Abstract
Factor analysis, often regarded as a Bayesian variant of matrixfactorization, offers superior capabilities in capturing uncertainty, modelingcomplex dependencies, and ensuring robustness. As the deep learning eraarrives, factor analysis is receiving less and less attention due to theirlimited expressive ability. On the contrary, contrastive learning has emergedas a potent technique with demonstrated efficacy in unsupervisedrepresentational learning. While the two methods are different paradigms,recent theoretical analysis has revealed the mathematical equivalence betweencontrastive learning and matrix factorization, providing a potentialpossibility for factor analysis combined with contrastive learning. Motivatedby the interconnectedness of contrastive learning, matrix factorization, andfactor analysis, this paper introduces a novel Contrastive Factor Analysisframework, aiming to leverage factor analysis's advantageous properties withinthe realm of contrastive learning. To further leverage the interpretabilityproperties of non-negative factor analysis, which can learn disentangledrepresentations, contrastive factor analysis is extended to a non-negativeversion. Finally, extensive experimental validation showcases the efficacy ofthe proposed contrastive (non-negative) factor analysis methodology acrossmultiple key properties, including expressiveness, robustness,interpretability, and accurate uncertainty estimation.