Reproduce “Deep Reversible Consistency Learning for Cross-modal Retrieval” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Ruitao Pu, Yang Qin, Dezhong Peng, Xiaomin Song, Huiming Zheng
Researchers from Sichuan University introduce Deep Reversible Consistency Learning (DRCL), a flexible, paired-free framework for cross-modal retrieval that addresses training complexity and semantic alignment issues. DRCL consistently achieved the highest average Mean Average Precision (MAP@all) scores across five benchmark datasets, outperforming state-of-the-art methods by up to 4.0%.
Key insights
- A carefully learned and selectively chosen non-orthogonal prior can more effectively guide semantic alignment and representation learning than randomly initialized or strictly orthogonal matrices.
- Ensuring reversibility in the mapping between the common subspace and category features, via the generalized inverse of the prior, guarantees semantic consistency across modalities.
- Feature augmentation (Mixup) in the embedding space enhances the model's ability to learn over a broader data distribution, contributing to improved robustness and generalization.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.