Bac Le

A novel training framework that enhances counterfactual contrastive learning for VQA and introduces a three-stage curriculum for stable multi-objective optimization, an enhanced Batch-Contrastive loss for more discriminative feature learning and two novel regularizers.