training set, regardless of the chosen target distribution (e.g., Gaussian or spherical). Empirical validation on benchmarks like CelebA shows that UOWReg significantly reduces Equalized Odds violations while maintaining competitive classification accuracy. Furthermore, on a novel "Synthetic Engraving Task," UOWReg successfully isolates fine-grained micro-signatures, preventing the subpopulation collapse often seen in standard SSL, even when these signatures are heavily entangled with dominant macro-structures.