BDCR: Distributional Contrastive Regularization for Separated Representations

Distinguishing semantically similar classes is challenging when class-conditional features exhibit substantial intra-class variation and inter-class overlap. We propose BDCR (Bhattacharyya-based Distributional Contrastive Regularization), a distribution-level regularizer that represents each class by a Gaussian distribution characterized by its mean and variance. Rather than contrasting individual samples or point prototypes, BDCR contrasts class-conditional distributions using the Bhattacharyya distance, jointly capturing differences in feature location and variation. To focus regularization on ambiguous class relationships, BDCR maintains momentum-smoothed class statistics and dynamically selects hard negative classes, while treating the discrete selection process as non-differentiable. The resulting regularizer can be integrated into standard classification objectives without modifying the network architecture. Experiments across diverse architectures and datasets show that BDCR consistently improves classification performance, with distribution-level hard-negative regularization providing complementary gains over conventional sample- and prototype-based objectives.